AEO 101 · the playbook that ships what is working in answer engine optimization, right now.

OpenAI began piloting Sponsored Agents in ChatGPT on September 16, 2026: after clicking an ad, users can open a clearly labelled conversation with a business-sponsored agent, kept separate from ChatGPT's own answer and from the user's original thread, for select advertisers in the United States, with Wayfair (limited scale) and Angi named. The same announcement packaged an ad stack: HubSpot as OpenAI's first CRM partner (live September 16), Shopify as its first ecommerce partner (free ChatGPT Ads app in the Shopify App Store since September 3, international from September 23), a natural-language Ads Manager plugin with a human confirmation gate, AI-generated ad copy and imagery, opt-in text customization that rewrites headlines per conversation and auto-translates them, impression-based conversion billing, configurable attribution windows (click 7, 14 or 30 days, view 0 or 1 day) and a 1 to 10 Event Quality Score. Pricing, tier restriction, the model behind sponsored agents and advertiser access to transcripts were not disclosed.

This is the first paid conversational placement inside an answer surface, and Angi's release states the sponsored agent can appear proactively in relevant conversations where its earlier ChatGPT app was opt-in. No AI visibility tool measures inside a sponsored-agent conversation, so a competitor can hold branded conversations with your prospects while your dashboard reads unchanged. With per-conversation text customization, ad creative no longer has a canonical form, so treat any captured ChatGPT ad copy as one rendering. Do not assume the Free and Go tier limit of the existing ads product carries over.

Cloudflare shipped 'Disallow AI Training' on September 15, 2026, publishing a site's no-training preference in robots.txt via Bot Preference Sync and splitting crawlers into two classes: accountable mixed-use operators (Apple, Google, Microsoft) that honour the preference and remain allowed for search, and training-only operators (Amazon, Anthropic, Meta, OpenAI) that are blocked outright with no search impact. Cloudflare reports 17% of sites enable some mechanism to block training against less than 1% blocking search bots. Bingbot currently supports only the NOARCHIVE meta tag, with Microsoft's robots.txt support targeted for early 2027.

The 'block AI training and you risk search visibility' trade-off now has a real answer for Apple and Google, no clean answer for Bing until early 2027, and no downside for the four training-only operators, which never sent search traffic. Audit whether any client property carries a legacy Training block or 'Block AI bots' toggle, and check whether its robots.txt is now Cloudflare-managed before reading intent into it. The 17% figure describes Cloudflare's network, not the web.

Profound raised a $180,000,000 Series D at a $1.8B post-money valuation on September 15, 2026, co-led by Sequoia Capital and Kleiner Perkins with Lightspeed, Khosla Ventures, Saga Ventures, Evantic and South Park Commons participating, taking total raised to $335M+. Its Series C was $96M at $1B in February 2026. Profound disclosed 2,500+ brands and one-third of the Fortune 100 as customers, including Estée Lauder, Walmart, Zoom, Royal Bank of Canada, Stripe, Ramp, Cursor and Figma, but no ARR, growth rate or headcount.

The category's best-funded product company nearly doubled its valuation in seven months, which settles whether investors think AI visibility is a feature or a company. A managed service competes on interpretation and methodology, the things a dashboard does not disclose. Do not draw a growth line from February's 'over 700 enterprises' to today's '2,500+ brands'; the units and denominators differ.

SEOmator scanned Cloudflare Radar's top 100,000 domains on September 13, 2026 and found 8,598 publish an llms.txt, 8.6% of the population and 11.4% of domains that responded. A further 10,775 domains return an ordinary HTML page at /llms.txt (a soft 404), which a naive status-code check counts as adoption, overcounting by 2.35x. At least 29.8% of files carry a platform or plugin signature (Shopify alone 18.2%, WordPress SEO plugins 9.6%), 95.1% open with a spec-correct H1, 16.4% contain no markdown links at all, and .ai domains adopt at 39.3% against 5.5% for .net.

Stop quoting llms.txt adoption as a settled figure unless the study validated that the file was a file; SEOmator did not audit other studies, so treat SE Ranking's 10.13% and Rankability's 8.7% as unvalidated rather than wrong. No major AI vendor has documented consuming llms.txt for discovery, so the file remains a low-priority artefact. If you ship one, make sure it carries links; a sixth of the corpus is an empty envelope.

Otterly published a stability study on September 11, 2026 covering 252,407 AI search answers across seven engines (ChatGPT, Perplexity, Gemini, Google AI Mode, Google AI Overviews, Microsoft Copilot, Claude), 520 prompts and 40 to 91 days per category. With 10 prompts tracked, 90% of Brand Coverage readings for the same brand fell between 39% and 71%; with 50 prompts, 49% to 62%; with 100 prompts, 51% to 60%, a roughly 73% reduction in variation from 10 to 100. Day-to-day source reuse over 30 days ran Perplexity 74.9%, Claude 62.3%, Copilot 46.2%, Google AI Overviews 36.1%, Gemini 28.4%, Google AI Mode 26.4% and ChatGPT 25.8%; across 45,919 distinct URLs, 48% were cited exactly once and the top 1% produced 44% of all citations.

Any before-and-after built on ten prompts sits inside the noise band; treat 50 prompts per brand as the floor for a reportable number and 100 as the target. Perplexity remembers its sources and ChatGPT does not, so expect ChatGPT citation wins to churn week to week. Otterly sells prompt tracking, so attribute the curve to Otterly, but it is the most useful measurement rule the category has published.

Amazon Ads announced on September 10, 2026 a partnership with OpenAI that lets Amazon Ads advertisers extend campaigns into conversational advertising experiences in ChatGPT, United States only, in a pilot phase with a select group of advertisers. Delta Vacations is the one brand named. The announcement discloses no ad formats, brand count, spend, CPMs or attribution mechanics.

This is the first major demand-side ad platform plugging its advertiser base into an answer engine: the buying relationship stays with Amazon while the inventory is OpenAI's. Expect ChatGPT ad inventory to be bought through the same DSPs that buy everything else, which makes ChatGPT ad volume less observable from outside, not more. Attribute any format or delivery detail to trade coverage, not to the primary.

Profound Research published a persona study on September 9, 2026: 71,147 responses collected August 11 to 24 across ChatGPT, Claude and Gemini in three categories (Credit Card, Clothing, Furniture), crossing gender, age, income and occupation personas against a same-persona repeat baseline. Only 1 in 4 brand mentions were shared across different personas against 2 in 5 across repeated same-persona responses; cited-source overlap was 20% across personas against 34% same persona. High-income personas received 7.5 brand mentions and 4.8 citations against 5.8 and 4.1 for low-income, and in Gemini clothing answers the high-income persona got 24 percentage points more brand-owned-site citations and 17 percentage points fewer earned-media citations.

Brand visibility is persona-specific, so any single owned-versus-earned percentage is an average over customers you may not target. Segment prompt sets by the buyer persona you actually sell to and report visibility per segment. Persona costs roughly as much visibility as raw nondeterminism, so measure the same-persona baseline before claiming a persona effect. Profound sells a Personas feature; attribute the study.

Otterly published its AI Search Index on September 8, 2026: 6,095,520 citations across 17 US industries and seven engines (ChatGPT, Gemini, Perplexity, Google AI Mode, Google AI Overviews, Microsoft Copilot, Claude), collected through July and August 2026. Company-owned pages take 53.7% of all citations, from 77.4% in Industrial/Manufacturing to 29.6% in Nonprofit; journalism 16.9%; government/NGO 8.1%. In ChatGPT after August 8, owned pages rose from 44.38% to 55.28%, journalism fell from 21.33% to 10.40%, Reddit fell from 4.42% to 0.39% and YouTube rose 239%. Google AI Mode produced 21.02% of citations and cited 69,390 distinct domains but used 436 domains for half its citations. Otterly also measured 107,633 ChatGPT ad units and 15,363 shopping cards from 6,592 distinct advertisers, finding almost none of the largest advertisers among the most-mentioned brands organically.

About half of what AI search cites is the company writing about itself, which converts into a per-vertical budget argument: how much of your AEO ceiling you control directly. In ChatGPT, what replaced Reddit was owned pages and YouTube, not journalism. Paid and organic populations inside ChatGPT are largely disjoint, so buying ads is not a route to organic mention share. Never quote the Reddit decline as a bare percentage; Qwairy, PromptWatch and Otterly measure the same event on different denominators.

Conductor's study of 167,867,680 Google AI Overview citations (June 15 to August 15, 2026, US, every industry, no keyword sampling, page last updated September 7, 2026) found 57% of AIO citations go to pages outside the organic top 10 and 43% inside. Position 1 supplies 8.5% of all citations, position 2 6.6% and position 3 5.3%, just over 20% for the top three combined. The likelihood that a page at a given rank is cited runs 24.9% at position 1, 21.1% at position 2, 17.9% at position 3 and 10.2% at position 10.

Rank still helps but weakly: a 2.4x spread from position 1 to 10 in citation rate, while share collapses because there are only ten positions and an unbounded citation pool. The much-quoted 'most AI citations come from outside the top 10' is mostly a statement about population sizes and about how many citations per answer a study counted, not evidence that ranking stopped mattering. Compare only figures that share a citation depth; Ahrefs' 76.10% counted only the top three citations per answer and its 37.9% counted all of them.

Semrush published a consumer survey on September 7, 2026, run by Exploding Topics: 2,338 US adults surveyed in July 2026. 57.5% of AI users have decided not to buy something based on information from an AI chatbot (69.7% among power users); 65.01% have at least partially replaced Google searches with AI chatbots; 74.15% would be at least somewhat less likely to buy if a chatbot flagged mixed or negative reviews; 73.58% of power users have bought on an organic AI recommendation; 41.63% dislike chatbot ads and 66.05% of those say ads make them doubt AI integrity.

This is demand-side evidence that AEO sentiment work is revenue defence: three-quarters of buyers say a single flagged negative moves them. The chatbot acts as a filter, not a brake; the same power users most often talked out of a purchase are the most likely to buy on a recommendation. Track what engines assert about your brand, not only whether you appear. These are stated preferences, so do not present the ads dislike as a prediction of ad performance.

Peec AI reported on September 4, 2026 that ChatGPT runs an internal search index family called Labrador, with verticals for general web, PDF, YouTube, news, arXiv, Wikipedia, local, finance, legal, medical, shopping and images. Server-side event data observed May 21 to July 21, 2026 showed a result_source field returning four values: Labrador, Bright, Oxylabs and SERP. Peec documented eight retrieval sources including OpenAI's own crawler, Bright Data, Oxylabs, Yelp, TripAdvisor and Microsoft's Web IQ, a crawl of 6 million pages at 35,000 requests per hour on a billion-page test site, and five live Shopping A/B experiments as of September 2, including prefer-index-over-serp-v3 affecting 8% to 20% of chats. Nick Turley's antitrust testimony placed the start of the index in 2023.

If ChatGPT retrieval increasingly prefers its own index over third-party SERP data, 'optimise for Bing' is optimising for a shrinking input; the target becomes being crawled and indexed by OpenAI directly and being present in the vertical sources it pulls from, such as Yelp, TripAdvisor, YouTube and arXiv. The finding is observational and vendor-authored with no OpenAI confirmation, so report the observed result_source values and experiment names and attribute the architecture claim to Peec.

Profound published a pricing study on September 3, 2026 with Kyle Poyar of Growth Unhinged: 7,600 AI responses on 77 Cloud 100 companies across six engines (ChatGPT, Google AI Mode, Google AI Overviews, Gemini, Microsoft Copilot, Perplexity), July 29 to August 10, 2026, using a single prompt template. 46% of responses cited a company-owned pricing page somewhere but only 12% cited it as the primary source; ChatGPT was the most brand-favourable engine, citing company pages first 38% of the time; third-party sources led with Vendr 18.7%, Reddit 18.6% and G2 15.9%. 57 of 77 public pricing pages were machine-readable, and pricing queries are 1% to 2% of all LLM queries.

Being cited and being the source of record are different: a brand can appear in nearly half of pricing answers while the substance comes from Vendr, Reddit and G2. Make the pricing page machine-readable first, then work the third-party sources engines actually draw from. Profound sells AEO software, so attribute the figures.

Google's generative-AI performance reports and its Search generative AI control rolled out to all websites worldwide as of August 31, 2026, per a note on Google's Search Central blog. The reports show impressions from AI Overviews, AI Mode and generative AI in Discover by page, country and date (device in the Search report); the property-level control in Search Console settings excludes a site from those three surfaces while it stays indexed and ranked normally. The reports carry impressions only, no click data, and the Gemini app is excluded from the opt-out.

Every site now has a first-party AI impression signal and a domain-level kill switch; the June 17 opt-out enforceability entry is no longer UK-subset only. Keep the standing read that almost no B2B brand should opt out, since the control is all-or-nothing and forfeits the impression. Use the reports as a visibility signal, not a traffic instrument, and do not read Gemini app presence from them.

ChatGPT Ads went live in India on August 27, 2026 on the ChatGPT Free and Go subscription tiers (Go at ₹399 per month), with ads shown at the bottom of responses. OpenAI said it would start by showing ads for 50 brands, has partnered with agencies WPP and Omnicom, and would launch an ad manager the following month with a daily minimum budget of ₹725 ($7.60); Indian outlets reported the self-serve Ads Manager opening on September 4.

