How to Optimize for AI Retrieval, Not Just Rankings

The question every B2B marketing lead is asking right now is the same one. The page ranks. Sometimes top three. Sometimes position one. And yet Google AI Mode never cites it. ChatGPT never cites it. Perplexity never cites it.

The answer is buried in two pieces of evidence that landed in the same month. The first is court testimony from the Google antitrust trial. The second is a Google Research paper called TurboQuant. Read together, they say something simple and uncomfortable.

Rankings are a downstream filter. Retrieval is the gate. And the gate is about to get wider, not narrower.

Why top 10 rankings stopped predicting AI citations

Most B2B marketing teams still treat "rank in the top 10" as the goal post. That mental model was built for a Google that physically could not afford to do anything else. The hardware is changing. The mental model has not caught up.

Reason 1: RankBrain only sees the top 20 to 30 results

Court testimony from Paul Nayak, Google's VP of Search, confirmed during the antitrust trial that RankBrain runs on only the top 20 to 30 candidates per query. Everything else gets culled by classical retrieval before the deep-learning ranking layer even sees it. Search Engine Land's analysis summarized the testimony bluntly: the deep-learning component SEOs built a decade of theory around is deliberately withheld from the bulk of the index because Google cannot afford to apply it more broadly.

The candidate pool size is a cost decision, not a relevance decision.

Reason 2: TurboQuant removes the memory ceiling on retrieval

In March 2026, Google Research published TurboQuant, a vector compression algorithm with two material properties. It cuts memory for vector representations by 4x to 4.5x. And it does that with near-zero indexing time and zero accuracy loss.

VentureBeat reported the broader implication: an 8x speedup on KV-cache workloads, with infrastructure costs cut by 50 percent or more. InfoQ noted the algorithm needs no calibration data and no dataset-specific tuning, which is what makes it deployable at index scale instead of as a research curiosity.

If indexing is virtually free and memory per vector drops by 4x, the economics that capped the candidate pool at 20 to 30 documents no longer hold. Google can plausibly evaluate a candidate set several times larger on the same hardware.

Reason 3: AI Mode and AI Overviews already widen the pool

Google AI Mode and AI Overviews do not run on the classical top-10 SERP. They run their own retrieval pass against a much larger candidate set, score passages instead of pages, and synthesize answers from the passages that survive. We covered the mechanics in why Google rankings no longer predict AI citations.

Rank tracking misses most of what AI surfaces actually cite. Conductor and Profound have reported tracked-keyword rank correlation with AI citation appearance below 20 percent for B2B SaaS portfolios.

Reason 4: Retrieval and ranking are two different scoring layers

Classical SEO collapses both into one number. AI retrieval does not. A page can fail to enter the candidate pool, fail to extract a passage, fail to rank inside the pool, or fail to survive the synthesis pass. Each failure mode has a different fix.

Treating "where do I rank" as the single diagnostic question hides three out of four failure modes from your reporting.

What retrieval optimization asks that SEO does not

The two disciplines look similar from the outside. They share tools, vocabulary, and a lot of working files. They diverge on the questions they put first.

Classical SEO asks:

• •What keyword should this page rank for?
• •How many backlinks does it have?
• •What is its position on the SERP?
• •Does it beat the competing top 10?

Retrieval optimization asks:

• •Is this page in the AI candidate pool at all?
• •Can a clean passage be extracted from it?
• •Does the brand get cited consistently across ChatGPT, Claude, Perplexity, Gemini, and Google AI Mode?
• •When the candidate pool widens, will this page be in or out?

The first set of questions is about competing inside a known 10-slot window. The second set is about whether you enter the retrieval candidate set in the first place. They are not the same question.

Retrieval optimization is what SEO becomes once the retrieval pool stops being scarce. That moment is happening now.

How to optimize for AI retrieval in five steps

Here is the practitioner workflow we run for B2B SaaS clients. Five steps. Each one is something a marketing team can start this quarter without rebuilding the stack.

Step 1: Audit which AI crawlers are reaching your site

Open your server logs and filter for user agents containing GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, Claude-Web, anthropic-ai, PerplexityBot, Perplexity-User, Google-Extended, and Bytespider. Bucket the hits by URL, by week, and by user agent. If a key product page or comparison page is not getting hit by GPTBot or ClaudeBot at all, that page is not in the candidate pool. No amount of on-page optimization fixes a page that the crawler never sees.

Most B2B SaaS dashboards do not surface AI crawler activity by default. That is the single biggest blind spot in B2B retrieval reporting right now. We walk through the full audit in the GEO crawlability audit for AI retrieval.

Step 2: Build retrieval-friendly passages, not just ranked pages

AI surfaces cite passages, not pages. A passage is a 40 to 80 word block that answers one question cleanly, sits under a heading phrased as that question, and contains the specific data point the AI needs. If your top-of-page contains a long brand narrative before the answer block, the extraction fails.

Three concrete rules:

• •One question per heading, phrased as the user would ask it.
• •One direct-answer paragraph under each heading, 40 to 80 words, leading with the answer.
• •One data point per paragraph, named source, linked source.

