How AI search works · Stage 1

What happens during query interpretation, the first stage of AI search?

Stage one is the only stage of the pipeline your website cannot touch. It is still worth understanding, because almost every unreliable AI-visibility test fails here rather than anywhere further down.

Deep dive on stage 111 min readUpdated 30 Aug 20267 sources, 6 primary and 1 vendor-described

Back to the roadmap Next: Query fan-out

On this page

The short answer

Query interpretation is the stage where an AI engine turns a conversational message into a machine-readable intent: what kind of question this is, which entities it refers to, which constraints apply such as locale and recency, and whether to run a search at all. It happens before any document is touched, so its only inputs are the user’s words, the conversation so far and the engine’s own priors, which is why nothing on your website reaches it and the one stage-one input you control is the prompt you choose when you measure. A 2026 measurement study of 55,393 trending Google queries recorded an AI Overview on 13.7% of them overall, rising to 64.7% of question-form queries against 9.5% of non-question queries.²

Key takeaways

Stage one runs on three inputs and you own none of them: the user’s words, the conversation history, and the engine’s trained priors. No page, schema block or file at a well-known path is read here.
Phrasing decides whether an AI answer appears at all. Question-shaped queries trigger one several times more often than non-question queries, in three studies using three different methods.
An ambiguous brand name is resolved to a single entity before any search runs. If that resolution goes the wrong way, stages two through five work flawlessly on a question that was never about you.
A follow-up is rewritten against the conversation before it is searched, so the same prompt in a fresh session and in a long thread are two different experiments.

Definition

What happens at the query-interpretation stage?

Query interpretation is the conversion of a human message into a machine intent that the rest of the pipeline can act on. Four things come out of it: an intent type, a set of resolved entities, a set of constraints, and a routing decision about whether to retrieve anything at all. The distinction that matters is that your prompt is not a query. “Is this worth it for a two-person team?” is a conversational turn; what stage two receives is a structured intent, and what stage three receives is a set of search strings that may share very few words with what you typed.

Google describes the step immediately downstream in its own documentation: AI Overviews and AI Mode “may use a ‘query fan-out’ technique” that issues “multiple related searches across subtopics and data sources” to develop a response.¹ The subtopics that fan-out expands into are drawn from the interpretation, not from the literal string. That is the whole reason a page can be retrieved for a question that shares no vocabulary with it, and the reason a page can miss a question it seems to answer perfectly.

The routing decision

Does every question even trigger an AI answer?

No. On Google, whether an AI Overview appears at all is decided at this stage, and the rate depends heavily on how the query is phrased. A 2026 preprint tracking 55,393 trending queries over 40 days measured AI Overview activation at 64.7% for queries phrased as questions against 9.5% for non-question queries, a 6.8-fold difference; the all-query rate was 13.7%, but that figure already contains the questions, so the question-versus-non-question pair is the sharper contrast.² An 11,500-query study presented at SIGIR 2026 using a representative rather than trending sample found an AI Overview shown for 51.5% of queries.³ The two numbers are not in conflict; they sample different populations, which is itself the lesson about reading this literature.

A large vendor study drawing on 146 million search result pages, published January 2026, reports the same direction from a third method: AI Overviews appeared on 57.9% of question queries versus 15.5% of non-question result pages, and 99.9% of the keywords that triggered them were informational.⁶ Three independent methods, one direction. This is the most robust structural finding in the whole field, and it is a stage-one finding: it is about how the query is shaped, not about how your page is written.

The operational consequence is a measurement one. If your test prompt does not trigger an AI answer, you have learned something about your prompt, not about your visibility. Record the trigger rate separately from the citation rate, or you will keep averaging zeros that never had a chance to be anything else.

Ambiguity

How does the engine decide which entity you meant?

The engine picks one reading of each ambiguous name and the entire rest of the pipeline inherits that choice, with no appeal and usually no visible signal that a choice was made. A product name that is also a common noun, a company that shares a name with a larger company in another industry, an acronym with three meanings in adjacent industries: each of these forces a resolution at stage one, and if the resolution goes the other way, stages two through five run flawlessly on the wrong entity. Your page was not rejected. It was never in the running, because the question the engine actually searched was about something else.