India lands three days after the 31 European markets that opened on August 24, so ChatGPT ads are now a multi-market surface with a low self-serve floor. Advertisers in these markets should expect the paid population to sit largely apart from organic mention leaders, per Otterly's later measurement, so treat ads and AEO as separate lines. OpenAI's own ads page was unreachable for verification, so treat the September 4 date as secondary-confirmed.

OpenAI retired o3 from ChatGPT on August 26, 2026, on schedule after a 90-day sunset that began with a May 28 announcement. o3-pro is not retiring and remains in ChatGPT for Pro, Team, Enterprise and Edu, and the API is unaffected, with the o3-2025-04-16 and o3-pro-2025-06-10 snapshots carrying December 11, 2026 removal dates. The Assistants API shut down the day before, on August 25, with no automated migration path.

Treat August 26 as a hard boundary in any ChatGPT citation series: the model left consumer tiers but stayed on paid ones, so a panel that mixes tiers is measuring a composition change, not a behaviour change. Ask any vendor reporting a late-August ChatGPT movement to state its tier mix across that date. If you run your own prompt harness, tag every run by model and tier from now on.

ChatGPT Ads went live on August 24, 2026 across 31 European markets, the 27 EU member states plus Norway, Iceland, Liechtenstein and Switzerland, taking the live footprint from 9 markets to 40 in a single day. OpenAI's announcement, dated August 18, calls it the company's largest expansion to date, says tens of thousands of marketers have now advertised on ChatGPT, and lists an advertiser stack of conversion-optimization bidding beyond CPM and CPC, geo-targeting, custom audiences, an OpenAI Pixel, a Conversions API and third-party measurement integrations, with self-service Ads Manager access to follow later this summer. Ads run on Free and Go only, Plus, Pro and Enterprise stay ad-free, and the European launch is widely reported as contextual-only (conversation topic, approximate location, device type, time of day and language, no chat history), although OpenAI's own text states neither that nor any GDPR legal basis.

If you sell into Europe, ChatGPT is now a paid surface as well as an organic one, and the paid side has a pixel and a conversions API while the organic side still does not. Baseline your brand's presence in ChatGPT answers in your European markets now, before ad load builds, so you can separate organic movement from paid displacement later. Do not quote any single ad-load percentage: vendor panels measuring US ad frequency disagree by more than 2x on overlapping windows.

Profound published 'Claude and Claude Code are distinct Answer Engines' on August 24, 2026: across 24,135 responses to 1,724 prompts (prompt window July 13 to 23), Claude searched the web in over 93% of responses while Claude Code searched in 13%, despite identical underlying models. The two interfaces mention only 1 in 5 of the same brands, worse than each system's own repeat consistency of 50% (Claude) and 40% (Claude Code). Claude Code responses average 322 words against Claude's 459, with lists in 94% versus 56% and tables in 54% versus 11%.

'How visible are we in Claude' is an underspecified question; specify the interface. If your buyers are developers, measure Claude Code separately and expect its crawler to hit documentation, informational and pricing pages, where nearly three-quarters of its agent visits landed. Single vendor, no confidence intervals and a coding-heavy prompt mix, so quote the 13% versus 93% search rate and the 1-in-5 overlap, and do not extrapolate the overlap to other model and interface pairs.

Press Ranger and OtterlyAI released a study on August 20, 2026 of 129.3 million citations across more than 20 million cited URLs, captured in June 2026 across ChatGPT, Google AI Overviews, Google AI Mode, Perplexity, Microsoft Copilot, Gemini and Claude, matched against 91 confirmed AI-publisher licensing deals signed since 2023. News pages from publishers with an OpenAI licensing deal earned 48% more citations on ChatGPT than pages from publishers without one, while Google-licensed publishers were cited on Google AI Overviews at a slightly lower rate than comparable unlicensed publishers and Perplexity's licensed publishers landed at parity. The most-cited news domains, NerdWallet, Healthline and Bankrate, hold no deals, and best-of lists, buying guides and product reviews make up 46.9% of licensed publishers' citations.

A licensing deal is a ChatGPT-specific tendency, not a universal lever and not a floor: Reddit holds an OpenAI deal and lost most of its ChatGPT citation share the same month. Format and commercial intent beat contracts, so build the buying guides and comparison pages rather than chasing a deal. When citing this, disclose that Press Ranger and Otterly have been commercial partners since November 2025, that the analysis covers one month, is not peer reviewed and publishes no method for selecting comparable unlicensed publishers; the study's 112% figure for OpenAI-exclusive publishers has no reachable methodology and should not be quoted.

Google began rolling out its August 2026 spam update on August 18, 2026, its third spam update of the year, expected to take several days. In the same window generative UI, which launched in AI Mode in November 2025 alongside Gemini 3 and is now global in English there, started reaching AI Overviews, building custom layouts, interactive tools and simulations matched to the searcher's question. Preferred Sources passed 600,000 unique sources selected by users, up from 345,000-plus in May 2026, rolling out across Search, Discover and Google News.

Any AI Overviews or AI Mode citation-share movement dated after August 18 has an index-level event inside its window, so do not attribute a late-August swing to your own changes without checking the rollout dates. A rendering change inside AI Overviews can reclassify citation formats, so re-baseline format-level tracking once it lands. Encourage your audience to add you as a Preferred Source, since the selection count rose 74% in three months. Separately, a logging error depressed Search Console's generative AI impressions for data from August 13 to about August 20, so that window is a bug, not a trend.

Profound published 'Where do inaccurate AI claims come from?' on August 17, 2026, analysing 158,000-plus claims via its FactCheck product with a snapshot dated August 14. The baseline inaccuracy rate ran 4.3% to 6.8% across answer engines, Claude was more than 1.3x more inaccurate than ChatGPT, and only 12.5% of returned claims and 11% of inaccurate claims overlapped between platforms. 54% of brands had inaccurate claims citing their own content, more than the 51% with an inaccurate claim from third-party earned media; pricing and billing was 12% of evaluated claims but 24% of inaccurate ones, and the top 100 earned-media domains account for only 40% of inaccuracies.

Audit your own pricing and billing pages first: a brand's own site is the single most common cited source of an inaccurate claim about it. There is no PR shortcut, because the sources of inaccuracy are a long tail and correcting ten domains fixes little. Measure accuracy per engine, since the claims and the errors barely overlap across platforms. Profound flags a possible selection effect on pricing, so treat the topic ranking as directional.

Reddit's share of ChatGPT Search citations collapsed in two steps on August 8 and August 14, 2026. PromptWatch measured Reddit at 3.83% of citations from July 18 to August 7, then averaging 0.52% from August 14 to 17, an 86.4% relative decline, with its trailing 7-day figure already back at 1.50%; Qwairy, on a fixed brand cohort across August 1 to 18, measured 2.05% to 0.07% with a median per-brand decline of 97.9% and no brand falling less than 50%. On August 8 the share of ChatGPT fanout queries using the site: operator jumped from 0.37% to 16.8% in a single day while fanouts per response also rose from about 1.08 to about 1.83; the decline was ChatGPT-specific, with Reddit flat to slightly up on Google AI Overviews over the same seven days, ChatGPT's total citation volume rose 3.5% over the window per Otterly, and the displaced share went to official brand sites, institutional sources, help centres and product documentation.

Report the endpoint, not the percentage: four panels give four different declines off baselines that differ by up to 8x, and PromptWatch itself says a data-collection issue cannot be ruled out. Date-stamp any deck that leans on Reddit's citation weight in ChatGPT, and expect the same shape on YouTube, TikTok, LinkedIn and Facebook, since a fifth panel reports the decline was social-wide on ChatGPT. Fund your help centre and product documentation, which absorbed the displaced share, but remember owned properties are still a minority of brand-mentioning citations. A September 2025 Reddit drop on Profound's corpus later rebounded, and Reddit's robots.txt has carried a blanket Disallow since at least December 2025, so neither a crawler block nor a permanent wipe is established.

Profound published 'Commercial conversations in ChatGPT more than doubled in a year' on August 13, 2026, analysing 7.5 million English-language ChatGPT conversations from June 23, 2025 to June 22, 2026 across 53 weekly aggregates. Commercial intent rose from 13.9% to 19.2% of all conversations, weekly commercial conversations rose from 243 million to 533 million (about 28 billion a year), and commercial conversations per active user per week rose from 0.38 to 0.54, a 41% increase. Automotive ran 59.1% commercial, Travel and Hospitality 56.1%, Telecom and Communications 47.9%, Hardware and Industrial 47.6% and Financial Services 47.1%.

Roughly one in five ChatGPT conversations now carries buying intent, and Digiday separately reported OpenAI seeing roughly 20% of queries with direct commercial intent, so the figure is safe to use. Build for the decision stage inside the chat, not the discovery stage: commercial conversations averaged 2.43 turns, down from 3.20, and 53% were single-turn, so the answer has to land in the first reply. Prioritise engines and prompts by vertical, since the commercial share varies by well over 2x across industries.

ChatGPT Ads went live in Mexico and Brazil on August 11, 2026, taking the Ads Manager footprint from 7 to 9 markets (US, Canada, Australia, New Zealand, UK, Japan, South Korea, Brazil, Mexico), alongside campaign optimisation for specific conversions, a carousel format showing multiple products from a single advertiser, and measurement integrations with Hightouch and Triple Whale. OpenAI's advertising policy document, updated July 31, 2026, defines 'sponsored agents' as conversations users hold with an AI representative of an advertiser's business. Automatic advanced matching, which captures identifiers users enter into website forms, hashes them client-side and sends them with the conversion event, became the default on pre-existing pixels on August 17 on an opt-out basis; reported economics include about $100M in ads annual recurring revenue as of May 2026, 7,378 distinct advertisers as of mid-July per Adthena with 60.1% US-based, and CPMs falling from $60 at launch to $25 by mid-April 2026.

Sponsored agents are a different unit from a labelled link under an answer: paid conversational placement now has a formal definition, so plan for a competitor's agent being the one that answers. If you run ChatGPT Ads, check your pixel's advanced-matching setting under Tools, Conversions, Data Source, because it is on unless you turned it off. OpenAI's own pages were unreadable to the research team for most of the month, so every figure here is secondary-sourced; verify against the primary before quoting economics to a client.

Perplexity blocked Time's markdown 'agent ads' on August 11, 2026, with chief communications officer Jesse Dwyer calling them deceptive and, in later comments, cloaking, and warning that publishers deploying them risk a reputational downgrade in Perplexity's search index including a hit to their trust score. Time had converted all of its webpages to markdown around June 2026 and, from July 30, began selling one FAQ-format ad per markdown page, generated by ad-tech platform Mobian from a brand brief, invisible to human readers and visible to AI agents, with Ally Bank and the Project Management Institute as first buyers. Time COO Mark Howard told Digiday that Time sees more bot traffic than human traffic on most days.

This is the first paid placement inside the document the model reads rather than beside the answer, and the first enforcement action against one; Perplexity is the only engine to have ruled, and the Google and Bing statements circulating as a response date from February 2026. Do not buy agent-only placements while the only ruling calls them cloaking. Audit your own markdown and agent-facing surfaces for anything that differs from what humans see, because the same mechanism is available to anyone converting pages to markdown. Note that Time is also a Perplexity revenue-share publisher partner.

PromptWatch published July 2026 citation-type breakdowns for ChatGPT Search and Google AI Overviews on August 11, 2026, from a panel it describes as 26 billion-plus data points. On ChatGPT Search, product pages were 32.8% of classified citations against listicles at 9.7%, a 3.4x gap, with product pages nearly doubling from an 18% baseline in March, and video was 0.1%. On Google AI Overviews listicles lost the top spot for the first time on record, with product pages overtaking on a daily basis from July 28 and ending the month at 17.9% versus 16.2%; full-month AIO shares were listicles 18.0%, product pages 16.3%, how-to 15.1%, news 13.5% and video 5.9%, with listicles down from about 26% in Q1 and product pages up from about 9% in January.

The two biggest engines have different format economies, so stop optimising for 'AI search' as one surface: video is 5.9% of AIO citations and 0.1% of ChatGPT's. The standard playbook of getting onto third-party best-of lists is losing share on AIO and is a publisher's playbook on ChatGPT; the page the brand controls is the rising format on both, which also makes Profound's finding that brands' own pages are the top source of inaccurate claims more expensive. Report the direction and the dated crossover, not the decimals: the classifier is undocumented and the ChatGPT categories sum to only 58.9%.

The canonical llms.txt specification at llmstxt.org shipped version 2 on August 10, 2026 (original publication September 3, 2024, Jeremy Howard of AnswerDotAI), rewritten based on two years of adoption. v2 adds standard link relations so agents can find a page's markdown version without guessing, rel='alternate' with type='text/markdown' pointing to the markdown version and rel='describedby' pointing to the governing llms.txt, deliverable as HTML link elements or HTTP Link headers; it blesses both page.html.md and page.md URL patterns, adds hierarchical most-specific-file-wins coverage, and removes llms_txt2ctx and the mechanical meaning of Optional sections. Chrome's Lighthouse now retrieves the file in an Agentic Browsing check and flags server errors, and the spec notes that OpenAI, Anthropic and Google's Gemini team publish llms.txt files for their own developer documentation.

