How an AI Search Engine Turns a Query into a Citation
For a brand page to appear in an AI-search answer, it must pass through more than a writing-quality test. This chapter traces the path from a user's question through search activation, retrieval, context allocation, generation, attribution, and user behavior so that a reported “visibility” result can be tied to the stage it actually measures. The practical judgment is simple: before asking how to improve a source, identify where in the chain the source was lost, used, or misrepresented.
Before the answer: activation, retrieval, and context
The first event is easy to overlook because it produces no citation at all: the engine decides whether to activate search. A user may ask a question and receive a response from a model's existing knowledge, a direct answer assembled from a product's internal system, or a web-grounded answer that retrieves current pages. Those are different information paths. If web search is not activated, a brand's page cannot be retrieved or cited in that interaction, regardless of how well the page is written.
This matters when a team reports a citation rate. Suppose a dashboard counts only answers that contain citations. It may describe how often a source is cited conditional on a search having run and the source having been available. It does not describe the probability that the brand is visible across all relevant user questions. The critical survey of GEO notes that, in one ChatGPT configuration, 57.8% of repeated outputs did not activate web search. That figure belongs to a particular product, query universe, and measurement design, but the underlying lesson is general: no-search outcomes are part of the visibility problem, not inconvenient observations to remove.
When search is activated, the engine has to assemble candidates. That process begins upstream of the generated prose. Pages may need to be crawled and indexed; an index may contain a page without returning it for a particular query; and a search system may reformulate the user's wording into several underlying searches. A product question such as “Which of these cameras is best for travel in wet weather?” may be decomposed into needs about weather sealing, weight, battery life, price, or independent testing. The brand's page is then competing not only for the exact words in the original question but for the engine's interpretation of the information need.
A source that is not in the candidate set has no downstream opportunity. This is a stricter condition than being “online.” A page can be publicly accessible yet absent from an index, excluded by a search mode, outranked for the relevant query, or represented by a different canonical URL. Conversely, a domain may be well known while the particular page that contains the needed evidence is not selected. The distinction between a page and its domain becomes important later, but the operational point begins here: eligibility for retrieval precedes answer influence.
The candidate set is then ranked or reranked. Conventional search ranking may provide the initial ordering; a generative system may apply another selection process based on query intent, source quality, recency, or the needs of the answer it is trying to compose. The source's position can matter because only a limited number of documents, passages, or tokens fit into the model's context. A page can therefore pass the retrieval gate and still receive too little exposure for its decisive evidence to be used.
Context allocation is the bridge between “the engine found this page” and “the generator saw enough of it.” Imagine two sources both discussing a product's return policy. One is near the beginning of the allocated context and includes a short, explicit statement; the other appears later and buries the condition in a long narrative. Even before the generator chooses its wording, the sources have different opportunities to contribute. Context position is not the same as a publisher-controlled search rank, and an observed position effect is not proof that a page owner can directly set that position. It is a condition that a measurement should record.
The same distinction applies to crawling and indexing. A publisher may know that a page is live, and an analytics system may show that people can reach it, without knowing whether the engine's relevant search surface has a usable representation of it. The page may be indexed under a canonical address, represented through a different page, or encountered only for some formulations of a query. This is why a technical audit that reports accessibility is not yet a citation audit. Accessibility may be necessary, but the engine still decides whether the page is a candidate for the information need.
Query reformulation introduces another layer of distance between the user's language and the source's opportunity. A user might ask for “a reliable accounting platform for a small nonprofit,” while the engine searches for security, nonprofit pricing, reporting, ease of use, and independent comparisons. A page that repeats the original phrase but does not explain those decision criteria may be less useful than a page that addresses them plainly. The issue is not that the engine has found a secret synonym list for the brand to imitate. It is that source relevance is judged against the need the engine is trying to solve, which can be broader or more specific than the initial wording.
Retrieval and ranking should also be treated as different observations. Retrieval asks whether the page or domain entered the candidate universe. Ranking asks how it was ordered within that universe. A study that compares only cited pages may never observe the sources that were eligible but not selected. A study that compares Google position with AI citations may find a strong association without showing whether Google position is the direct cause, a proxy for domain presence, or a shared consequence of relevance and authority. The position-controlled SSRN study in the evidence base is valuable precisely because it asks what page-level features predict citation among comparable ranking bands, but its publisher-supplied abstract does not turn predictive performance into a causal guarantee.
Context can also be transformed before it reaches the generator. A system may pass whole pages, selected passages, summaries, or other representations of the source. A long page may contain the right evidence but distribute it across sections that are difficult to allocate together. A short page may be easy to pass through the context while omitting the qualification that makes its claim safe. These possibilities do not justify a universal rule about page length or formatting. They explain why a source-readiness judgment must ask whether the information needed for a decision is explicit, coherent, and attributable in the material the engine is likely to use.
The stages can fail in combinations. A brand may have high domain presence but poor page-level relevance; it may be retrieved for a navigational query but not a generic comparison; it may enter a context but lose the position contest; or it may receive ample context and still be omitted from the final citations. A no-citation answer does not reveal which failure occurred. The engine may not have searched, the source may not have been retrieved, the generator may have used it without attribution, or another source may have covered the same point more prominently. Treating all of these as “the page did not rank” creates a diagnosis before the evidence exists.
