Every page currently ranking for agentic SEO describes the same thing: software that runs your SEO work for you. None of them measured anything. We went the other direction and pulled 29 days of our own edge logs, because there is a second meaning of the term that has actual numbers attached to it.
Those numbers turned out to be worse than we expected, and more useful.
What is agentic SEO?
Agentic SEO means two things. The sold meaning is AI agents doing SEO work autonomously: research, drafting, publishing, monitoring. The measurable meaning is optimizing for AI agents that fetch your pages during a live answer. Only the second one changes whether a buyer sees your brand, and only the second one shows up in your server logs.
First-party edge logs · cite.solutions
Only 1.14% of AI bot traffic came from a bot that can cite you
601,929 verified AI bot requests, August 17 to September 14, 2026.
Read pages to train future models. Cannot link back, cannot cite.
GPTBot, ClaudeBot, Meta-ExternalAgent
One declared agent doing conventional search and AI work at once.
Amazonbot, Bingbot, Applebot, Googlebot
Fetch the page while a person is waiting. The only layer that cites.
ChatGPT-User, OAI-SearchBot, PerplexityBot
Inside the 1.14%: which bots actually showed up
One site. Bots classified by declared user agent at the edge. Perplexity-User logged one request and is omitted from the lower chart.
Agentic SEO means two different things, and only one is measurable
Read the pages currently ranking for this term and you get one definition. Ahrefs calls it "applying AI agents to SEO workflows so they can act, adapt, and recover on your behalf." Frase frames it as a system that "doesn't just report, it acts." SearchAtlas describes agents that "plan, execute, and iterate SEO actions continuously."
All three are describing a product category. None of them mentions a single crawler user agent.
The vendor meaning asks:
- •Can the agent write the brief without me?
- •How many autonomy levels does it have?
- •Does it keep watching the page after publication?
- •How much of my workflow does it replace?
The log meaning asks:
- •Which agent fetched this page, and when?
- •Was that fetch triggered by a live human question?
- •Did the fetch succeed, or did my WAF drop it?
- •Can that particular bot produce a citation at all?
Both are real. The first is a purchasing decision about your team. The second is a technical condition of your site, and it is the one your competitors are quietly failing.
Agents doing your SEO is a tooling choice. Agents reading your site is a measurement you either have or you don't.
The rest of this post is about the second one, because that is where we have data and where nobody else is publishing any.
What 601,929 AI bot requests say about the agent layer
We log every AI bot request to cite.solutions at the edge, classify it by declared user agent, and store it. Between August 17 and September 14, 2026, that came to 601,929 requests across 29 days.
Bots split into three jobs, and the split is not close.
Finding 1: Only 1.14% of AI bot traffic came from a bot that can cite you
Training crawlers took 538,190 requests, or 89.41% of everything. Mixed-purpose crawlers took 56,855, or 9.45%. Answer-time fetchers, the only bots that can put your link inside a live answer, took 6,884. That is 1.14%.
This is the second independent measurement of the same shape. Otterly ran an eleven-month experiment on a set of glossary pages and logged 1,209 AI agents against them: 84.4% training and data scrapers, 9.5% search indexing, 6.1% on-demand fetchers. Different site, different method, same conclusion: most of the AI traffic hitting you cannot produce a citation.
Nine in ten AI bot requests to your site are reading you for a model that will never link to you.
Our number is lower than theirs because our training share is inflated by one vendor running a heavy loop during the window. Strip the volatility and the direction holds either way.
Finding 2: ChatGPT-User is the largest citation-capable bot on our site
Inside that 1.14%, the distribution matters more than the total.
ChatGPT-User alone is 41% of every citation-capable request we received. That bot is not an index crawler. It is the fetch that happens mid-conversation, while somebody waits for an answer.
RESONEO reverse-engineered this path and published the architecture on Search Engine Land, confirming that ChatGPT-User, not OAI-SearchBot, is the agent that pulls page content during conversational browsing. Our logs agree with the ratio that implies.
If agentic SEO has a single concrete referent, this is it. One bot, 2,825 requests, each one a live human question in progress.
Finding 3: The agent layer is the steadiest number in your logs
Daily volumes tell a sharper story than the totals.