This is a spec event, not an adoption event: a seven-month server-log study across about 900 domains logged 1,227 requests to llms.txt files and not a single real AI bot, Google says the file neither helps nor harms rankings, and no engine has committed to reading it in production. Publish one for agent readiness, because v2 now tells you exactly what to emit and it costs nothing, but do not expect citation lift. When someone says the AI labs have adopted llms.txt, note that they publish one for their docs, which says nothing about whether they read yours.

Google confirmed on August 10, 2026 that it is testing a redesigned Search homepage for signed-out desktop users: three buttons, create images, ask about files and brainstorm, appear beneath the search box behind a plus sign, an AI Mode button sits inside the search field, and 'I'm Feeling Lucky' survives; ask about files and brainstorm worked without an account while image generation still requires sign-in. Robby Stein, VP of Product, on X: 'This is a small test on desktop to help people find new things they can do with Search' and 'No impact to how the core search box works.' The treatment was observed in stable Chrome signed-out, in Microsoft Edge and in Perplexity's Comet browser, so it is a server-side test rather than a Chrome feature.

Record the event, not the headline: aggregators ran it as Google demoting search, and Google's own statement says otherwise. What is measurable is AI Mode entry-point prominence, which nobody has measured; if AI Mode's share of your Google-sourced AI visibility rises in the coming weeks, this distribution test is one candidate cause. Keep AI Mode and AI Overviews tracked as separate surfaces, since they cite the same URLs only about 13.7% of the time.

ChatGPT Atlas stopped working on August 9, 2026, nine months after its October 21, 2025 launch, following a deprecation announced around July 11, 2026. OpenAI's framing: 'We're deprecating Atlas and moving browser-based agentic capabilities into ChatGPT and Codex,' with multiple tabs, downloads, improved navigation and account login support named as the replacement capabilities inside ChatGPT itself. Coverage attributes the failure to Atlas never shipping beyond macOS, a documented prompt-injection vulnerability, and agent latency described as ten minutes to add three items to an Amazon cart.

Agentic browsing did not go away; it moved from an opt-in app into the surface with the most users, so any measurement design that treated Atlas as its own engine is broken and any that treated agentic browsing as niche is wrong the other way. Add agentic-browse versus standard answer as a segmentation axis in your ChatGPT tracking, because it changes the retrieval path entirely. Make sure your site works for an agent that needs to log in, navigate and download, not only for one that reads.

OpenAI made GPT-5.6 Luna the default model for ChatGPT's Free and Go tiers on August 7, 2026, replacing GPT-5.5, announced a 'Think' button giving Free users higher reasoning on demand the following week, upgraded GPT-5.6 Sol for Plus and Pro alongside the thinking slider, and removed limits on text-based chats for all users. OpenAI's internal figures put factual errors down 62% for GPT-5.6 Luna and 68% for GPT-5.6 Sol versus GPT-5.5-Instant.

This is a model swap under the largest user segment, not a feature toggle, and OpenAI's release notes were not directly readable at the time, so the details rest on secondary coverage. ChatGPT's citation surface is now segmented at least four ways, Luna versus Sol, slider position, Think on or off, and free versus paid, and no monitoring vendor segments by any of them; Semrush measured only 25.6% of sources surviving a reasoning-mode switch, so ask any vendor which tier and reasoning depth their ChatGPT numbers sample. The error-reduction figures are unaudited and describe factual accuracy, not citation behaviour.

OpenAI updated GPT-5.6 Sol in ChatGPT on August 6, 2026 for more direct responses, tighter formatting and less unhelpful elaboration, and gave Plus and Pro users a slider to choose reasoning effort across web, mobile and desktop, alongside universal search across chats, projects, images and documents. Reasoning depth had previously been assigned by OpenAI's router; from this date it is a user-controlled variable at consumer scale. Separately, on August 1 OpenAI announced 'Astra', its next major model, inside a mathematics blog post, with no release date, pricing or model card.

The same brand, same prompt and same day can now yield materially different citation sets depending on where a user left a slider; Semrush and Kevin Indig measured only 25.6% of sources surviving a reasoning-mode switch and Reddit's share halving from 15% to 7% under high reasoning. Sample across slider positions in your own testing or state that you did not, and treat any single-setting ChatGPT visibility score as a partial view. Tighter answers plausibly mean fewer cited sources per answer, which is unmeasured, so watch your citations-per-answer trend from this date.

Cloudflare added an AEO Visibility Dashboard to its AEO Suite on August 6, 2026, in early access requestable from the Cloudflare dashboard Overview tab, alongside the existing Agent Readiness tool, with four metrics: Citation Rate (the share of answers in your category that cite your site), Mention Rate (how often assistants name your brand whether or not you are cited), Prominence (how much of the answer is yours and how early it lands) and Share of Voice (your slice of citations against competitors). Cloudflare's engineering blog states it probes 'the leading assistants (today, Anthropic's Claude and OpenAI's GPT)', so currently no Gemini, Perplexity, Copilot or Grok, and no pricing has been disclosed. Agent Readiness is tiered from robots.txt, sitemaps and Markdown serving through Content Signals and API catalogs, then OAuth, MCP and agent cards, up to the x402, ACP, UCP and AP2 commerce standards.

The infrastructure layer in front of roughly a fifth of the web now measures AI visibility as a bundled feature from a vantage point every panel vendor has to approximate, and its readiness ladder ends at payment rails, so read it as a commerce onboarding funnel with monitoring on top. Request early access if you are on Cloudflare and compare its Citation Rate against your current vendor on the same prompts. Do not treat it as a replacement yet: two engines only, no price, and as of the end of August not one GEO vendor had publicly responded, even though Profound integrates with Cloudflare and Peec's Crawl Insights runs on Cloudflare data.

Adobe Brand Visibility reached general availability on August 4, 2026, per Adobe's release notes at abv.adobe.com (announced August 1 for August 3): 'LLM Optimizer is now Adobe Brand Visibility', combining Adobe's optimisation workflow with Semrush Enterprise AIO data. The metrics were redefined in the process: 'Brand Visibility' replaces the old Visibility Score and counts brand mentions only, a new Source Visibility counts how often your site is used as a source, 'Topic Volume' replaces Popularity with actual search-demand figures, and Share of Voice is now weighted by search demand and ranking position; an Impact Measurement Engine re-runs a fixed prompt set over about a two-week cycle, and a free no-login demo is live at play.bv.now. A further release note dated August 20 reversed part of this for paying customers, renaming Responses with Mentions and Responses with Citations to Mentions and Citations and removing the metrics that counted each occurrence within a response.

Any Adobe client comparing August numbers to July is comparing different metrics with the same names, and a series spanning August 4 to August 20 has a second definition break inside it, with paid and unpaid tiers now on different metrics. Re-baseline and date-stamp every vendor-score chart, and ask Adobe or any vendor for the metric definition in force on each date. Adobe's own statement that its measurement runs on fixed prompt sets rather than traffic is the honest answer to why a visibility score and analytics disagree; note that secondary sources still date the rename to August 21, which the primary contradicts.

The Interactive Advertising Bureau published 'Measuring Visibility in the AI Era' on August 3, 2026, a 36-page standard stating that more than 20 companies sell AI-visibility tools using different query sets, platform coverage, scoring rubrics and definitions, so two of them 'can measure the same brand in the same category during the same week and return materially different results.' It defines the 4 P's, Presence (Mention Rate, Citation Rate, Share of Voice, Visibility Momentum), Prominence (Position), Portrayal (Sentiment, Framing, Hallucination Rate, Factual Inaccuracy Rate) and Persuasion (Recommendation Strength, Post-Citation Click-Through Rate), plus a two-tier quality bar in which 'directional' data 'does not meet the bar for budget allocation or executive strategy' and 'decision-grade' requires all four intent types (informational, comparison, recommendation, transactional) with distribution disclosed, weekly-or-better cadence, enough responses per query to establish a stable distribution, and platforms representing a substantial majority of consumer AI traffic; fewer than 50 queries is classed as exploratory. The 13-member working group includes Tinuiti, IQRush, Walmart, EMARKETER, Acxiom, WPP Media, Microsoft Clarity, PMG, emberos and the Alliance for Audited Media.

You can now say 'IAB decision-grade' instead of arguing measurement theory from first principles: demand all four intent types, weekly cadence, a disclosed sampling depth and multi-platform coverage from any vendor before using its number for budget. The standard sets no variability threshold (deferred to a working group) and names no auditor, so expect vendors to self-certify; ask for the methodology, not the badge. Profound, Semrush and Adobe, Peec, Otterly, Scrunch and Conductor are not on the working group.

OpenAI cut GPT-5.6 API prices on July 30, 2026. Luna dropped from $1.00 / $6.00 to $0.20 / $1.20 per million input/output tokens (80% cut), Terra from $2.50 / $15.00 to $2.00 / $12.00 (20% cut), while Sol stayed at $5 / $30. A new API Fast mode for Sol runs up to 2.5x standard speed at 2x the standard price and replaces the Priority Processing tier.

Cheaper inference on the tier that serves Free and Go users removes cost as a reason to keep default reasoning shallow, and Semrush's June 30 study showed reasoning depth alone turns over 74% of cited domains. Treat this as a testable prediction, not a finding: if default reasoning rises, expect Reddit's citation share to fall and documentation and institutional sources to rise. Re-measure ChatGPT citation source mix over the next few weeks rather than assuming the answer surface is unchanged.

Profound published 'Where do AI citations come from?' on July 30, 2026, analyzing 11.84 billion citations across 8 models and 29 industries from April 16 to July 16, 2026, mapped against a 3.02-million-domain classification covering 98.3% of citation volume. Roughly 57% of citations point at brand sites overall, ChatGPT 47% versus Gemini 69%; cybersecurity has the highest median brand share at 74% and government/nonprofit the lowest at 15.9%; pharma draws 59% of citations from earned media while SaaS/software draws only 11.4%; Google AI Overviews cite a social source about 1 in 4 times, Copilot about 1 in 29. Profound defines a brand site as any company's own web property, so competitor domains count as brand, and in 24 of 29 industries the median customer gets more citations from brand sites (competitors, adjacent companies) than any other bucket.

Do not blend the 57% brand figure with the 82% to 98% third-party citation band; they use different denominators and both are correct. For B2B SaaS, lead with the 11.4% earned-media share: PR and editorial placement are structurally weak citation levers in software, and the citation graph is dominated by other companies' properties and social/UGC. Report brand share per platform, since a 22-point ChatGPT-versus-Gemini swing disappears in a blended number.

Otterly published 'The 2,000 AI Blogs Experiment' on July 28, 2026: two new domains were each filled with about 1,000 fully AI-generated articles in under a day and tracked from April 1 to July 27 across Google and seven AI platforms, yielding 482,394 tracked citations. Google algorithmically deindexed both sites with zero manual actions (one in about 8 days, the other in about 12 weeks), while AI citations to the second site rose from 100 in April to 1,666 in 27 days of July. Microsoft Copilot supplied 2,718 of that site's citations (69%), Claude 714, Perplexity 501, and Google AI Overviews, AI Mode and Gemini combined cited it exactly zero times; ChatGPT-User crawled 4,808 pages but cited almost nothing.

Google's AI surfaces inherit Search's quality judgment while Copilot, Claude and Perplexity do not, so AI visibility cannot be one number or one strategy. Never scale unedited AI content on a domain you value, and monitor indexed page counts rather than manual actions, because algorithmic suppression gives no appealable signal. Add Copilot to the tracked set: it is the least-filtered major surface and the one most clients ignore. Caveat: two sites, one synthetic topic, and the intended author-bio test failed.

Otterly's 'Dead Citations' study, published July 24, 2026, re-fetched 20,092,769 URLs cited across seven AI engines over one month and found 19.3% do not return a live page. Per engine: ChatGPT 25.1% (5,779,229 URLs), Google AI Mode 23.9%, Microsoft Copilot 20.0%, Gemini 18.7%, Perplexity 18.0%, Claude 16.7% (67,191 URLs), Google AI Overviews 12.6% (6,260,034 URLs). 7.7% of cited Wikipedia URLs and 15.1% of cited Reddit URLs were no longer live, and 77% of cited PDFs carried no date.

Roughly one in five citations an engine hands a user is already broken, so every raw citation count carries an error bar, including your own audits; say so before a competitor does. Engines answer from stored snapshots, which makes redirect hygiene a GEO task: migrated URLs rot silently and nobody gets a 404 report from ChatGPT. The 12.5-point spread between ChatGPT and AI Overviews means citation counts are not comparable across engines.

Seer Interactive published 'Content Recency's Impact on AI Visibility in 2026' on July 24, 2026, covering 7,683 cited pages generating 47,097 citations across ChatGPT, Gemini and Perplexity from March to June 2026 for four brands in four verticals. 75% of cited pages were updated within the past year and 88% within two years; on the 4,124 pages carrying both dates, 72% were updated within a year but only 42% were published within a year. Gemini was strictest (78% of citations from pages updated within a year), ChatGPT 73%, Perplexity 65%. Pages cited in all four months were the oldest cohort (median update age 0.47 years). Owned content was about 2% of cited material, 98% was third-party.