Here the evidence runs out, and it is worth being explicit about that. No engine publishes its entity resolution, and there is no public benchmark measuring how often AI search resolves a brand name incorrectly. What follows is inference from how retrieval works rather than a measured lever: use one consistent name for the product everywhere, state the category in a flat sentence on your own pages rather than implying it, and make sure the third-party pages that describe you use the same name and the same category words. If the corpus consistently associates a name with a category, resolution has more to work with. That is a plausible mechanism, not a proven one, and anyone selling “entity optimisation” with a percentage attached is selling you a number nobody has measured.

Context carry-over

How does conversation history change the query?

A follow-up question is rewritten against the conversation before it is searched, so the query that actually ran is not the query you typed.

The two columns below are the 2026 critical survey’s minimum study checklist: what to record before a run, and what quietly invalidates one.⁴

Record before every test run

The exact product and mode: chat, search mode, or AI Mode
The model version, where the interface exposes it
Date, time and locale
Whether search was enabled on that turn
A fresh, logged-out session with no stored memory
The prompt verbatim, unchanged between runs

Quietly invalidates the run

Asking the next test in the same thread as the last one
Letting account memory or history personalise the answer
Rewording the prompt slightly between runs
One run per engine, then reading it as a standing
Averaging several engines into one visibility score
Reading “no AI answer appeared” as “we are invisible”

Ask “what about pricing?” after a question about three products and the engine resolves the pronoun, the subject and the comparison set from the conversation, then searches that rewritten version rather than your six words. That is helpful for a user and fatal for a test: two people running the same script in two threads have run two different experiments. It is also why the same prompt in a fresh session and in a long thread can produce answers with almost no sources in common.

Stochasticity compounds it. The 2026 critical survey reports Jaccard overlap of 0.34–0.42 between repeat runs of the same query on commercial engines, and 9–28% changes in repeated decisions on surfaces where temperature can be controlled.⁴ A 2026 statistical treatment that sampled several engines daily over nine days and at ten-minute intervals concluded that citation distributions follow a power law with substantial variability and that “many apparent differences between domains fall within the noise floor of the measurement process”.⁵ Control the session, then repeat the run, then report a distribution.

No leverage

Why does nothing you own influence stage one?

Because stage one has exactly three inputs and you own none of them: the user’s words, the conversation history, and the engine’s own trained priors about language and entities. No document has been fetched yet. There is no slot in this stage where a page, a schema block, a robots directive or a file at a well-known path could be read, and Google’s own documentation confirms the negative on the file front: “you don’t need to create new machine readable files, AI text files, or markup to appear in these features.”¹

This is a useful diagnostic to keep. When you meet a GEO claim, ask which stage it acts on. If the honest answer is stage one, meaning anything phrased as making the engine “understand your brand” before it searches, then either the claim is misdescribed work on a later stage, or there is no mechanism behind it. The pipeline does not have a door there.

Application

What can you actually do about stage one?

Choose the prompts you measure as carefully as you would choose keywords, because the prompt is the only stage-one input in your hands. Write them the way a buyer types, in full sentences, since question-shaped queries trigger AI answers at roughly four to seven times the rate of non-question queries. Cover the whole decision, not just the branded question: what the category is, what the options are, what it costs, what goes wrong. A prompt set made only of “best X tool” queries measures one narrow slice of a buying conversation.

Then treat the trigger decision as data rather than noise. Log whether an AI answer appeared at all, separately from whether you were cited in it, and separately again from whether you were described accurately. Those are three different failures with three different fixes, and the survey’s own recommendation is to “measure retrieval, citation, and fidelity separately.”⁴ Everything else in the how AI search works stage is about the stages after this one, where you actually have inputs.

The honest limit of this article

Stage one is the least-documented stage in the pipeline, and this article is correspondingly the most inferential in the set. No engine publishes its query understanding; there is no public benchmark of how often AI search resolves a brand name to the wrong entity; and the naming advice in the entity section is a plausible mechanism with no measured effect size behind it. What is solidly evidenced here is narrow: the activation rates for question-shaped queries, and the run-to-run variance figures. Everything else is a working model, and you should treat a confident number about stage one from anybody as a claim about something nobody has measured.

Where a product fits, and where it does not

The whole method here is free and manual: write 20 prompts a buyer would actually type, open a fresh logged-out session per engine, paste them verbatim, and record three columns: did an AI answer appear, were you cited, were you described correctly. Repeat on three separate days, because one run is a sample of one. A spreadsheet does this perfectly well; the only thing that breaks is your willingness to keep doing it weekly. Bavior automates that repetition, running a fixed prompt set across five engines on a schedule and recording the cited sources per run so you see a distribution instead of a screenshot, and where a cited source is a live discussion thread it drafts a reply on an account you control, which you approve before anything posts. It cannot influence stage one, because nothing can: it does not change how an engine parses a question, resolves your brand name, or decides whether to search. The free AI visibility check and the free GEO audit run without a paid plan; paid plans are from $99/mo billed monthly, or $79.17/mo billed annually (as of 29 Aug 2026).