For that reason, a useful observation record follows the source through the pipeline. It notes the query as asked, whether search was active, the search mode and date, the candidate page or domain, its observed position, the context representation when visible, the final citation, and the support for the attributed claim. A brand cannot necessarily collect every field from a commercial product, but the fields it lacks are part of the uncertainty. The absence of a visible retrieval trace should reduce confidence in a causal explanation; it should not be silently filled with an assumption about what the engine did.
The sequence is a model of dependencies, not a claim that every commercial engine exposes these modules or performs an explicit internal support check. Some systems fetch pages during an interaction; others rely more heavily on prebuilt indexes. Some may reformulate a query, summarize documents, or allocate context in ways that are invisible to the publisher. The diagram's value is that it makes failure paths visible. A source can disappear before generation, after entering the candidate set, or after contributing material that is not attributed clearly.
The foundational GEO paper used a simplified but useful version of this arrangement. It supplied a simulated engine with the top five Google search results for each query and asked a language model to generate an answer grounded in those sources. Because the source set was supplied in advance, the experiment could isolate a later question: if a source is already present, do changes to its text alter its share of the answer? It intentionally held much of the upstream retrieval problem still.
After retrieval: generation, citation, and what the user receives
Once the generator receives context, it must turn several documents into one answer. This is not a simple act of copying the highest-ranked page. A response may combine a definition from one source, a statistic from another, and a qualification from a third. It may attach citations sentence by sentence, group them at the end of a passage, or omit an explicit attribution even when a source influenced the wording. The source's visibility inside the answer consequently has several dimensions: whether it is cited, where the citation appears, how much text is associated with it, whether its contribution is unique, and whether the answer actually depends on it.
The foundational GEO work made this difference measurable with metrics designed for generative responses. Its Position-Adjusted Word Count gave more weight to source-attributed material that appeared earlier in the answer. Its Subjective Impression measure asked a language model to assess factors including relevance, influence, uniqueness, position, amount of material, likelihood of a follow-up, and diversity. These measures were attempts to describe exposure in a synthesized answer where a linear result rank no longer captures what the user sees.
They should not be confused with a direct reading of the model's internal reasoning. A source can be important without receiving a long citation, and a long attributed passage can be copied without being especially useful. The survey calls the harder question absorption: did the source contribute facts, language, or structure to the answer? Citation is observable; absorption is usually inferred through similarity measures, ablations, or changes in the answer when the source is removed. A composite absorption score can be useful, but it remains a proxy unless a matched comparison shows what would have changed in the source's absence.
Fidelity is a separate test again. A citation can be attached to a sentence that sounds plausible but goes beyond the source. It can support only one clause of a compound claim. A source can be named accurately while its qualification is dropped, or a brand can be cited for a statement that reflects a competitor's evidence. The critical survey reports an earlier audit in which 51.5% of answer sentences were fully supported and 74.5% of citations supported the propositions associated with them across four historical generative engines. Those figures are not a current universal error rate, but they establish why citation count alone is an inadequate quality measure.
After the answer comes the user's response, which may be acceptance, comparison, another question, a visit to a source, or the end of the browsing session. A citation is not a guaranteed handoff. In a Pew Research Center analysis of Google searches, only about 1% of visits to pages with an AI Overview led to a visit to a source cited inside the Overview. Any result link was selected on 15% of visits without an Overview and about 8% of visits with one; sessions ended on 16% and 26% of pages respectively. The panel and model controls make these observations informative, but the study is observational and explicitly does not establish that the Overview caused every difference. The safe conclusion is narrower and more useful: exposure in an answer and a source visit are separate outcomes that must be measured separately.
These stages create two kinds of effect that are often mistaken for one another. A conditional post-retrieval effect asks whether changing a document alters its citation or answer share when the document is held inside the context. A total pipeline effect asks whether the changed document is more likely to be crawled, retrieved, allocated context, cited, faithfully represented, and followed by a valuable user action. The first can be positive while the second is neutral or negative.
The survey's account of SAGEO Arena illustrates the possibility. In a testbed that reinstated retrieval and reranking, body-only optimization reportedly reduced average top-20 presence by about 9%, top-10 presence after reranking by 16%, and final citation by 6%, across 171,003 documents and 2,700 queries. These are testbed-specific results from work the survey identifies as a preprint, so they are not a universal claim that rewriting harms retrieval. They are a warning about composition: an edit that makes an already supplied page easier for a generator to use can make the page less competitive before the generator ever sees it.
A simple probability decomposition makes the logic clear. The overall chance that a source is cited is the chance that search activates, multiplied by the chance that the source is retrieved once search activates, multiplied by the chance that it is cited once retrieved. A high value for the last term cannot compensate for a low value for the first two. If a study reports only the last term, it is answering a narrower question than a marketing team may think it is answering.
This is the operational meaning of a stage-based model. It does not tell a brand that it can control every engine decision. It tells the brand what to ask of every result: Was search active? Was the relevant page or merely the domain retrieved? How much context did it receive? Did the answer use the page, cite it, and preserve its qualifications? Did the user act? Each answer narrows the interpretation of the metric.
Once the pipeline is visible, page-level work can be judged more intelligently. The next issue is not how to decorate a page for a language model, but which forms of evidence and structure help a brand answer a real information need while remaining clear, verifiable, and useful to people.