How answer engines decide who to name.

When an assistant answers a buying question by naming three companies, something specific happened. It wasn’t a lookup. It wasn’t a ranked list. A model interpreted the question, issued a set of searches the user never saw, pulled back a few dozen documents, and wrote a paragraph conditioned on whatever text ended up in front of it.

Most advice about “getting cited by AI” skips that sequence and jumps straight to tactics. We think the sequence is the whole story. Once you know where the decision actually gets made, the tactics sort into two piles: the ones that touch the retrieval step, and the ones that don’t.

Here is the mechanism, as precisely as the public record supports.

The answer is assembled, not recalled

The architecture behind every mainstream AI answer is retrieval-augmented generation. The term comes from a 2020 paper by Patrick Lewis and colleagues at Facebook AI Research and University College London, which combined a language model with “a dense vector index of Wikipedia, accessed with a pre-trained neural retriever” and found the hybrid produced more specific and more factual output than the model alone.[] Consumer assistants are a much larger, messier version of the same idea: a model that has been trained on text, bolted to a live search system it queries at runtime.

Google describes its own version in its documentation. Its generative features use retrieval-augmented generation to “review the specific information from those retrieved pages to generate a more reliable and helpful response, showing prominent, clickable links to relevant web pages.”[]

The step most people miss sits before retrieval. Google calls it query fan-out, and defines it as “a set of concurrent, related queries generated by the model to request more information and fetch additional relevant search results to address the user’s query.”[] Elsewhere it describes the same technique as “issuing multiple related searches across subtopics and data sources,” which lets the system “display a wider and more diverse set of helpful links” than a conventional results page.[]

That has a blunt consequence. You are not competing for the query the user typed. You are competing for a set of machine-generated sub-queries you will never see, whose composition changes as the underlying model changes. A page that is the single best answer to the literal question can lose to a page that happens to be the best answer to one of the six questions the model decided to ask instead.

Every engine is reading a different web

The second reason answers diverge is that the engines are not looking at the same corpus.

Google’s AI features run on Google’s index and, in Google’s words, “are rooted in our core Search ranking and quality systems.”[] Eligibility is the ordinary eligibility of Search: the page must be indexed and permitted to show a snippet. Google states plainly that “there are no additional technical requirements.”[]

OpenAI runs a different stack. Its documentation distinguishes three user agents, and only one governs whether you can appear in an answer with a link: “OAI-SearchBot is for search. OAI-SearchBot is used to surface websites in search results in ChatGPT’s search features.” GPTBot is for training data. ChatGPT-User handles live fetches triggered by a specific user action.[] A site that blocks GPTBot for training reasons and accidentally blocks OAI-SearchBot has removed itself from ChatGPT’s answer surface — a distinction that is invisible in a normal SEO audit. OpenAI also says ChatGPT search “leverages third-party search providers, as well as content provided directly by our partners,” which means part of its retrieval layer is licensed rather than crawled.[]

Perplexity operates its own index. In its September 2025 engineering post it describes “an index covering hundreds of billions of webpages” and, more interestingly, a retrieval design that “divides documents up into fine-grained units” which “are individually surfaced and scored against the original query parameters.”[] That is a meaningful architectural detail: the unit of competition is a passage, not a page.

Three corpora, three crawlers, three rankers, three chunking strategies. Before any question of quality or authority enters the picture, the engines have already disagreed about what documents exist.

Ranking is a prior, not a gate

The most common assumption in the market is that AI citations are downstream of rankings — get to page one and the citations follow. That was closer to true than it is now, and the change is documented.

In July 2025, Ahrefs analyzed 1.9 million citations drawn from one million AI Overviews and found that 76.1% of cited pages ranked in Google’s top 10, with 86% appearing somewhere in the top 100. The authors were careful about the causal reading even then, noting that ranking first gives you “a coin flip at best” at being cited.[]

In March 2026 the same authors re-ran the analysis across 863,000 keyword SERPs and roughly four million AI Overview URLs. Top-10 citations had fallen to 37.9%. Positions 11 through 100 accounted for 31.2%, and pages ranking nowhere in the top 100 accounted for 31.0%. They attribute the shift substantially to more aggressive query fan-out, which pulls citations from the results of sub-queries rather than the original one.[]

Two honest caveats. Part of the gap is measurement — Ahrefs notes improved detection between the two studies, so the true decline is smaller than 76 to 38. And other measurements disagree: Search Engine Journal, reporting the update, noted a separate BrightEdge analysis putting top-10 overlap near 17% on a different methodology, and called the whole metric “a moving target.”[]

The directional conclusion survives the disagreement, and it is the one that matters. Ranking is a strong prior on being retrieved; it is no longer a gate. Roughly a third of cited URLs in the 2026 sample rank nowhere in the top 100 for the query being answered. Something other than classical ranking is putting them in the context window.