Sources, all checked 29 Aug 2026

Google Search Central, “AI features and your website” (query fan-out, no special files or markup; first-party): developers.google.com/search/docs/appearance/ai-features
Xu, Iqbal & Montgomery, 2026; 55,393 trending queries over 40 days, collected 13 March – 21 April 2026; AI Overview activation 64.7% for question-phrased queries against 9.5% for non-question queries, a stated 6.8× difference; 13.7% across all queries (preprint): arxiv.org/abs/2605.14021
Grossman et al., SIGIR 2026; 11,500 queries, AI Overview shown for 51.5% of a representative sample: arxiv.org/abs/2604.27790
“Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023–2026)”, 15 Jul 2026, arXiv:2607.14035 (survey preprint): arxiv.org/abs/2607.14035
Sielinski, “Quantifying Uncertainty in AI Visibility: A Statistical Framework for Generative Search Measurement”, Mar 2026, arXiv:2603.08924 (preprint): arxiv.org/abs/2603.08924
AI Overview trigger study, Jan 2026: 146 million search result pages, 76.7 million AI Overviews; AI Overviews on 57.9% of question queries versus 15.5% of non-question result pages, 99.9% of triggering keywords informational. Vendor-published; described rather than linked, per this curriculum’s sourcing rule.
Aggarwal et al., “GEO: Generative Engine Optimization”, KDD 2024, arXiv:2311.09735 (the retrieval-plus-synthesis framing this pipeline model starts from): arxiv.org/abs/2311.09735

Continue reading

Learn GEO GEO roadmap Learn GEO Previous: The five-stage pipeline Learn GEO Next: Query fan-out

FAQ

Frequently asked questions.

Why did an AI engine answer a completely different question than the one I asked?

Because stage one resolved your question into an intent, and the resolution went somewhere you did not intend. Ambiguous product names, acronyms shared across industries, and pronouns inherited from earlier in the thread all get settled at interpretation time, before any document is fetched, and every later stage inherits that decision with no visible signal that a choice was made. The answer is then built correctly for the wrong question. Rerunning the prompt in a fresh session with the entity named unambiguously is the fastest way to tell interpretation failure apart from a retrieval failure.

Do question-shaped prompts really trigger AI answers more often?

Yes, and it is the best-corroborated structural finding in this field, from three independent methods. A 2026 preprint tracking 55,393 trending queries over 40 days measured Google AI Overview activation at 64.7% for queries phrased as questions against 9.5% for non-question queries, a 6.8-fold difference. An 11,500-query SIGIR 2026 study using a representative sample found AI Overviews on 51.5% of queries. A January 2026 vendor study of 146 million search result pages reported 57.9% on question queries versus 15.5% on non-question pages. The samples differ, the direction does not.

Can I make an AI engine understand my brand before it searches?

There is no documented mechanism for it, because stage one takes no input from your site. Interpretation runs on the user's words, the conversation history and the engine's own trained priors, before any document is fetched, and there is no slot where a page, a schema block or a file at a well-known path is read, and Google's documentation states directly that no new machine-readable files or markup are needed to appear in its AI features. Consistent naming across the web plausibly helps resolution, but no one has published an effect size for it, so treat any percentage attached to that claim as unmeasured.

Why does the same prompt give different sources in a new chat versus an old thread?

Because a follow-up is rewritten against the conversation before it is searched, so the query that actually ran is not the query you typed. Pronouns, comparison sets and implied subjects are all resolved from history at stage one. Add the engine's own run-to-run variance, since the 2026 critical survey reports Jaccard overlap of 0.34–0.42 between repeat runs of the same query on commercial engines, and two threads become two different experiments. Test in fresh, logged-out sessions with memory off, and repeat each prompt several times before reading anything into the result.

Bavior Editorial

The team that researches and maintains Bavior’s writing on Reddit marketing and AI search visibility. Every figure here is attributed to a named source with the date it was checked, and none of our links are affiliate links.

Found a number that looks wrong? Tell us and we will re-check it: support@bavior.com