The freshness engines reward is manufactured by updates, not new publishing, so a systematically refreshed existing library will out-earn a net-new content calendar. Reframe the work as maintenance rather than a treadmill: established plus maintained gets staying power, newest only gets the spike. Prefer this over the AirOps December 2025 83% figure for 2026 recency claims, and remember it rests on four brands, so do not generalize the vertical splits.

OpenAI shipped eight ChatGPT Ads features on July 24, 2026, announced by product email to Ads Manager beta participants: conversion-optimized bidding (oCPC) with a new Conversions objective, automatic advanced matching using hashed customer data, a seven-day rolling average budget model replacing fixed daily caps, a bulk API for asynchronous campaign operations, enhanced product cards showing pricing and star ratings, AppsFlyer and Adjust mobile attribution integrations, and geographic exclusion targeting.

The conversions objective moves ChatGPT Ads from an awareness placement toward a performance channel that will compete directly with organic AI-visibility budget. Watch product cards with star ratings closely: review-platform authority now carries commercial weight on both the paid and organic sides of the answer. None of this changes citation behavior, and the confirmed live footprint is still five markets (US, Canada, Australia, New Zealand, UK); ignore the circulating '51% of US replies' figure, which has no primary source.

Yelp signed a licensing deal on July 23, 2026 giving ChatGPT real-time access to its reviews, ratings, photos and business information, reported as about 330 million cumulative reviews and 8 million-plus business listings. Yelp branding and links appear when OpenAI uses the content, though OpenAI controls presentation; a 'Request a Quote' feature is coming to ChatGPT local-services searches. The deal is non-exclusive and financial terms were undisclosed. Yelp shares rose about 8% on the news.

ChatGPT's historical weakness on local and 'near me' queries just got patched, so local and multi-location brands have a new ChatGPT-visible surface where Yelp review authority is the ranking substrate. It is also the clearest signal yet that engines will pay to license third-party review corpora rather than build them, which reinforces the off-domain citation thesis. Do not tell clients to optimize their Yelp page and assume they win ChatGPT; no citation-share or click data exists yet, so measure it first.

Profound published a reverse-engineering study of ChatGPT Shopping on July 23, 2026, covering 201,137 unique prompt runs and 812,190 product cards collected June 18 to 25, 2026. 87.3% of product cards (707,752) came from web crawl and only 12.7% (102,523) from direct merchant feeds; Reddit supplied about one third of all shopping citations; ChatGPT runs a separate product query fanout from its text fanout. Rank-1 cards versus rank-4-and-below cards showed GPT tag presence up 143.97% and median review count up 123.58%, with rank-1 cards averaging 787 reviews versus 352.

Feed enrollment is not the path onto the shelf; ordinary crawlable web presence is, and the feed matters for offer accuracy rather than selection. Treat review acquisition as GEO work, not just conversion work, since review volume is a first-class ranking input. Audit the product carousel separately from the text answer because they are selected by different machinery. Ratios are solid; absolute levels are window-bound and pre-date GPT-5.6.

Notified, in partnership with PRWeek, published an AI Search report on July 22, 2026 analyzing more than 250,000 press releases and nearly 250 million AI citations. A sub-study of 8,000 GlobeNewswire-distributed releases from June 2026 found 99.3% were cited by ChatGPT or Claude, with an average time to first AI citation of about 8 hours and most citations within the first 24 hours. The report proposes a SOAR framework (Structure, Originality, Authority, Recency) for what AI systems select and cite.

Syndicated, well-structured announcements can be cited by ChatGPT or Claude within a business day, which makes the press-release channel a near-instant citation surface. Attribute this as a PR-wire vendor report measured on its own distributed releases: the syndication footprint is likely why citation is near-universal, so do not tell clients any press release gets cited. The 8-hour ingestion figure is the novel number, and the recency dependence argues for continuous distribution plus measurement over one-off releases.

Google made Gemini 3.6 Flash, Gemini 3.5 Flash-Lite and Gemini 3.5 Flash Cyber generally available on July 21, 2026, confirmed on the Gemini API changelog and the Google blog. Gemini 3.6 Flash uses 17% fewer output tokens than 3.5 Flash, advances its knowledge cutoff from January 2025 to March 2026, and is priced at $1.50 / $7.50 per million tokens; 3.5 Flash-Lite is priced at $0.30 / $2.50 per million and is rolling out in Google Search for agentic search. Gemini 3.5 Pro did not ship and Google teased Gemini 4 instead; temperature, top_p and top_k sampling parameters were deprecated.

Google's answer surface has been governed by 3.5 Flash since May 19, so any Gemini or AI Overviews baseline captured before July 21 is potentially stale; re-run Gemini-surface audits once the Search rollout lands. Google has not disclosed what share of AI Overviews or AI Mode traffic runs on Flash-Lite or a completion date, so frame any re-audit as a dated re-measurement, not a causal test. Do not write that Gemini 3.5 Pro launched.

Conductor published its AI Brand Recommendation Study on July 17, 2026, built on 14,000 API calls across 10 industries, 7 intent types, 4 LLMs (ChatGPT, Perplexity, Claude, Gemini) and 5 simulated personas at 50 stateless runs per cell, measuring consistency as Jaccard brand overlap. Overlap ranged from 40% on Purchase intent to 63% on Comparison intent, and lead-brand stability from 30% on Education to 91% on Comparison. Perplexity was the most consistent engine and Gemini the least, returning 9.2 brands per response versus ChatGPT's 5; Cloud Computing was the least consistent industry at 36%. Conductor explicitly positions the result against the January 2026 SparkToro finding that recommendations are essentially random.

Query intent, not industry, predicts how stable AI recommendations are, so a blended visibility score across mixed intents is arithmetic nonsense; report per intent. This validates running MOFU and BOFU prompts only, since Comparison and Recommendation intents are the measurable ones and Education is the noisy band. Caveats: vendor research, simulated personas, API calls rather than consumer interfaces, and Jaccard penalizes Gemini's longer lists.

Conductor with BOL Agency published the first direct AI-citation-to-organic-traffic attribution study on July 17, 2026: 87 domains (filtered from 184) tracked weekly from September 2025 through February 2026, giving 26 data points per domain and 1,300-plus paired observations, with Spearman correlations at four lag windows. Non-branded organic traffic correlated more strongly with AI citations than branded traffic at every lag, peaking at the one-week lag (non-branded 0.2494 versus branded 0.2210), all eight correlations at p below 0.0001. AI citations explained only about 5% to 6% of the variance in non-branded organic traffic.

This is the first defensible answer to whether AI citations move traffic, and the honest headline is the modest effect size, so quote the 5% to 6% first, not the correlation. Because citations move non-branded traffic hardest, AI visibility belongs in the category-discovery and demand-gen budget rather than brand defense. Correlation not causation, and the window pre-dates GPT-5.5 and GPT-5.6 entirely, so the relationship needs re-measuring at current model generations.

Google made Grounding with Parallel Web Search generally available on the Gemini Enterprise Agent Platform on July 16, 2026, subscribed via Google Cloud Marketplace and configured in Agent Studio. Its advertised differentiator over Google's own grounding is 'expanded licensing flexibility': developers may 'extract and cache web data' and 'post-process results using other LLMs'. Google's own Gemini API Additional Terms of Service, effective March 23, 2026, forbid using Links from Grounded Results 'to build an index' and forbid caching, analysing or training on Grounded Results.

The official Gemini grounding route cannot legally be used to build a citation index, while Google sells shelf space to a third party whose selling point is that you may. Ask any vendor reporting Gemini citation shares which interface produced the numbers and whether it was permitted to keep them; observing the consumer Gemini app is a different activity, so name no vendor as in breach. If you build your own Gemini measurement, use a licensed route.

A critical survey of generative engine optimization, arxiv 2607.14035 by Olivier Martinez, was submitted July 15, 2026, reviewing 45 studies published between November 2023 and July 2026. It argues GEO is a stochastic, partially observable pipeline rather than a single ranking task, and finds the original GEO paper's gains were conditional on a source already being present in a fixed context, that topical relevance and context position are the most reproducible levers, that generic heuristics transfer poorly, that competition erodes individual gains, that citation-oriented rewrites can impair retrieval, and that no reviewed technique demonstrated stable cross-platform causal effects on organic discoverability.

Use this as the independent academic reference when a client asks whether GEO works: the levers that reproduce are topical relevance and context position, and everything else is noisy and context-dependent. Drop generic GEO checklists from deliverables and measure with confidence intervals, since run-to-run variability is documented. Attribute it as a single-author literature synthesis about the state of the evidence, not fresh measurement.

arxiv 2607.13304 by Dmitrij Zatuchin, submitted July 14, 2026, decomposed the variance in LLM brand recommendations across 12,933 responses, 20 Central and Eastern European brands, 8 languages and 3 models (GPT-5.2, Gemini 3 Flash, Perplexity). Query language explained 26.5% of variance while brand identity alone explained just 1.5% (resampling 34.8%, brand-in-context 29.6%, brand-by-language 8.6%). Brand-ranking reliability from a single answer was about 0.01, rising to only about 0.36 with a fully-crossed design, and adding a fifth repeat of the same prompt cut relative-error variance by only about 0.0003.

A single AI answer carries almost no brand-discriminating signal, so single-prompt visibility scores are noise and the industry habit of running five prompts and averaging barely helps. Reliability comes from spreading measurement across languages and models, not from repeating one prompt. Cite the variance structure rather than the absolute numbers, since the model mix pre-dates GPT-5.6 and Gemini 3.5 Flash.

arxiv 2607.10341, 'From Stochastic to Stable' by Ronald Sielinski (IQRush), submitted July 11, 2026, proposes a sequential convergence framework for deciding when an AI-visibility measurement has enough answers. Across 30 platform-topic combinations on Gemini, SearchGPT and Perplexity, stable rankings required between 33 and 94 answers depending on platform and topic, 3 of 30 tests never stabilized even after 125 questions, and one topic showed a 17% no-citation miss rate. A separate University of St. Gallen validation reproduced the approach.

When a prospect asks how many prompts you run, the defensible answer is now 33 to 94 answers per platform-topic determined by convergence, not a round number; most competitors run 5 to 20. Disclose that about 10% of measurement targets are structurally unmeasurable at practical sample sizes. The method transfers but the thresholds do not; queries were ChatGPT-generated and the paper is an unreviewed preprint.

On July 10, 2026, the day after OpenAI released the GPT-5.6 family, feed-integrated retrieval in ChatGPT Shopping went from 8.26% to 61.54% of product recommendations in a single day and overtook web-search retrieval as the primary source, according to Profound's analysis of 1,757,723 daily prompt runs in July and 97,725 monthly prompt runs. Top-10 merchant share rose from 22.5% to 41.8% (July 7 to 9 versus July 10 to 12), unique merchants fell from 13,524 to 10,607, and Shopify accounts for about 35% of all feed-integrated retrieval. Profound states that brands will not be considered for 65% of product recommendations in Shopping unless integrated.

Any ecommerce client whose ChatGPT Shopping visibility moved in mid-July has a platform explanation, not a content one. Product feed integration (Shopify is the practical route for smaller retailers) is now the gate to the majority of Shopping recommendations; the 65% figure is Profound's framing of its own data, so attribute it. Concentration nearly doubled at the same time, so late feed adopters compete inside a smaller merchant set.

OpenAI moved GPT-5.6 from a government-gated limited preview of about 20 organizations to broad general availability across ChatGPT, the API and Codex on July 9, 2026, about 12 days after the June 26 preview was held at the request of the White House Office of the National Cyber Director. Three variants ship: Sol (flagship), Terra (advertised at roughly GPT-5.5 quality for about 2x lower cost) and Luna (lowest cost). Terra became the Free and Go default answer model; press-reported per-million-token pricing was Sol $5 / $30, Terra $2.50 / $15, Luna $1 / $6, and the bare gpt-5.6 API alias routes to Sol.

Most consumer ChatGPT answers now run on the 5.6 family, the trigger condition for measurable brand-recommendation shifts of the kind GPT-5.5 Instant caused in late June. Re-baseline ChatGPT visibility now and segment by tier, because Sol, Terra and Luna can return different sources for the same prompt and an aggregate mention rate can look stable while presence in one tier erodes. Note that as of the end of July no vendor had published a measured, tier-segmented GPT-5.6 brand study.

Profound launched Aim on July 2, 2026, described as the first background agent for marketing. It continuously monitors a brand's citations, sentiment, accuracy, prompt volumes and agentic traffic across ChatGPT and Claude, and when it detects a meaningful shift it explains what changed, writes a memo and auto-creates a Project with prioritized tasks routed to Profound Agents while keeping marketers in approval control. Named customers at launch included Figma, Walmart, Ramp, MongoDB, Chime and U.S. Bank; pricing tiers were later listed at $99 per month Starter and $399 per month Growth (billed yearly), plus custom Enterprise.