What the engines actually lean on

There is now decent public data on which domains show up in AI answers, and it is more useful for what it reveals about engine personality than for the domain names themselves.

Semrush analyzed 230,000 prompts across ChatGPT search, Google AI Mode, and Perplexity between July 14 and October 12, 2025, covering more than 100 million citations. Across the set, Reddit, Wikipedia, LinkedIn, Forbes, and Medium ranked highest overall — but the per-engine mixes were sharply different. Wikipedia’s share of ChatGPT responses ran near 55% at the start of the window. On Google AI Mode it held near 3%, and on Perplexity near 0.8%.[] Perplexity’s top sources skewed toward Reddit, LinkedIn, NIH, Microsoft, and Google.[]

A larger sample tells a similar story. Peec AI analyzed 30 million sources across ChatGPT, Google AI Mode, Gemini, Perplexity, and AI Overviews; Reddit was the most-cited domain overall, followed by YouTube and LinkedIn, with Wikipedia and Forbes also in the top five. The per-engine tendencies again diverged — ChatGPT toward Wikipedia, Reddit, and editorial outlets like Forbes; Google AI Mode toward platforms including Facebook and Yelp; Perplexity toward Reddit, LinkedIn, and G2 on B2B questions.[] That coverage reports rankings rather than frequencies, so we will not attach percentages to those tendencies.

The volatility is as instructive as the ranking. In that same window, ChatGPT’s Reddit citation rate went from close to 60% of prompt responses in early August to around 10% by mid-September, and its Wikipedia rate fell from roughly 55% to under 20% — while both held steady on AI Mode and Perplexity.[] That is not the web changing. That is one vendor changing a retrieval policy.

A strategy built on being present in whatever source type an engine currently favors is built on a surface that can move tens of percentage points in six weeks, without notice, on one engine and not the others.

Why corroboration beats anything you publish about yourself

This is the part that is hardest to sell and easiest to defend.

Start with the mechanism. Fan-out generates many sub-queries. Retrieval scores documents — or, on Perplexity, passages — against each of them independently.[] Your own website is one document, or a handful. It can win one sub-query. Ten credible independent pages that mention you can win ten, and can win the ones about comparison, cost, reliability, and alternatives that your own site was never going to be retrieved for. Corroboration is not a trust signal in some abstract sense. It is a larger retrieval surface.

Then there is what happens once documents are in the context window. When retrieved passages agree, the model produces a confident, specific answer. When they conflict, it hedges or drops the disputed detail. Two independent sources saying the same thing about you resolves that conflict in your favor in a way one self-published claim cannot.

Google’s own instructions to its human quality raters make the underlying preference explicit. The reputation research section directs raters to seek “independent reviews, references, recommendations by experts, and other sources of credible information about the website,” explicitly excluding material produced by the company itself, and instructs them: “Be skeptical of claims that websites make about themselves, particularly when there is a clear conflict of interest.” The guidelines go further: “When there is disagreement between what the website or content creators say about themselves and what reputable independent sources say, trust the independent sources.”[]

Be precise about what that document is. It governs human raters evaluating Search quality. It is not a description of AI Overview internals, and Google has never published one. The connection is an inference — a well-supported one, given Google’s own statement that its generative features are rooted in the same core ranking and quality systems those guidelines exist to evaluate[] — but an inference, and we would rather label it than let it pass as documented mechanism.

