AI Search Visibility Is a Retrieval Problem, Not a Markup Problem

The GEO checklist is aimed at the wrong layer

Most generative-engine optimization advice starts with the page: add an FAQ block, expose structured data, write a summary, improve the author box, or tune a page-speed score. Those are reasonable website practices, but they are a weak theory of how an AI answer finds a source.

The better starting point is retrieval. An engine has to decide whether to search, rewrite the user’s request, retrieve a candidate page, rank or filter that candidate, place evidence into context, and finally display a citation. A page cannot be cited if it never enters that pipeline. A page can also be retrieved and still not be cited.

That distinction is now supported by several independent studies. A 90-day study of 158,099 answers and 468,037 citations found that retrieval explained 32% of within-website citation variation, while page properties explained only 2.8% for ChatGPT. A separate audit of four live AI search surfaces found a mean exact-URL Jaccard overlap of 0.0079 across matched prompts; 96.4% of observed URLs appeared in only one engine. These are not results from the same experiment, so they should not be merged into one universal number. Together, however, they make a useful argument: “AI visibility” is not one page score.

The practical consequence is uncomfortable for SEO tooling. A score that inspects HTML without observing an engine can estimate query-page fit or content quality. It cannot honestly call itself a probability of being cited everywhere.

Separate exposure from citation before optimizing either

The first measurement mistake is treating retrieval and citation as synonyms. They are different events with different levers.

Let E mean that a page is exposed to the answer-generation process and C mean that the page is shown as a citation. A minimal model is:

P(C | page, query, engine, time)
  = P(E | page, query, engine, time)
    × P(C | E, page, query, engine, time)

The first term includes crawling, indexing, search activation, query rewriting, fan-out, retrieval, ranking, filtering, and context allocation. The second includes relevance, evidence quality, source selection, answer composition, and citation formatting. A fixed-candidate experiment sets exposure to one by design; it can study conditional selection, but not whether a production engine discovers the candidate in the first place.

This is why a page can be perfectly crawlable and still be invisible for a commercially important question. Google’s official AI features documentation says that pages need to be indexed and eligible for a normal search snippet, but there are no additional technical requirements or special schema needed for AI Overviews and AI Mode. It also explains that these surfaces may use query fan-out and that their models and techniques can differ.

That guidance does not make technical SEO irrelevant. It changes its role. Robots.txt, internal links, rendered text, accurate structured data, page experience, and Merchant Center data help maintain eligibility and discoverability. They are admission controls, not a citation guarantee.

The distinction also changes reporting. Instead of one “AI score,” report at least four quantities:

Quantity	What it asks	What it can support
Crawl and index eligibility	Can the engine access and index the page?	Technical remediation
Retrieval exposure	Did the engine return or consider the page for this request?	Query and content alignment
Citation selection	Was the exposed page shown as a source?	Evidence and source-quality work
Business outcome	Did the mention produce a click, lead, or sale?	Budget and prioritization

A page may be strong in one row and weak in another. Combining the rows hides the decision you need to make.

Query fit beats generic “AI-friendly” formatting

The strongest page-level result in the University of Toronto study is not “add more structure.” It is semantic alignment. Pages whose titles and paragraphs matched the questions people asked were cited more often. FAQ blocks, structured data, summary boxes, author boxes, and page speed did not survive the same within-website tests as reliable explanations of citation variation.

That finding should not be overread. The study does not prove that structured data has no value for every search product. It shows that popular checklist items are poor standalone explanations once website-level differences are controlled. A schema implementation can clarify an entity or a product without causing an engine to retrieve the page for a particular question.

The right editorial test is therefore simple: can the page answer a real sub-question in the language of the user, without forcing an engine to infer the connection?

For a B2B software page, “inventory management platform” may be a necessary category phrase, but it is not a complete retrieval target. Users and engines may ask:

• How do I reconcile available stock after a partial shipment?
• Can one SKU have different sellable states across marketplaces?
• What is the safest workflow for a bin audit?

A page that explains those operations with concrete vocabulary gives a retrieval system more meaningful matches than a generic introduction followed by an FAQ assembled for keywords. It also gives a human buyer evidence that the company understands the problem.

This is where our earlier analysis of SEO visibility as an asset-level measurement problem becomes operational. Measure individual pages against the questions they are intended to answer, rather than treating domain-wide authority as a substitute for page-level relevance.

Engine divergence makes universal scores unreliable

The second measurement mistake is assuming that visibility transfers between engines. It often does not.

In a June 2026 benchmark of 15 commercial prompts across ChatGPT, Microsoft Copilot, Google, and Perplexity, Benjamin Tannenbaum recorded 589 citation observations, 528 unique URLs, and 356 domains. Mean pairwise exact-URL overlap was 0.0079; the median was zero, and 84.9% of engine pairs shared no cited URL. A single engine captured between 11.4% and 42.6% of the four-engine URL union. The study’s scope is limited—15 prompts on a particular date—and its author warns that the intervals describe that benchmark rather than every commercial query. The direction is still important: the engine is part of the result.