We cover the structural rules in passages beat pages: how to structure content for AI citation.

Step 3: Track citations across all major AI surfaces

Rank tracking and citation tracking are different jobs. Rank tracking checks position one through ten on Google SERPs. Citation tracking checks whether your brand appears as a named source inside ChatGPT, Claude, Perplexity, Gemini, Google AI Mode, and Copilot answers to the prompts your buyers actually use.

For each prompt your category produces, log:

• •Was the brand cited? Yes or no.
• •Which URL was cited?
• •Which competitors were cited?
• •How did the answer frame the brand vs competitors?

This is a measurement layer most stacks do not have. Share of voice in AI search walks through how to build the baseline.

Step 4: Fix structured data so the candidate pool can pull you in

Schema is now a retrieval signal more than a ranking signal. Perplexity, ChatGPT Search, Gemini, and Google AI Overviews all parse FAQPage JSON-LD as a primary extraction signal. Otterly's controlled experiment found pages with FAQ schema received 2,379 citations vs 529 baseline, a 350 percent lift. That is why we wrote the case against removing FAQPage schema when Google killed the SERP rich result.

Schema priorities for a B2B SaaS page going forward:

• •FAQPage JSON-LD on product, pricing, and comparison pages.
• •Product schema with verified attributes.
• •Organization schema with sameAs links to LinkedIn, Crunchbase, G2.
• •Article schema with author and datePublished on every blog post.

Step 5: Close the loop with server log signals and refresh cadence

The audit is not a one-time job. AI crawlers update fingerprints. Retrieval pools widen. Models change. The B2B SaaS teams winning at retrieval visibility today are running a monthly close-the-loop cycle:

• •Pull server logs, recheck crawler coverage by URL.
• •Re-run the prompt panel, recheck citation share.
• •Refresh underperforming pages with the citation-rate data attached.
• •Track which refreshes lift citation rate within 14 days.

That is the loop. The teams running it see citation rate lifts of 30 to 60 percent within a quarter. The teams not running it stay flat because their reporting still tells them they rank.

What changes if Google flips the TurboQuant switch in production

The piece nobody knows the answer to is when. Google has not announced that TurboQuant is live in production retrieval. The paper, the Google Research blog, and the TechCrunch coverage treat it as an infrastructure capability, not a shipped product change. Production rollout could be next quarter. It could be next year. It could be silent and partial.

What is not in doubt is the direction. The cost ceiling that pinned the retrieval candidate pool at 20 to 30 documents is being removed. Once removed, it does not come back.

The teams that prepared for a wider pool by auditing crawler coverage, building extractable passages, and tracking citations across all surfaces will pick up the marginal traffic when the switch flips. The teams that kept optimizing for top 10 will not know it happened until the rank report still looks fine and the AI citations still go to the same handful of competitors.

You do not need to predict the timing. You just need to be ready for the pool when it opens.

FAQ

Did Google confirm TurboQuant will run in production search?

Not yet. Google Research published the algorithm at ICLR 2026 with benchmark results and infrastructure cost claims. Google has not stated a production rollout date for ranking or retrieval. The infrastructure economics, however, are explicit: 4x to 4.5x memory reduction, near-zero indexing time, zero accuracy loss. That is the kind of property that gets deployed.

Will my top 10 ranking still matter?

Yes, for classical Google search results. No, as a reliable predictor of AI citations. Top 10 rank correlation with AI citation appearance is already below 20 percent for B2B SaaS in tracked portfolios. If the retrieval pool widens, that correlation drops further. Rank reporting alone is no longer a complete visibility report.

How do I know if my pages are in the AI retrieval candidate pool?

Server log analysis is the cleanest diagnostic available today. Pull a 90-day log window, filter for AI crawler user agents (GPTBot, ClaudeBot, PerplexityBot, Google-Extended, OAI-SearchBot, ChatGPT-User), and bucket by URL. Pages with regular crawler hits are in the candidate pool. Pages with zero hits over 90 days are not.

What is the difference between retrieval optimization and SEO?

SEO optimizes for ranking position inside a known, narrow candidate pool. Retrieval optimization optimizes for entering the candidate pool in the first place, then for being passage-extractable once inside. They share tools and vocabulary. They diverge on diagnostics, KPIs, and which failure modes get fixed first.

Should I drop my old SEO playbook?

No. Keep it for traffic that still comes through classical Google SERPs, which is still a meaningful share for most B2B SaaS portfolios. Add the retrieval optimization layer on top so you also win citations on AI Mode, AI Overviews, ChatGPT, Claude, Perplexity, and Copilot. The two stacks are additive, not substitutes.

The takeaway

Top 10 rankings made sense in a world where running deep-learning ranking on a wider candidate pool was too expensive. TurboQuant and the broader vector-compression wave are removing that cost ceiling. AI Mode and AI Overviews already ignore the top 10 framing.

Audit crawler coverage on your top revenue pages this week. If GPTBot or ClaudeBot has not visited the page in 90 days, that is your starting point.