The leading measurement vendor is completing the move from measure to act, which commoditizes the task-generation and low-level execution layer. A background agent can hand you a prioritized to-do list and draft the pages; it cannot earn the third-party credibility that makes a model recommend you by default, so anchor managed-service positioning on strategy and earned authority. Expect the 'we tell you what to do' ground to get crowded.

Profound published 'The AI Mention Effect' on July 1, 2026, a causal study built on more than 2 million AI conversations paired with real browsing behavior (January to June 2026, US, across ChatGPT, Gemini and Google AI Overviews). A brand mentioned in an AI answer saw its site-visit rate rise to 1.5x to 2.5x its forecast baseline over the following 7 days: Gemini 5.42% versus 2.21% (about 2.5x), Google AI Overviews 7.79% versus 4.83% (about 1.6x), ChatGPT 6.39% versus 4.33% (about 1.5x). 20.5% of downstream visits landed within 1 hour and 42% within 24 hours, and more than 97% of those visits carried no UTM parameter even after ChatGPT's May link changes raised direct attribution from 1.0% to 2.47%.

An AI mention is not a vanity metric; it measurably lifts baseline traffic for a week, and the biggest per-mention payoff is on Gemini, an engine most brands under-monitor. Standard analytics sees only about 2.5% of the true impact, so dedicated AI-visibility monitoring, not Google Analytics, is how you know whether you are winning. Measure over a 7-day window, since a same-session view misses more than half the effect; attribute as a vendor study and quote the range.

Anthropic redeployed Claude Fable 5 globally on July 1, 2026 after a June 12 access suspension, priced at $10 / $50 per million input/output tokens. Claude Sonnet 5 had shipped on June 30, 2026 with native 1M-token context at introductory pricing of $2 / $10 per million tokens through August 31, and was set as the default model on July 1. Neither release changed Claude's web-search or inline-citation surface.

The Claude model cohort turned over first among the three major labs in July, but the citation mechanics stayed the same, so no Claude-specific re-baseline is forced by these ships alone. The lesson is cadence: a client who checked AI visibility in June was already looking at a stale picture on one engine, with GPT-5.6 and the Gemini Flash update following within weeks. Keep Claude in the tracked set even though its citation footprint is small relative to ChatGPT.

Cloudflare announced on July 1, 2026 that from September 15, 2026, for all new domains onboarding to Cloudflare, crawlers classified as Training or Agent are blocked by default on pages that display ads while Search remains allowed, with an opt-out in Security settings. On the same date multi-purpose crawlers that combine Search with Training, which Cloudflare names as Googlebot, Applebot and BingBot, are evaluated under 'the most restrictive applicable rules', so customers who selected block-Training, including via the legacy one-click 'Block AI bots' service, would block them. The switches to manage Search, Agent and Training traffic separately went to all customers including the Free tier on July 1. Pay Per Crawl is replaced by Pay Per Use, paying publishers when content appears in an AI answer rather than when it is fetched, with Ceramic.ai, You.com, beehiiv, Condé Nast and Patreon as partners. Cloudflare says over 50% of AI crawler traffic is spent re-fetching unchanged pages, that 52% of crawler requests were for AI training as of June 2026 (up from 22% in Spring 2025) and that mixed-use crawlers are over 36% of activity.

The September 15 default barely affects existing sites; the exposure is a Training block or 'Block AI bots' toggle ticked in 2024 or 2025 that could take out Googlebot, Applebot and BingBot. Check every property and set the opt-out in Security settings if needed. Pay Per Use makes citation a billable event at the CDN layer, so 'did we get cited' starts to look like a revenue line, not only a marketing metric. Cloudflare's September 15 'Disallow AI Training' release later gave Apple, Google and Microsoft a way to honour the preference and stay allowed for search.

Profound formally launched the Profound Index at Zero Click New York on June 29, 2026 (soft-previewed June 16), a public AI-search visibility benchmark built on 1.5 billion-plus real user prompts across 50-plus industries and all major LLMs, ranking brands on six metrics: Visibility Score, Share of Voice, Mention Position, Citation Share, Co-citation Share, and Co-mention Share. Headline finding: only about 19% of ChatGPT's answers overlap with Google's top results. Brand swings tied to the GPT-5.5 update included IKEA up 19% on ChatGPT but down 11% across AI platforms overall, Travelocity, Orbitz, and Priceline each down 35% to 56% on ChatGPT, and Ambetter and Humana each down 50%-plus while Aetna gained.

Stop treating Google rankings as a proxy for AI visibility; roughly four in five ChatGPT answers draw on sources outside Google's top results, and Brandlight separately reports Google-to-AI source overlap falling from about 70% to below 20%. Adopt the six-metric vocabulary in client reporting so 'mentioned' is understood as step one, not the finish line. Use the free preview to baseline each client's category, and monitor continuously, since a single model update moved major brands by 35% to 56%.

Semrush released its Expanded 2026 AI Visibility Index on June 26, 2026, analyzing 126 million real US AI search prompts from January to April 2026 across 22 industries and 4 AI platforms, scaled up from a 2,500-prompt pilot, on a global prompt database of 261 million. Category concentration ranges from winner-take-most (top-3 brands hold 82.9% of visibility in News and Media, 76.9% in Consumer Electronics) to distributed (41.4% in Finance, 42.2% in Industrial), and only 36 global brands held top-100 visibility across all 4 platforms in every month. Semrush also reports 45% of marketing leaders cannot accurately measure their brand's AI-answer visibility and only 9% have tools to track all relevant metrics across platforms.

Check a client's category concentration in the live Index before setting targets: distributed categories like Finance and Industrial are genuinely contestable, while winner-take-most categories need sustained differentiated effort, not a tool toggle. Use the 45% cannot-measure and 9% fully-tooled gap to frame measurement as the first deliverable. Attribute the figures to Semrush and pull the specific per-industry split before quoting a percentage to a client.

OpenAI updated GPT-5.5 Instant, the most-used model in ChatGPT, on June 24, 2026 to be better at decisions, advice, planning, research, and shopping. Per OpenAI's update notes it reads intent more carefully, handles requests that stack several constraints, adapts when users push back, makes sharper use of location for shopping and local queries, and spells out why a pick fits the stated need, rewarding pages that clearly say who they are for and back it up while sidelining thin, generic content.

Audit every commercial and comparison page for an explicit audience statement (who this is for, and for whom it is not) backed by checkable evidence, since the model now explains its picks and drops generic pages. Expect recommendation share to move without warning after model updates; re-baseline AI visibility measurements whenever a default model changes and schedule a re-check of top prompts after each update.

Profound published a ChatGPT Shopping Deep Dive on June 24, 2026 analyzing about 1 million ChatGPT shopping offers over 30 days in June, comparing 406,639 top-ranked citations against 659,868 bottom-ranked ones. Product feeds are only about 20% of retrieval sources but about 99.9% of them appear as the first product offer, feed reliance has grown about 15 times this year from near zero in November 2025 to about 20%, and product detail pages account for 88.29% of all instances but underperform in top rankings. Even merchants with feed integration still derive 75.81% of their offers from PDPs.

For any client selling through ChatGPT Shopping, submit a product feed; feed-supplied metadata (checkout image, brand, merchant subtitle, Best Price tag) is what wins the first slot. Keep optimizing PDPs anyway, because three quarters of offers still come from them. For B2B SaaS, port the structural lesson: provider-supplied structured pricing, spec, and comparison data is being weighted over crawl-discovered content, and the shift is fast.

At ChatGPT Ads' Cannes Lions debut on June 23, 2026, OpenAI head of global ads solutions David Dugan disclosed the first engagement data for the product: about 20% of ChatGPT queries show direct purchasing intent, ad cross-out (dismissal) rates have dropped 50% since the February launch, ChatGPT has about 1 billion users, and 2,000-plus brands now advertise through Criteo. Ads sit in a sponsored box beneath the organic response and perform better when the call to action is framed as a direct answer to the user's question. Markets held at 7 with Brazil and Mexico still pending.

Treat the 50% drop in dismissals as the first sign the format is settling in, but note every metric is consumer and aggregate: there is still no firmographic targeting and no measured B2B CPA, and Adweek reported on June 26 a ChatGPT ad CTR as low as about 0.91% versus about 6.4% for Google Search. Keep B2B spend to a fast-loop lead-gen pilot at most, and write any ad creative as a direct answer, not a product listing.

Peec AI published 'Translated Reddit Is Winning AI Citations' on June 22, 2026, analyzing 64.77 million Reddit citations across 20 countries and 4 AI engines (ChatGPT, Google AI Overview, Google AI Mode, Gemini) from March 1 to June 10, 2026, of which 3.7 million carried Reddit's ?tl= machine-translation parameter. In non-English markets, 40% to 73% of Reddit citations in Google's AI answers point to machine-translated English threads (below 1% in English-speaking markets), with Google AI Overview at 52% translated Reddit in Germany and above 70% in Sweden and Norway, while ChatGPT cut translated-Reddit citations from 6.14% in April to 0.30% in June.

Instrument AI citations by engine and by market separately; Google is leaning hard into translated English Reddit while ChatGPT is pulling back, so one Reddit strategy is wrong. For clients with EU or other non-English exposure, genuine English-language Reddit presence now competes directly with local-language content in Google's AI answers. Treat the per-market splits as directional vendor data and validate before hard-coding them in a deck.

OpenAI confirmed sponsored content live in ChatGPT for users in South Korea on June 19, 2026, the 7th market after the US, Canada, Australia, New Zealand, the UK, and Japan. Ads run on Free and Go tiers only, are labeled 'sponsored content' with a hide-ad option, and advertisers get access only to aggregated performance metrics such as impressions and clicks. No Korea CPM was disclosed, and Brazil and Mexico remain pending.

Two market launches in two days confirm the geographic cadence, but the Korea fine print (aggregated impressions and clicks only) is fresh confirmation that there is still no firmographic targeting and no B2B-grade reporting. Hold the lead-gen pilot stance for fast-loop SaaS and keep enterprise ABM budgets out until a measured B2B CPA exists.

ChatGPT Ads launched to users in Japan on June 18, 2026, embedded in-conversation on Free and Go tiers, the first non-English-language market and the first Asian market after the US, Australia, New Zealand, and the UK. Reported CPM was about 5,000 yen per 1,000 impressions (about $33), with Korea named next and Brazil and Mexico not yet live.

Use the first non-English go-live as evidence the channel is scaling fast and strategically funded, while keeping the caveat that it remains shopping, retail, and travel skewed and consumer-tier only. Do not price a B2B test off the Japan CPM; wait for a measured B2B CPA before recommending anything beyond a small pilot.

arXiv 2606.20065, 'Generative Engine Optimization at Scale: Measuring Brand Visibility Across AI Search Engines' (Pratyush Kumar, Ranqo), was submitted June 18, 2026, covering 100,000-plus prompt responses across 100-plus brands tracked March to May 2026. Global household names appear in 73% of relevant AI answers, mid-market and regional brands in 44%, and niche or small brands in just 11%; about 78% of citations point to corporate or owned websites and about 21% to best-of listicles; and sentiment framing fluctuates about 6.7 times more often than mention frequency.

Set client expectations off the 73/44/11 maturity gradient: a niche B2B brand starting near 11% baseline visibility has a different ramp than a category leader. Treat listicle placement as a measurable earned-media lever worth roughly one in five citations, not a nice-to-have. Monitor how a brand is framed, not just whether it is mentioned, since framing is far less stable than presence. Cite as a vendor-affiliated preprint, directional rather than peer-reviewed.

Google's AI-search opt-out became enforceable on June 17, 2026. Opted-out sites stop appearing in AI Overviews, AI Mode, and AI Overviews in Discover, while organic rankings are unaffected. It is the first binding conduct requirement on Google under the UK DMCC 2024 and the first regulator-enforced AI-search opt-out anywhere, limited to a UK subset of site owners, with Gemini excluded from scope.

Advise almost all B2B clients NOT to opt out, since opting out forfeits the impression entirely. For any UK-subset client even considering it, pull an AI-impressions baseline from the GSC Gen-AI report first. Note the Gemini exclusion means you still cannot regulate your way out of multi-engine visibility.

Google launched Search agents on June 12, 2026, starting with information agents in AI Mode for Google AI Ultra subscribers. The agents monitor news, blogs, social, finance, shopping, and sports data around the clock and push updates with links. Follow-up from an AI Overview into a conversational AI Mode session also went live worldwide.

Treat persistent citation as the new bar, since always-on agents keep re-pulling sources, so freshness and sustained authority matter more than a one-time citation. Optimize AI Overview presence as a funnel into multi-turn AI Mode sessions.

Fractl and Search Engine Land published a study on June 15, 2026 showing consumer trust in AI search collapsed from 82% in 2025 to 54% in 2026, a 28-point drop. Buyers now check an average of 2.4 platforms before validating a purchase, trust splits Google 39% / Reddit 15% / AI tools 14%, and a cross-referenced Ahrefs 75,000-brand study found branded web mentions and YouTube impressions correlate most with AI visibility (0.50 to 0.74 Spearman) while backlinks and ad spend correlate least (under 0.30).