There is also a harder, less comfortable version of the argument. The Tow Center for Digital Journalism at Columbia ran 1,600 queries across eight generative search tools in early 2025, feeding each an excerpt from a known article and asking it to identify the source. The tools answered more than 60% of queries incorrectly, with error rates ranging from 37% on Perplexity to 94% on Grok-3. Notably, the engines repeatedly credited syndicated republications instead of originals: “On some occasions, chatbots directed us to syndicated versions of articles on platforms like Yahoo News or AOL rather than the original sources.” In one case, though USA Today blocks ChatGPT’s crawler, the chatbot “still cited a version of its article that was republished by Yahoo News.” Formal licensing did not fix it — despite a partnership with OpenAI, ChatGPT correctly identified only one of ten San Francisco Chronicle excerpts.[]

Read that as a statement about the machine rather than about journalism, and it says something clear: engines cite what they retrieved, not what is canonical. Presence in the wider graph of pages that discuss you is doing more work than control of the page you own.

What structured data actually does

Google’s position here is unambiguous, and it contradicts a large amount of agency marketing. From the documentation: “Structured data isn’t required for generative AI search, and there’s no special schema.org markup you need to add.”[] And: “There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary.”[]

So we are not going to tell you schema causes citations. There is no public evidence for that claim, and the platform that would know says otherwise.

What it does do is narrower and still worth doing. It makes unambiguous, machine-readable assertions about entity facts — legal name, location, hours, price, identifiers, and the sameAs links tying your entity to its profiles elsewhere. Those are exactly the facts that get contradicted across the web, and exactly the contradictions that make an engine hedge. It also remains the mechanism for rich result eligibility in conventional Search, which still feeds the index the generative layer retrieves from.

Adoption is broad but shallow. HTTP Archive’s 2024 Web Almanac found JSON-LD on 41% of pages, up from 34% in 2022, with the most common types being WebSite at 12.73% of mobile pages, Organization at 7.16%, and LocalBusiness at 3.97%.[] Most sites are marking up that they exist, not what they do.

The closest thing to controlled evidence on content-level tactics is academic. The GEO paper by Aggarwal and colleagues, published at KDD 2024, tested content modifications against a benchmark of generative-engine responses and reported visibility gains of up to 40%, with citations, quotations, and statistics among the more effective additions — and effectiveness varying substantially by domain.[] That was measured on the authors’ own benchmark, not on live commercial engines, and it predates most of the systems people care about now. Suggestive, not decisive.

Why the same question returns different names

Pulling it together, there are at least six independent reasons two engines name different companies for the same question:

Different indexes and different crawler permissions decide which documents exist at all.[][][] Different fan-out behavior decides which sub-queries get run. Different retrieval granularity — whole page versus scored passage — decides what fragment competes.[] Different source-mix policies, which we now know can shift by tens of percentage points inside a quarter, decide which kinds of domains get pulled.[] Personalization, location, and recency shift results for the same user across sessions. And generation itself is sampled, not deterministic, so identical retrieved context can still produce a different set of names.

Which is why a spot-check is worthless as measurement. One query on one engine on one day tells you nothing you can act on. Anyone reporting AI visibility from a handful of manual prompts is reporting noise.

It is also why the citation displayed next to a claim should not be read as provenance. Liu, Zhang, and Liang’s audit of generative search engines found that only 51.5% of generated sentences were fully supported by their citations, and only 74.5% of citations supported the sentence they were attached to.[] That work is from 2023 and evaluated engines that have since changed or shut down, so treat the numbers as historical. The Tow Center’s 2025 findings suggest the failure mode did not go away.[]

What we do not know

We would rather be clear about the edges of this than pretend to more than the record supports.

No engine publishes its retrieval scoring, how many documents enter the context window, or how candidates are ranked once retrieved. Google confirms fan-out exists but not how many sub-queries it generates or how they are chosen.[] Nobody outside these companies can distinguish a source that was retrieved and used from one retrieved and discarded. Every third-party citation study is an observational study of a black box, on a query sample the researcher chose, at a moment in time — which is why credible studies of the same question return numbers ranging from 17% to 76%.[]

What is well-supported: answers are assembled from retrieved documents, not recalled from memory. The retrieval step is where the decision is made. Engines read different webs. Ranking helps and no longer decides. Independent corroboration expands the retrieval surface in a way self-published content cannot. Structured data is hygiene rather than a lever, on the platform owner’s own account. And the whole system moves under you, quietly, on a schedule nobody publishes.

Anyone claiming more precision than that is selling something.