Training crawl volume moved 56-fold across four weeks. It is driven by whichever lab happens to be running a corpus refresh, and it tells you nothing about your visibility. Teams watching a bot dashboard see that line spike and conclude something improved.
Nothing improved. A crawler got hungry.
A training-crawl spike is a vendor's schedule, not your performance.
The answer-time line sat between 93 and 565 requests a day for the entire window while the training line went up 56x. That is the line worth alerting on, and most bot reports bury it inside an "AI crawlers" total that the training layer dominates by two orders of magnitude.
Finding 4: Answer-time bots read the blog, mixed crawlers ignore it
The two layers do not read the same site.
Answer-time fetchers spent 29.9% of their requests on blog posts and touched 2,300 unique paths. Mixed-purpose crawlers spent 4.7% on the blog across 36,683 paths, mostly images and dated study pages. One layer is looking for an answer; the other is enumerating a site.
Our Markdown twins took 9.9% of answer-time requests, which lines up with what we found when ClaudeBot fetched a Markdown twin for nearly every HTML page it took. A per-page alternate that every page declares gets used. That result sits oddly next to the finding that across 500 million AI bot visits, only 408 requests ever fetched /llms.txt. Both are ours, both are true, and the difference is discovery rather than format.
Finding 5: The bots most likely to lock you out are the mixed ones
The 9.45% middle layer is Amazonbot, Bingbot, Applebot and Googlebot. Each crawls under one declared agent for more than one purpose.
That single fact is the live risk in agentic SEO right now. Cloudflare's September 2026 default preserves declared search crawlers and blocks undeclared ones. A crawler that does conventional search indexing and AI work under one user agent can be caught by a rule aimed at AI training, and the crawler most likely to be caught is the one sending you organic traffic.
We have not confirmed that failure mode at a Cloudflare primary, and the vendor's own release does not address conventional search crawlers. Treat it as the thing to check on your properties this week rather than as an established fact.
The labs, for what it is worth, already split cleanly. OpenAI publishes three separate agents with their own IP ranges: GPTBot for training, OAI-SearchBot for search, ChatGPT-User for live fetches. Anthropic shipped the same three-way structure as a bots.json manifest. The taxonomy exists. Most robots.txt files we audit still treat all of it as one thing, which is the same error we wrote up in which AI crawlers get you cited.
Do you know which AI bots reach your revenue pages?
We read your edge logs, separate answer-time fetches from training crawls, verify each bot by IP, and show you which buyer-critical pages the citing bots have never seen.
Book a Technical GEO AuditHow to run agentic SEO on the side you control
You cannot make ChatGPT-User visit more often. You can make sure that when it arrives, the fetch succeeds and the page answers. That is the whole job.
Step 1: Split your bot logs into three layers before you read them
Classify every AI request as training, answer-time, or mixed-purpose. A combined "AI crawler traffic" number is dominated by training volume and will move for reasons that have nothing to do with you. Our full method is in the AI crawler log audit guide.
Step 2: Verify every bot by IP before you count it
A user agent string is a claim anyone can type. We once logged a single request calling itself Claude-SearchBot that asked for /.env from an IP in none of Anthropic's published ranges. That was a credential probe wearing a crawler's name. Check requests against the vendors' published ranges and Cloudflare's verified bot list, or your agent numbers are an upper bound rather than a measurement.
Step 3: Make the answer-time fetch cheap to serve
ChatGPT-User fetches while a person waits. A slow response, a JavaScript-dependent render, or a bot-challenge interstitial costs you the citation with no error anywhere in your analytics. Serve the answer in the HTML, and check that the bot sees the same content a browser does.
Step 4: Put the claim where a single fetch can reach it
The agent takes one pass. It does not click through your navigation or read your pillar page for context. Whatever you want quoted has to be on the page it landed on, stated plainly, near the top. This is the same reason passages beat pages for AI citation.
Step 5: Re-audit robots.txt and your WAF for mixed-crawler damage
Look for any blanket AI block, any legacy "block AI bots" toggle, and any rule written before the labs split their fleets. Confirm that ChatGPT-User, OAI-SearchBot, Claude-User and PerplexityBot are allowed and that nothing aimed at training is catching Googlebot or Bingbot on the way past. If you would rather not run this internally, a managed GEO agency can audit the fleet and the robots rules together.