Lead with accuracy and credibility, not volume, and feed engines checkable sourced facts. Prioritize branded mentions and YouTube presence over backlink or ad-spend tactics, and instrument across multiple engines since buyers triangulate 2.4 platforms.

OpenAI launched ChatGPT Ads product-feed campaigns on June 12, 2026, letting advertisers upload product feeds of up to 2 million items to power ad placements, a move toward shopping and commerce-style ads.

Apply B2B caution since this is retail/shopping-skewed, but treat machine-readable product feeds and structured catalogs as the price of entry for AI commerce surfaces.

Adobe Brand Visibility, combining Adobe LLM Optimizer with Semrush AI Optimization data, surfaced drawing on 289M-plus AI search prompts, described as the largest global database of its kind, with reported AI traffic to US retail up 1,324% (Oct 2024 to May 2026) and travel up 2,215%.

Use the post-acquisition enterprise-suite consolidation as a qualifying point: these suites are blind to engines they do not instrument. Position vendor-neutral multi-engine triangulation as the objective alternative.

Google Analytics 4's native AI Assistant default channel reached broad availability around June 7, 2026 (channel introduced May 13), auto-recognizing traffic from ChatGPT, Gemini, Claude, Copilot, Grok, and DeepSeek without manual UTM setup. It still routes Perplexity to Referral, counts AI Overviews as Organic Search, and misses referrer-less AI traffic.

Enable the GA4 AI Assistant channel now to start a baseline, but pair it with server-side and own-funnel tracking since it undercounts and is blind to Perplexity and AI Overviews.

Optimizely launched a full enterprise AEO platform on June 10, 2026 via an exclusive partnership with Conductor, embedding Conductor's SEO/GEO/AEO intelligence plus a new Agent Visibility Analytics module that uses log-level data to classify AI requests by intent (retrieval, indexing, training) and three turnkey agents (AEO Gap Finder, Competitive AI Share-of-Voice, AI Brand Visibility Report).

Build fluency in the agent-behavior and bot-log axis, not just citation share, before clients ask. Pitch reading both citation outcomes and server-log agent intent as a managed-agency advantage over single-axis dashboards.

ChatGPT Ads was confirmed to ship native B2B conversion events in the oaiq() pixel/SDK, including lead_created (labeled best for lead gen, agencies, B2B), appointment_scheduled (service businesses, consultants), and trial_started (SaaS), closing the long-standing no-native-B2B-event measurement gap. By June 11, lead_created and appointment_scheduled appeared available for CPA bid-optimization too, though targeting remains contextual-only with no firmographic/ABM audiences and zero measured B2B performance data.

Retire the old you-cannot-even-measure-a-B2B-demo line. Stand up the oaiq() pixel plus Conversions API, fire the correct B2B event, and run a capped $2-5K/mo CPA pilot for fast-loop lead-gen/SaaS measured against the client's own Google/Meta CPA. Advise enterprise ABM clients to wait until firmographic targeting ships.

SparkToro published a study on June 9, 2026 using Similarweb's US desktop and mobile clickstream panel from January through April 2026, finding that 68.01% of US Google searches now end without a click, up 7.56 percentage points from 60.45% in 2024. Of every 1,000 US Google searches, only 276 clicks now reach the open web, down from 374 in 2024. Google AI Overviews, appearing on more than 20% of all searches, were identified as the primary driver and found to cut click-through rates by nearly 60%.

With 276 of every 1,000 Google searches reaching the open web, the organic referral channel is structurally smaller than it was two years ago and the gap is widening. Treat citation inside the AI answer as the primary visibility unit for any query where an AI Overview is likely to appear. In budget conversations, use the 276-per-1,000 figure to show why optimizing for blue-link position alone is an increasingly incomplete brief. Track GSC Gen-AI impression share as a separate KPI from organic clicks, since the two now move on independent tracks and a page can rank without being cited.

GNW Consulting and Demand Metric published the 2026 State of GEO in B2B Marketing on June 3, 2026 (n=225 B2B marketing and revenue leaders). It found 92% experimenting with or operationalizing GEO, 78% of investors reporting measurable ROI, and 22% saying AI traffic exceeds 5% of site traffic, but 88% of SEO agencies claim GEO services with 37% loosely defined, and fewer than 15% of companies have a dedicated GEO owner.

Use the demand-rich, ownership-poor, GEO-washed-supply gap as the core managed-service pitch: position Cite as the defined owner of the function against a market where almost no one owns it and a third of agency offerings are vague.

First Page Sage published an AI-chatbot market-share estimate dated June 5, 2026 showing ChatGPT fell to 53.1% from about 61.8% the prior period, with Claude 21.1%, Gemini 13.1%, Copilot 8.7%, and Perplexity 2.7%. The figures are modeled estimates and the ChatGPT decline is the defensible finding.

Reinforce a multi-engine posture in client recommendations: with ChatGPT sliding from roughly 62% to 53%, single-engine optimize-for-ChatGPT advice is increasingly wrong. Treat the exact percentages as directional.

First Page Sage's ChatGPT Ads Conversion Rate report (last updated April 27, 2026) surfaced as the first independent, non-Criteo B2B benchmark, projecting B2B SaaS ChatGPT-Ads conversion at about 1.1%, below its own SEO benchmark of 2.1% and roughly even with Google Ads at 1.0%, across a 19-industry spread from 0.3% (Financial Services, about $1,500-plus CPA) to 6.0% (Higher Education, about $50 CPA). The numbers are modeled, not measured live-campaign actuals.

Counter the loud 2x-better-than-search retail claim with this independent B2B projection, and use the CPA-by-vertical table in should-we-run-ChatGPT-Ads advisories. Tell clients to instrument their own demo/trial conversion before trusting any headline multiple.

Search Engine Journal framed the citation-to-transaction shift around Google's announcement that Chrome auto-browse reaches Android phones starting at the end of June 2026 (Pixel 10 and Galaxy S26 first), and documented eight auto-browse failure modes (client-side rendering, cookie walls, unlabeled forms, div-based buttons, modal traps, CAPTCHAs, timeouts, sign-in walls) plus a silent-loss dynamic where failed agent transactions throw no analytics signal. A June 9 correction clarified the Android rollout is gated to AI Pro and Ultra subscribers, Android 12-plus, en-US, with confirmation before sensitive tasks.

Productize an agent-transaction-readiness audit: verify real button elements, semantic form labels, and a conversion path free of modal traps, CAPTCHAs, and sign-in walls so AI agents can complete bookings. Lead client conversations with the silent-loss risk since failed agent transactions generate no error or bounce.

ChatGPT Ads went live in the UK on June 6, 2026, the first market served outside the US, Australia, and New Zealand, with expansion to Japan, South Korea, Brazil, and Mexico slated next. OpenAI projects about $2.5B in 2026 ad revenue targeting $100B by 2030, with ads serving only logged-in adult Free and Go users in shopping, retail, and travel categories.

Flag to clients that paid AI placement is internationalizing and strategically funded, but it is consumer-tier and retail/travel-skewed with zero measured B2B CPA, so the fast-loop lead-gen pilot stance holds and enterprise ABM should wait.

OpenAI launched conversion-based billing inside the ChatGPT Ads Manager on June 5, 2026, replacing impression-only and click-only pricing with a pay-per-action model. Advertisers now pay when a user completes a defined action. The platform ships no native book-a-demo conversion event, so B2B advertisers must configure a custom action.

Before activating the new billing model, define a conversion event that reflects real purchase intent. B2B SaaS operators cannot use a purchase pixel and must pick a proxy action such as a pricing page visit or demo form submission. Retailers tied to checkout events can activate more directly. Verify that your conversion math closes at ChatGPT audience costs before committing budget.

Burson published 'The Credibility Paradox' on June 4, 2026, the first large study to separate AI visibility from AI believability, conducted in partnership with Profound across 85 companies on 7 AI answer platforms and generating 55,000-plus believability forecasts. Fact-based claims tied to innovation, products, and workplace culture scored as the most believable across every industry studied. Claims tied to leadership, governance, and citizenship scored as the least believable. Business decision-makers rated AI-generated answers 10% more believable on average than the general population.

Citation share is a partial metric: an AI answer that surfaces your brand with a claim your audience does not believe is worse than no mention, and business decision-makers are the audience most likely to act on what AI says. Build every AEO page around sourced, checkable, fact-based claims such as product specifications, measurable outcomes, and dated case-study results. Retire soft reputation narratives about leadership vision or citizenship from the pages your citation program depends on, because those are exactly the claim types that land with the lowest believability among the buyers who trust AI answers most.

Anthropic released Claude Opus 4.8 and announced a $65 billion Series H at a $965 billion post-money valuation on May 28, 2026, the highest disclosed valuation for a private AI lab. Published benchmarks put Opus 4.8 at 88.6% on SWE-bench Verified (up from 80.8% for Opus 4.6) with 0% hallucination on flawed data and the lowest incorrect rate of six models tested on every hallucination benchmark, while fast-mode pricing fell to $10 per million input and $50 per million output tokens (from $30 and $150) and standard pricing held at $5 and $25. Anthropic also said Mythos-class models would reach all customers in the coming weeks.

Treat Claude as a broad-enterprise citation surface, not a regulated-industry niche: the price objection is gone and the accuracy story is the strongest in the market. Flag any Claude baselines captured before May 28 as pre-Opus-4.8 and expect the visible delta to be more abstention rather than new citation behavior, so do not over-read short-term Claude citation swings during the 30 to 60 day default transition. Keep Mythos on the watch list as a second re-baselining trigger.

CNN filed a copyright lawsuit against Perplexity on May 28, 2026 in the U.S. District Court for the Southern District of New York, the first television network to sue an AI company. The complaint alleges Perplexity scraped more than 17,000 CNN stories and falsely advertised a CNN relationship through its Comet Plus tier, after 2025 licensing talks failed. Perplexity CCO Jesse Dwyer responded: "You can't copyright facts."

Perplexity's publisher citation pool now carries litigation risk through the second half of 2026, so hedge Perplexity-priority budgets toward ChatGPT, Claude, and Gemini for clients whose AEO plan leans on video, webinar, or executive-interview content. Use the case in procurement conversations with risk-averse legal teams, and watch the SDNY docket for motion-to-dismiss timing since the same court is hearing NYT v OpenAI.

Peec AI published a 500,000-prompt study by Tomek Rudzki on May 28, 2026 finding AI Overviews appeared on 86.7% of searches in April 2026, up from 56.9% in April 2025, and on 88.5% of decision-stage commercial queries. Appearance rates ran 91.4% in the US, 91.7% in the UK, 75.6% in Germany, 76% on average across the EU, and 0% in France, and rose with query length from 64.6% for 2-word queries to 89.1% for 11 to 15 word queries. Prompts were pulled April 13 to 20, 2026 from Peec's business-oriented prompt dataset.

Stop treating AI Overviews as informational-only: Peec frames it at 2.5 billion monthly users versus 1 billion for AI Mode, and it triggers on nearly nine of ten commercial queries. Put AI Overviews citation share ahead of ChatGPT for any client with general search-intent traffic, build content for long, multi-clause questions, and drop AIO from the plan for France-targeting programs.

Google began rolling out the May 2026 Core Update on May 21, 2026, its second core update of the year and fourth confirmed ranking update, and completed it on June 2 after roughly 12 days. Coverage described it as larger and more volatile than the March 2026 update, with thin AI-generated content, YMYL, and aggregator or affiliate-review properties hit hardest, and click declines decoupling from ranking declines as AI Mode and AI Overviews absorbed traffic on queries where rankings held.

Do not make reactive content changes off mid-rollout data; take the first clean Search Console read from about June 9. Reset client expectations that holding rank no longer means holding traffic and report citation share as its own KPI. The pattern favors first-hand, original-research content over AI-rewritten aggregation, so keep the human editing pass on everything you publish for clients.

At Google I/O on May 19, 2026 Google disclosed AI Mode has surpassed 1 billion monthly active users, with queries more than doubling every quarter since its May 2025 launch and the average AI Mode search running 3 times longer than a traditional Google search. Gemini 3.5 Flash became the default model in AI Mode globally the same day, and AI Overviews gained conversational follow-ups that flow into AI Mode with context preserved. Follow-on I/O materials put AI Overviews at more than 2.5 billion monthly users.

Move AI Mode into the must-measure tier for any B2B program serving general search intent; it is now a 1 billion user surface growing faster than any other AI surface by query volume. Flag every AI Mode baseline captured before May 19 as pre-3.5 and re-baseline in early June, pair AI Overview citation share with AI Mode follow-up citation share in reporting, and write for long, nested, multi-clause questions rather than short answer-box copy.

Google shipped Gemini 3.5 Flash at I/O on May 19, 2026, not the Gemini 3.2 Flash that leaks predicted, claiming it surpasses Gemini 3.1 Pro on coding, agentic, and multimodal benchmarks with 4 times faster output tokens per second than other frontier models, and said Gemini 3.5 Pro would follow next month. Per-token pricing published the next day was $1.50 per million input and $9.00 per million output tokens, roughly 3 to 5 times Gemini 2.5 Flash's $0.30 and $2.50 and about 6 times the leaked $0.25 and $2.00 figures. Google also cut its AI Ultra plan from $250 to $100 per month.