The same study found that 96.4% of URLs appeared in only one engine, while adjacent-day turnover within the same engine was much lower than cross-engine divergence. That suggests two different kinds of volatility:

1. Product divergence: each engine has its own retrieval sources, ranking behavior, and citation policy.
2. Temporal drift: the same engine’s candidate set changes as indexes, models, prompts, and product settings change.

A universal GEO score collapses both. It might be useful as a static content diagnostic, but its name should say what it measures. “Question-page alignment score” is more honest than “AI visibility score” when no live engine is being observed.

The implication for competitive research is direct: do not infer that a win in ChatGPT proves a win in Google AI Mode. Run matched prompts per surface, preserve the date and locale, and report exact URLs separately from domains or entities. The production ChatGPT citation study found that external proxies matched ChatGPT Web better at hostname and entity levels than at exact-URL level, and that Google Search plus the OpenAI Responses API recovered only about one quarter of cited URLs in the tested settings. A domain-level presence signal may be useful, but it is not the same as a page citation.

Brand memory is a separate path to visibility

Retrieval is not the whole story. Models also carry prior knowledge about brands, and that memory can shape which searches they issue and which companies they name.

The University of Toronto researchers found that ChatGPT wrote known brands into its own searches, preferred websites it knew, and sometimes named brands without a supporting page. In their reported observations, 46% of brand mentions lacked a supporting page. This does not mean a brand can ignore evidence. It means the page-citation path and the brand-recognition path can produce different outcomes.

That distinction matters for measurement. A company may be named in an answer but not cited. It may be cited through a publisher, marketplace, or review site but not through its own domain. It may appear in a response because the model already knows the brand, even when the prompt did not retrieve a first-party page. These are different forms of visibility with different risks.

Treat brand memory as an earned-entity problem, not a markup hack. Build a consistent public record: accurate product facts, durable naming, useful first-party documentation, and independent sources that describe the company without repeating the same boilerplate. Then measure brand mentions, supporting citations, source domains, and factual accuracy separately.

Do not turn this into a campaign to seed names everywhere. The evidence supports monitoring how engines behave, not manufacturing mentions. Google’s guidance for generative AI features explicitly emphasizes unique, non-commodity, people-first content and warns against producing pages for every query variation merely to influence generative responses.

A measurement stack that survives product changes

A useful AI-search program should be designed like an experiment, not a one-time audit. Start with a fixed question set grouped by intent: discovery, comparison, troubleshooting, recommendation, and brand lookup. Store the exact prompt, locale, device, engine, mode, date, and conversation state. A follow-up prompt is not equivalent to a standalone prompt because it carries context.

Then capture four layers of evidence:

1. Eligibility: crawl response, robots directives, index status, rendered text, canonical URL, and relevant structured data.
2. Exposure: retrieved candidates or observable search results, including query rewrites where the product exposes them.
3. Citation: exact URL, hostname, entity, citation position, and the claim the source appears to support.
4. Outcome: referral sessions, engaged time, assisted conversions, leads, and revenue where attribution is possible.

Repeat each prompt. A single answer is an anecdote; a distribution is a measurement. The ChatGPT study found that candidates retrieved more often across repeated runs were more likely to be cited, while retrieval behavior differed across model settings. That is a reason to report frequency and variance, not just a binary “mentioned once.”

Use a compact experiment table for each target question:

Field	Example decision
Query family	“best inventory workflow for partial shipment”
Intended page	Fulfillment guide, not the homepage
Engines	Google AI Mode, ChatGPT Web, Perplexity
Exposure metric	Candidate retrieval frequency across runs
Citation metric	Exact URL and hostname citation rate
Quality check	Does the citation support the claim?
Business metric	Qualified referral or assisted conversion

Finally, attach an intervention to the layer it can plausibly change. Fix blocked crawling for eligibility. Rewrite a page around a real user question for retrieval fit. Add evidence and clearer source boundaries for citation selection. Improve independent documentation and product consistency for brand recognition. Do not add schema and then claim that every downstream movement came from it.

The recommendation: optimize the observable bottleneck

The best near-term strategy is not to abandon technical SEO or to chase a new GEO syntax. Keep the fundamentals that make pages eligible, then spend optimization effort where measurement shows a bottleneck.

If a target page is not indexed, fix access and rendering. If it is indexed but never retrieved for the intended questions, improve its topical coverage, internal links, and direct answers. If it is retrieved but rarely cited, test evidence density, source clarity, and whether the page actually supports the answer an engine is composing. If it is cited but produces no business value, investigate intent, snippet-to-page continuity, and conversion design.

That sequence is less exciting than a universal checklist, but it is more defensible. The current evidence says generative visibility is a multi-stage, engine-specific process. Measure the stage you want to change, name the uncertainty you cannot observe, and keep page quality accountable to people rather than to a speculative score.