What agentic SEO does not fix
Two things get oversold here, and both are worth saying plainly.
The first is attribution. A preprint published in September 2026, Purchase Advice and Observable Buyer Responses in Real AI Conversations, worked from 317 licensed conversations with commercial AI assistants and isolated 67 purchase-directed episodes. In 52 of those 67, the assistant offered options, channels or preferences. The purchase outcome itself was not observable in the channel. Agents advise inside a surface that does not report what happened next, so anyone selling you an agentic attribution number is modelling it.
The second is volatility. Our own 63-day study of 90,132 AI answers found the category leader flipped on only 18.7% of day pairs and never changed at all in 4 of 10 categories. The average category leader appeared in 78.2% of its category's answers. Agent behaviour is steadier than the daily-tracking dashboards imply, which is another reason to treat a moving bot line as noise until you have separated the layers. The full figures are on our AI search statistics page.
Your competitors are not your benchmark. The fetch log is.
FAQ
What is agentic SEO?
Agentic SEO refers to either AI agents performing SEO tasks autonomously, or the practice of optimizing a site so AI agents can fetch and use its pages during live answers. Vendor content almost always means the first. The second is the one that produces measurable changes in citations.
Is agentic SEO the same as GEO or AEO?
No. GEO and AEO are about being selected as a source in a generated answer. Agentic SEO, in the log sense, is about the mechanical step before that: whether the agent's fetch of your page succeeds and returns usable content. You can pass the agentic layer and still lose on relevance.
How do AI agents crawl websites?
They make ordinary HTTP requests under declared user agents with published IP ranges. OpenAI runs GPTBot, OAI-SearchBot and ChatGPT-User; Anthropic runs ClaudeBot, Claude-SearchBot and Claude-User. Only the live-fetch and search bots can lead to a citation. Training crawlers cannot link back.
Do AI agents actually buy things?
Not observably. The September 2026 arXiv analysis of real assistant conversations found purchase-directed advice in 52 of 67 episodes but no observable purchase outcome in the channel. Treat agentic commerce conversion claims as models rather than measurements.
Should I block AI agents from my site?
Block training crawlers if bandwidth or training-data policy matters to you, since they cannot cite you either way. Do not block answer-time bots such as ChatGPT-User, OAI-SearchBot, Claude-User or PerplexityBot. Those are 1.14% of AI bot volume and 100% of your AI citation surface.
The part worth acting on
The vendor definition of agentic SEO will keep selling, and some of those tools are genuinely useful. But it describes a change in how your team works, not a change in how you get found.
The log definition describes something you can check this afternoon. Pull 30 days of bot requests, split them into three layers, and look at what the answer-time bots actually fetched. On our site that was 6,884 requests out of 601,929, concentrated in one bot, aimed mostly at blog posts, steady while everything around it swung 56-fold.
If that layer never reaches your pricing page, no amount of autonomous content generation will fix it.
Find out what the citing bots see on your site
A technical GEO audit separates your three bot layers, verifies each agent by IP, and identifies the buyer-critical pages that answer-time fetchers have never successfully retrieved.
Book a Discovery CallContinue the brief
AI Crawlers: Which Ones Get You Cited?
Seven weeks of first-party logs, 761,885 hits. Most AI crawlers cannot cite you at all. Here is which ones can, and what to do about the rest.
How to Optimize for Agentic Search in 2026
Agentic search does not read one page. It fans out, opens sources, and checks claims. Here is how to optimize for agentic search in 2026.
What Is ClaudeBot and Should You Block It?
ClaudeBot sent 106,107 requests to our site in 100 days. Anthropic's answer-time bot sent 172. Here is what that gap means for blocking it.
Framework
Learn the CITE framework behind our GEO and AEO work
See how Comprehend, Influence, Track, and Evolve turn AI visibility into an operating system.
Services
Explore our managed GEO services and AEO execution model
Audit, prompt discovery, content execution, and ongoing monitoring tied to AI search outcomes.
Audit
Start with an AI visibility audit before execution
Understand prompt coverage, recommendation gaps, source mix, and where competitors are winning.