Retire any cost-floor assumption that Gemini-grounded tooling will undercut ChatGPT- or Claude-grounded tooling: Flash no longer means cheap, and expect GEO vendors that migrate to 3.5 Flash to pass through per-query cost increases within 60 to 90 days. Schedule re-baselining of Gemini and AI Mode citation data for the week of June 1 to 7, and watch AI Ultra adoption as a driver of consumer Gemini share.

Google published its first official guide to optimizing for generative AI features in Google Search on May 15, 2026, a roughly 2,000-word document announced by John Mueller on the Search Central Blog under a new 'Generative AI fundamentals' section. It states sites do not need new machine-readable files such as llms.txt, do not need to chunk content into small pieces, and do not need to write in a special way for AI, frames AEO and GEO as 'still SEO', and includes Google's first guidance on optimizing for AI agents.

Re-audit every service line that leans on llms.txt evangelism, chunking-for-retrieval, or AI-specific writing templates against Google's stated position, and reframe llms.txt as an agentic-browsing and IDE-agent readiness tactic rather than a Google AI search tactic. Use the guide's existence as category validation in client conversations while keeping the emphasis on unique, non-commodity content. Note that Chrome Lighthouse 13.3 (May 7) audits llms.txt by default, so Google's own products disagree.

Peec AI published 'The Listicle Rank Effect' on May 14, 2026, analyzing 5.7 million data points across nearly 200,000 AI responses on 8 engines and three markets from September 2025 to March 2026. Ranking #1 in a cited listicle lifted mention probability by 16.5 percentage points in B2B SaaS and 13.4 points in Emerging MarTech, and moved brands 1.80 positions earlier in US Finance answers, 1.17 earlier in B2B SaaS, and 0.82 earlier in MarTech. ChatGPT, GPT-5 Search, and Copilot showed larger lifts from top listicle placement than Google AI Overviews and AI Mode.

Pursue a handful of high-rank placements in listicles the engines actually cite rather than volume; the study's own framing is that five strong placements outperform fifty weak ones. Bind listicle work to ChatGPT-grounded query intent where the lift is largest, and present the effect as observational association, not causal proof.

Anthropic launched Claude For Legal on May 12, 2026 with 12 practice-area plugins, 20-plus legal tech integrations including Thomson Reuters Westlaw, LexisNexis, Harvey, DocuSign, and Box, a cross-app Microsoft 365 agent, and named launch customers Freshfields, Quinn Emanuel, Holland & Knight, and Crosby Legal, at no separate price for paying Claude customers. Anthropic disclosed that legal became the #1 power-user job function in Claude Cowork at over 3 times the usage of any other function. The same day SAP made Claude the primary reasoning and agentic capability across its Joule-powered AI portfolio and Thomson Reuters connected Claude to 1.9 billion Westlaw and Practical Law documents via MCP; Claude For Small Business followed on May 13.

Treat vertical density, not horizontal coverage, as the planning unit: Claude is now embedded in Big Law matters and across SAP shops, so buyers in legal and SAP-centric procurement will research vendors inside a Claude-grounded surface. Add a legal-grounded Claude surface to tracking for legal-tech clients, and for enterprise clients selling into SAP accounts track Joule plus Claude as its own citation surface.

Trustpilot published a study on May 12, 2026 analyzing 800,000-plus AI responses across 15,000-plus prompts on ChatGPT, Gemini, Perplexity, and Google AI Mode. Brands with no Trustpilot profile were cited in 1% of AI answers, brands with a profile in 53.5%, and brands with 80 or more reviews and active engagement in 75.3%. Review and trust platforms ranked as the #2 cited source type at 14% of all citations in the sample.

Audit every client's Trustpilot, G2, Capterra, and TrustRadius presence within 30 days against an 80-review, actively-managed threshold, since the jump from profile-exists to actively-managed is where roughly 22 percentage points of citation lift sits. Position managed review-pipeline work as a differentiator that monitoring-only tools do not deliver.

OpenAI quietly shipped a branded-link change on May 7, 2026 that embeds brand homepage URLs inline as clickable hyperlinks in ChatGPT responses, with no announcement. Profound's monitored-brand data, published May 19, showed OpenAI referrals to brand sites up 60 to 65% overnight and sustained, homepage share of OpenAI referrals rising from about 4% to about 24%, and the URL embedding rate in responses rising from about 4.5% to 20 to 24%. B2B software and SaaS daily referrals rose over 200%, Financial Services rose 60%, and E-commerce and Retail were essentially flat.

Retire the objection that AI cites you but nobody clicks, at least for ChatGPT, and treat brand mentions as a referral channel with B2B SaaS the biggest winner. Report a dual KPI of citation share plus branded-link click-through, make sure the homepage the link lands on is built for a first-touch buyer, and expect e-commerce clients to see little change because product flows route through shopping surfaces.

Otterly published 'The URL AI Citation Study 2026' on May 7, 2026, analyzing 1,028,959 unique URLs and 1,932,200 citation instances over a 24-hour window across ChatGPT, Google AI Overviews, Google AI Mode, Perplexity, Gemini, and Microsoft Copilot. URL length, domain length, path depth, and hyphen count all showed near-zero correlation with citations (r between -0.025 and +0.039), while guide pages averaged 2.7 citations versus 1.5 for pricing pages, clean URLs earned 24% more citations than URLs with query strings (2.1 versus 1.6), and the top 15.8% of URLs captured 50% of all citations.

Stop optimizing slugs and start shipping guides; page type is the lever, not URL structure. Run a shared-link hygiene pass so tracking parameters are not fragmenting one canonical page into low-citation variants, and concentrate editorial effort on a few flagship pages rather than a wide spread of thin posts, since citations follow a power law.

Muck Rack published the May 2026 edition of Generative Pulse on May 7, 2026, its third quarterly study, analyzing 25 million-plus links across ChatGPT, Claude, and Gemini in 17 industries. Earned media held at 84% of all AI citations, paid and advertorial content at 0.3%, and journalism at 27%. ChatGPT cited in 96% of responses with about 5 sources each, Gemini in 82% with about 8, and Claude in 55% with about 13, and press releases appeared 3.5 times more often in industry-trend responses than in best-of queries.

Reframe PR retainers as AI visibility investment: earned coverage drives citations and paid placements essentially never do. Tag press release programs to category-trend topics for early-funnel visibility, and run two citation-profile templates, broad deep source coverage for Claude and frequent evenly-distributed citations for ChatGPT.

Google rolled out five citation and link features to AI Mode and AI Overviews on May 6, 2026, announced by VP of Search Hema Budaraju: Further Exploration links at the end of responses, 'Subscribed' labels on links from a user's news subscriptions, Community Perspectives previews with creator names and handles, granular inline links next to the relevant claim, and desktop hover previews showing site name and page title, plus larger favicons and full author names. Google said people were significantly more likely to click links labeled as their subscriptions.

This is the first 2026 Google change that raises the per-citation traffic value rather than lowering it. Have clients with paid newsletters opt into subscription linking, atomize flagship content into discrete claims so each can earn an inline link, and have named executives rather than anonymous accounts participate in Reddit and forum threads now that creator attribution is surfaced.

Snap and Perplexity confirmed on May 6, 2026, during Snap's Q1 earnings, that their $400 million one-year integration deal announced in November 2025 had 'amicably ended'. The deal would have put Perplexity inside Snapchat's Chat interface for 483 million daily active users; the companies had not agreed on a path to broader rollout beyond tests with select users.

Deprioritize Perplexity for consumer-facing clients relative to ChatGPT, Claude, and Gemini, while keeping it as the priority surface for B2B research-intensive citations. Do not build a GEO plan around a brand reaching LLM users through a third-party app integration; organic citation presence is the durable route.

OpenAI made GPT-5.5 Instant the default model for all ChatGPT users on May 5, 2026, replacing GPT-5.3 Instant, with 52.5% fewer hallucinated claims on high-stakes medicine, law, and finance prompts and a 37.3% reduction in inaccurate claims on conversations users had flagged. Memory Sources began rolling out to all consumer plans, showing users what information personalized a response, with memory now pulling from past chats, saved memories, files, and connected Gmail. Paid users can keep GPT-5.3 Instant for three months.

Flag ChatGPT citation baselines captured before May 5 as pre-5.5 and re-baseline once the rollout settles, since the prior two Instant transitions each cut unique cited domains by about 20%. Note that brand presence in a customer's Gmail is now a personalization input, so subject lines and sender names in email programs are a citation-adjacent signal worth auditing.

OpenAI removed the $50,000 minimum spend on ChatGPT Ads on May 5, 2026, opening the Ads Manager beta to any verified U.S. advertiser, adding CPC bidding with recommended starting bids of $3 to $5 per click alongside CPM, releasing a conversion tracking API, and promising CPA bidding and third-party measurement. Minimum spend had fallen from $200,000 to $250,000 at the February 9 launch to $50,000 on April 10 to zero, and current ChatGPT ad CTR was reported at about 1.3% versus 29.2% for Google Search.

Mid-market and early-stage B2B SaaS clients can now test ChatGPT Ads with $5,000 to $10,000 a month; pair a capped pilot with the organic visibility audit to compare paid and organic citation lift. Use the 1.3% versus 29.2% CTR gap as the case for organic investment, and prepare to defend organic budget once CPA bidding ships.

Peec AI published 'Patterns we see in ChatGPT query fanouts' by Tomek Rudzki on May 5, 2026, analyzing 5 million query fanouts across ChatGPT, Perplexity, and Grok from April 1 to 21, 2026. ChatGPT ran 2.1 hidden fanout searches per user query, Perplexity 1.4, and Grok 6.8; ChatGPT inserted the word 'best' into 24.3% of advice-style fanouts and added current-year modifiers in 5.44%, it searched for reviews even when the user did not ask, and Grok hit Reddit in 10.5% of chats.

Carry explicit best-of framing on flagship category pages or roughly a quarter of ChatGPT's discovery queries will not find them, and keep review platforms current because they are checked on non-review queries. Add Grok to the standard coverage matrix, since each visible query triggers 6.8 background checks. Build pages that answer several fanout angles at once: best-of, reviews, comparisons, and current-year information.

At an invite-only New York event on May 5, 2026 Anthropic unveiled 10 pre-built financial services agent templates covering pitchbooks, earnings analysis, credit memos, underwriting, KYC screening, month-end close, statement audits, and insurance claims, made its Microsoft 365 add-ins for Excel, PowerPoint, and Word generally available with Claude for Outlook in beta, and added 8 data partnerships including Moody's as a native app, Verisk, Dun & Bradstreet, and Experian. FactSet fell as much as 8.1% and Morningstar more than 3% in May 5 trading, and Anthropic's 2026 run rate was reported above $30 billion with 1,000-plus customers spending $1 million or more a year. It followed the May 4 announcement of a roughly $1.5 billion enterprise AI services joint venture with Blackstone, Hellman & Friedman, and Goldman Sachs.

Elevate Claude to the top priority surface for fintech, insurance-tech, and compliance clients; it is now native inside Excel, Word, and PowerPoint and deployed inside banks and insurers. Add an internal-document readiness audit so client case studies and technical PDFs sit in customers' repositories in a form Claude will cite, and evaluate a Claude MCP connector for any API-first product.

5W Public Relations released the AI Platform Citation Source Index 2026 on May 1, 2026, synthesizing 6 published citation studies covering 680 million-plus citations across ChatGPT, Claude, Perplexity, Gemini, and Google AI Overviews from August 2024 through April 2026. The top 15 domains capture 68% of consolidated AI citation share, Reddit is the #1 source on every engine at about 40% citation frequency, Wikipedia accounts for 26 to 48% of ChatGPT's top-10 share, G2 ranks #19 and dominates B2B SaaS recommendation queries, and only 36% of Claude's journalism citations are from the past 12 months versus 56% for ChatGPT.

Plan earned presence on or adjacent to the top 15 domains rather than more owned blog volume, and put G2, Reddit, LinkedIn, and Wikipedia at the center of B2B SaaS strategy. Claude rewards evergreen documentation and analytical journalism, so clients with deep but older content libraries have a real opening there. Because 5W finds citation patterns shift within months, sell continuous monitoring rather than one-time audits.

Profound published a query fan-out study on April 30, 2026 tracking 10,000 prompts over 14 days across ChatGPT, Perplexity, and Copilot. 91% of ChatGPT's sub-queries were unique across runs, versus 14% for Perplexity and 47% for Copilot. Copilot converted 'how' prompts into keyword strings 93% of the time and Perplexity 78%, while ChatGPT preserved the question structure 31% to 40% of the time.

A single ChatGPT visibility check is one draw from a distribution, not a stable reading, so report rolling citation rates rather than point-in-time snapshots. For ChatGPT, build content clusters that cover adjacent phrasings of the same question. For Perplexity and Copilot, keyword-dense pages with clear product attributes win, and monitoring can run less often. Own the 'best [category]' framing everywhere, since Perplexity kept that format in 65% of runs and ChatGPT in only 39%.

OpenAI released GPT-5.5 (codename Spud) on April 23, 2026 to ChatGPT Plus, Pro, Business, and Enterprise, with GPT-5.5 Pro for Pro, Business, and Enterprise and API access following on April 24. OpenAI positioned it as a step toward an agentic super app that completes multi-step tasks across tools; it matches GPT-5.4 latency at a higher intelligence level and uses significantly fewer tokens on the same Codex tasks. Two days later an independent AA-Omniscience benchmark run by Karo Zieminski put GPT-5.5's hallucination rate at 86%, against 50% for Gemini 3.1 Pro Preview, 36% for Claude Opus 4.7, and 17% for Grok 4.20.

This is the third ChatGPT model transition in under six months, so any citation baseline captured before April 23 needs a fresh pull before it is used in reporting. Run brand-description accuracy checks now and compare what GPT-5.4 said against GPT-5.5, because an 86% hallucination rate on uncertain questions means brands with thin third-party coverage are more likely to be described confidently and wrongly. Treat G2, press, Wikipedia, and Reddit coverage as hallucination mitigation, not just citation building. Watch which work products (briefs, comparisons, spreadsheets) cite your brand as agentic use ramps.

OpenAI launched Workspace Agents in research preview on April 22, 2026 for ChatGPT Business, Enterprise, Edu, and Teachers plans: shared, long-running cloud agents built on Codex that persist across tasks, run on schedules or triggers, and connect to Slack, Google Drive, Microsoft apps, Salesforce, and Notion. Users describe a workflow in plain language and ChatGPT turns it into an agent. Free until May 6, 2026, then credit-based pricing; not available on Plus or Pro at launch.

OpenAI's own examples include a Lead Outreach Agent that researches inbound leads and updates the CRM and a Weekly Metrics Reporter, meaning ChatGPT's citation pool now feeds scheduled sales intelligence inside enterprise workflows. If your brand is absent from ChatGPT's search and training data, it is absent from every AI-generated lead brief those agents produce, on repeat. Frame ChatGPT visibility to buyers as a pipeline variable, not an awareness metric, and check that your product pages render cleanly for ChatGPT's fetcher.

Google launched Deep Research and Deep Research Max, two autonomous research agents powered by Gemini 3.1 Pro, in Gemini API public preview on April 22, 2026, with Google Cloud rollout to follow. Deep Research is tuned for speed and lower cost; Deep Research Max uses extended test-time compute to iteratively reason, search, and refine. For the first time either agent can combine open-web search with a company's private data streams in a single API call and return a fully cited report with embedded charts, with MCP support included.

Enterprise vendor evaluations and competitive analyses will increasingly be generated by these agents, comparing internal CRM data against Gemini's view of the public web. A brand that is not in Gemini's citation pool does not appear in the synthesized report that lands on the buyer's desk. Move Gemini citation work up the priority list for B2B clients and test category prompts against Deep Research outputs directly, not only against the Gemini chat interface.

Profound published multilingual citation research on April 21, 2026 covering 3.25 billion citations across 7 AI models, 14 countries, and 5 domain classifications, drawn from March 2026 prompt volumes. English-baseline social citation rates ranged from 15.3% on Google AI Overviews and 14.5% on Google AI Mode down to 9.1% on ChatGPT, 3.99% on Claude, and 3.6% on Gemini. In Portuguese queries YouTube took 65% of social citations while Reddit fell to 7%, in Arabic queries Instagram reached 29%, and ChatGPT kept Reddit at 51% to 76% of social citations across all 14 languages studied.

A single English-optimized playbook does not transfer to non-English markets, and can point at the wrong platform entirely. Build a per-market platform mix: YouTube for Portuguese-language visibility, Instagram for Arabic, and social investment for Spanish AI Overviews queries where social content surfaced at 1.4x the English rate. Reddit is the one constant, but only on ChatGPT. Any client selling into EU, LATAM, MENA, or APAC needs a localized channel plan, not a translated one.

A Resoneo study built on Meteoria data, reported by Search Engine Journal in April 2026, measured 27,000 comparable ChatGPT responses (400 prompts monitored daily for 14 weeks) and found that after GPT-5.3 Instant became the default ChatGPT experience, unique domains cited per response fell from 19 to 15 and unique URLs from 24 to 19, a 21% decline on both. URLs per domain held constant at 1.0, and Jerome Salomon of Oncrawl confirmed via server logs that ChatGPT-User crawl volume and frequency declined over the same period. The mechanism: GPT-5.3 Instant triggers fewer web searches and leans on training data more often.

Roughly 1 in 5 previously cited sources dropped out of ChatGPT's citation pool, so brands that were marginal candidates are now likely outside the window. Refresh every ChatGPT baseline captured before the GPT-5.3 Instant transition before using it in a report. The value of being in the top-cited tier went up, and the addressable live-search space is smaller than most audits assume. Pair this with training-data presence work, since fewer answers touch the live web at all.

Kevin Indig published 'The Ghost Citation Problem' in Growth Memo on April 20, 2026, analyzing 3,981 domains across 115 prompts, 14 countries, and 4 AI search engines: 62% of AI citations were ghost citations that link a domain without ever naming the brand, 74.9% of domains received a citation link at least once but only 38.3% received a brand mention, and just 13.2% got both in the same response. Gemini named brands in 83.7% of appearances but generated a citation link only 21.4% of the time, while ChatGPT cited 87% of the time but named brands in only 20.7% of answers. Evaluation and comparison content generated 30x more brand mentions than informational content.

Citation counts overstate visibility because most citations give the platform a link and the buyer no brand name. Split client reporting into three KPIs: citations (URL appears), mentions (brand named), and recommendations (actively suggested). Prioritize 'best tools for X' comparison and evaluation pages, which is where the 30x mention lift lives. Gemini and ChatGPT need different content architectures, since one is brand-first and link-light and the other is the reverse.

Google updated AI Mode in Chrome desktop on April 16, 2026 so that clicking a citation opens the publisher's page side-by-side with the AI conversation in a split pane instead of navigating away from the AI interface. The same update added cross-tab search, the ability to add images and PDFs as query context, and enhanced Canvas and image tools. US only and Chrome desktop only, with no expansion timeline given.

This is the first Google UI change that lowers the cost of clicking a citation, so a citation is now more likely to earn a visit. Page titles and meta descriptions do the selling inside the pane, so tighten them on cited pages. A side-pane load may not register like a full-page view, so clients using AI referral traffic as a KPI should verify GA4 and Search Console attribution. Enterprise clients running display on cited pages should audit ad rendering at narrow widths.

AirOps published a ChatGPT citation study via Search Engine Land on April 16, 2026 covering 16,851 unique queries run 3 times each for 50,553 responses across 353,799 pages in 10 verticals. Pages ranking #1 on Google were cited 58.4% of the time versus 14.2% at position #10; pages whose headings directly matched the query reached a 41% citation rate versus about 30% for loosely related headings. The optimal length was 500 to 2,000 words with pages under 500 and over 5,000 words both underperforming, the freshness sweet spot was 30 to 89 days old, and pages with JSON-LD schema were cited 38.5% of the time versus 32% without.

Google rank is still the single strongest lever for ChatGPT citation, so GEO and SEO are not decoupled for ChatGPT. Write headings that match the query wording exactly, keep pages in the 500 to 2,000 word range, and reconsider 5,000-plus word ultimate guides in favor of clusters of focused pieces. Put refreshes on a 60 to 90 day cadence so pages stay inside the 30 to 89 day citation window, and add JSON-LD for the modest but consistent lift.

OpenAI switched ChatGPT Ads from CPM to cost-per-click pricing on April 15, 2026, with bids of $3 to $5 per click and the minimum spend cut to $50,000 from the $200,000 to $250,000 managed-pilot floor; a self-serve ads manager with a Google Ads-like layout had gone live for a limited group of advertisers on April 13. The $60 CPM at the February 9 launch had degraded to $25 within 10 weeks, and OpenAI disclosed ad revenue targets of $2.5B for 2026, $11B for 2027, and $100B for 2030. Ads run on Free and Go tiers only, shown as promoted responses at the end of answers, and Sam Altman stated OpenAI will not accept money to influence the answer ChatGPT gives.

At $50,000 minimum and $3 to $5 per click, growth-stage B2B SaaS can now test ChatGPT ads without a six-figure commitment, and beta advertisers in B2B SaaS reported CPA 20% to 40% below Google Search campaigns. The current pricing is a market-building window ahead of global auction rollout (Canada, Australia, and New Zealand pilots went live April 16 to 17), so early tests build auction history before competition intensifies. Keep paid placement and organic citation as separate workstreams in client plans, since ads sit outside the answer itself.

Google expanded Personal Intelligence globally across AI Mode in Search, the Gemini app, and Gemini in Chrome between April 14 and 17, 2026, launching in India on April 14 and everywhere except the EEA, Switzerland, and the UK. Responses are personalized using connected Google account data such as past purchases, favorite locations, Gmail history, and Google Photos; it reached Google AI Plus, Pro, and Ultra subscribers first with free users following over the next few weeks. It had gone to all US Gemini users on March 17, off by default.

AI Mode answers are no longer uniform across users, so citation monitoring run in anonymous mode systematically under- or over-counts real exposure for your buyers. Flag this personalization gap in every client audit and stop treating aggregate citation rates as any individual's experience. For brands selling to existing customers, presence in a buyer's Gmail threads, calendar, and purchase history now feeds their AI Mode results, which makes owned channels like email a citation input.

Seer Interactive published on April 13, 2026 an analysis of 82,000 Gemini responses monitored since November 2025 across 20 brand workspaces showing Gemini's citation rate fell from 99% in February 2026 to 76% in March, with the sharpest drop between February 16 and March 2. One e-commerce brand fell from a 96% citation rate to 3.7% in a single week, YouTube citations in Gemini fell from 18% to 3%, and editorial sites such as Forbes, Medium, and NYMag declined most while Reddit and Wikipedia held steady at 44% and 33%. Over the same period headings appeared in 99.5% of Gemini responses (up from about 3%), markdown tables in 52% (from 0%), and average response length fell 15% from about 559 words to about 477.

Any Gemini citation baseline captured before mid-February 2026 is no longer valid, so re-pull before reporting. Gemini has shifted toward shorter, heading-heavy, table-structured answers that need fewer sources, and it is de-prioritizing editorial perspective in favor of reference-grade and community content. Clients leaning on thought-leadership blog posts for Gemini visibility should expect losses; restructure key pages with tables, clear headings, and verifiable facts.

Anthropic released Claude Sonnet 4.6 on April 13, 2026 with a 1M token context window in beta; it now powers the core API and Claude.ai across all tiers. The same day Anthropic announced Claude Mythos Preview, its most capable model to date at 93.9% on SWE-bench versus 80.8% for Opus 4.6, but withheld it from public release, giving access only to the Project Glasswing consortium (AWS, Apple, Microsoft, Google, and other major tech firms) with $100M in usage credits for vulnerability patching. Claude Opus 4.7 followed on April 16, topping SWE-bench Verified at 87.6%.

Do not count Mythos as a citation surface in audits; it is not public. Sonnet 4.6 is now the default behind Claude.ai and Claude-powered surfaces such as Yahoo Scout, and a 1M token window means Claude can read an entire brand knowledge base or docs set in one session, so deep, well-structured long-form content matters more for Claude than short answer blocks. Opus 4.7 measured a 36% hallucination rate on AA-Omniscience against GPT-5.5's 86%, so treat Claude as the more citation-reliable model when checking how your brand is described.

AI startup Oumi and The New York Times published a joint accuracy study of Google AI Overviews on April 8, 2026, running 4,326 SimpleQA queries against Gemini 2 (October 2025) and Gemini 3 (February 2026). Raw accuracy rose to 91% in February from 85% in October, but 56% of correct answers were ungrounded, meaning the cited sources did not support the answer given, up from 37% in October. Google spokesperson Ned Adriance called the study flawed because SimpleQA is an OpenAI-built benchmark that itself contained errors.

AI Overviews are getting more confident and less grounded at the same time, so being the source that actually supports the claim it is cited for is now a differentiator, not just hygiene. Publish specific data with disclosed methodology and conclusions the page visibly supports. Monitor what AI Overviews say about your brand and not only whether you are cited, because a citation attached to an unsupported claim is a reputation risk. Note the average AI Overview now exceeds 1,200 pixels in height, pushing organic results further below the fold.

OpenAI announced an overhauled ChatGPT shopping experience on March 24, 2026 built on a new Agentic Commerce Protocol that integrates retailers directly into ChatGPT's commerce layer, with Target, Sephora, Nordstrom, Lowe's, Best Buy, Home Depot, and Wayfair as retail partners. Users can upload images to find similar items and compare products side-by-side on price, reviews, and features; it replaces the underperforming Instant Checkout feature and was confirmed rolling out to all ChatGPT users in the week of April 26.

ChatGPT is now a product discovery engine, not only a research tool. Peec AI's analysis of 43,000 ChatGPT carousel products found 83% originate from Google Shopping organic results, with 60% from the top 10 Google Shopping rankings and 84% from the top 20, so for product queries the lever is Google Shopping rank, not on-page GEO. Resellers that outrank the brand's own site on Google Shopping will own the brand's carousels, so e-commerce and D2C clients need a Google Shopping feed track separate from content work.