How AI search works · Deep dive

What is an AI answer engine, and how does it decide what to cite?

Three different things get called “AI search”, and only one of them responds to work you do on your own pages. Sorting them out is the cheapest hour you will spend on this subject.

Deep dive on stage zero12 min readUpdated 30 Aug 202611 sources, 10 primary, 1 vendor-described

Back to the roadmap Next: The five-stage pipeline

On this page

The short answer

An AI answer engine is a retrieval system with a language model attached to its output: it turns your question into searches, reads the documents that come back, writes an answer over them, and attaches links to some of the claims. That is what separates it from a traditional results page, which ranks links and leaves the reading to you, and on the largest surface the written answer is now the default rather than the exception: an 11,500-query study presented at SIGIR 2026 found Google generating an AI Overview above the organic results for 51.5% of representative real-user queries.⁹

Key takeaways

Seven systems are in scope here: ChatGPT and ChatGPT Search, Perplexity, Google AI Overviews, Google AI Mode, Gemini, Microsoft Copilot, and Claude with search enabled. Every one of them retrieves first and writes second.
A results page ranks links and leaves you to read them; an answer engine writes the prose and footnotes part of it. The source pools differ measurably: 29.8% of the domains an AI Overview cites appear nowhere on the first page of results shown beside it.¹¹
A model answering with search switched off is not an answer engine. Its weights froze months before you published, so nothing you ship this quarter can reach that answer.
An agentic browser is not one either. It drives a real Chromium session on a person’s behalf, so it reads pages a crawler cannot and arrives in your logs looking like a visitor.
A citation is a link the engine attaches to a claim, not a recommendation and not a ranking position. Being cited and being named in the answer text are separate outcomes, worth very different amounts to a brand.

Definition

What is an AI answer engine, exactly?

An AI answer engine is a retrieval system with a language model attached to its output. It does not answer from memory: it issues searches, reads the documents it gets back, writes a summary over them, and attaches links to some of the sentences it wrote. The academic literature calls the same object a generative engine and models it as exactly that pair, a retrieval component followed by a synthesis component, and that pair is the entire reason this subject exists separately from search engine optimisation.²

Seven systems are in scope for this curriculum: ChatGPT and ChatGPT Search, Perplexity, Google AI Overviews, Google AI Mode, Gemini, Microsoft Copilot, and Claude with search enabled. For its two surfaces Google states the entry requirement in a single sentence: to be eligible as a supporting link in AI Overviews or AI Mode, “a page must be indexed and eligible to be shown in Google Search with a snippet, fulfilling the Search technical requirements.”¹ That is the only authoritative statement anyone has about eligibility on the largest of these surfaces, and it says the entry ticket is ordinary search hygiene.

Note the word eligible. Being retrievable makes you a candidate. It does not make you a citation, and the rest of the how AI search works stage is about the three stages that stand between the two.

Architecture

What does each half of the architecture actually do?

The retrieval half decides which documents the engine is allowed to see; the synthesis half decides what it says about them. Every intervention available to you acts on one half or the other, and the two respond to completely different work. Retrieval responds to being indexed, being fetchable, and being genuinely relevant to a query the engine actually issued. Synthesis responds to how a surviving passage is worded, and only to passages that already survived.

This is not academic hair-splitting. It is the reason two well-known results look like they contradict each other and do not. The 2024 KDD paper that named generative engine optimisation measured page rewrites against a purpose-built engine that fetched the top five Google results and had GPT-3.5 write an answer over them; the optimised page was already inside the five-document context before the experiment started.² A 2026 KDD benchmark that reinstated retrieval and reranking over 171,003 documents and 2,700 queries found body-only optimisation reducing final citation by about 6% and average top-20 presence by about 9%.³ Neither measurement is wrong. One measures the synthesis half in isolation; the other measures the composition of both halves, and the composition can be negative.

Not an answer engine

How is this different from a model answering without search?

A model answering without search is doing recall from its training data, and nothing you publish this quarter can reach it. The weights were frozen at a cutoff that predates your work by months; the next training run is not something you can schedule, verify or appeal; and there is no documented mechanism by which a page published today enters a model’s parameters on a timeline you could plan around. When the same product runs with search enabled it becomes an answer engine, because it re-reads the live web at question time. That is the only reason any of the work in this curriculum has a mechanism at all.

The practical test is whether the answer carries source links, and it is imperfect: most chat products decide per turn whether to search, and the user is not always told which happened. An answer with no links is either a memory answer or one whose engine chose not to show its work; either way it sits outside the reach of on-page work. This matters for measurement more than it matters for writing: a visibility test run in a product that silently answered from memory has measured the training corpus, not your site.

It also gives you a clean question for anyone selling influence over what a model “knows” about you: which stage does that act on, and how would it be verified?

Terms

What counts as a citation in an AI answer?

A citation is a link the engine attaches to a claim in its answer, and nothing more. It is not a recommendation, not a ranking position, and not a promise that the sentence beside it reflects what your page says. Being cited and being named in the answer text are different outcomes, and only one of them does much for a brand. A vendor study published in June 2026, covering 115 prompts and 3,981 domain appearances across four platforms in 14 countries, found roughly 62% of source appearances were links the answer text never named in prose, and reported an almost inverted profile between two engines: one with a high mention rate and a low citation rate, the other the reverse.⁵ The sample is small enough that the direction is the finding and the decimals are not.

Counting is also engine-specific. In a 602-prompt academic sample published in April 2026, the mean was 6.88 citations per ChatGPT response, 12.06 for Google’s AI surfaces and 16.35 for Perplexity; mean influence per fetched page was 0.2713, 0.0584 and 0.0646 respectively.⁶ One ChatGPT citation and one Perplexity citation are not the same unit of anything. A report that adds them together has produced a number with no referent.

The third thing

Is an agentic browser an AI answer engine?

No. An agentic browser drives a real Chromium session on a person’s behalf instead of retrieving from an index, so it reads pages a crawler cannot and arrives in your logs looking like a visitor rather than a bot.

Three systems get called “AI search”. They read different things, and they leave different traces in your logs.

	AI answer engine	Model without search	Agentic browser
What it reads	the live web, retrieved at question time	its training data, frozen at a cutoff	your page in a real Chromium session
Attaches source links	Yes, to some claims	No	Yes, to what it visited
Reached by work you ship this quarter	Yes, through stages 2–5	No, not within a release cycle	Yes, JavaScript included
How it appears in your logs	a declared crawler token, or nothing	nothing	close to indistinguishable from a person

An agentic browser is a browser, not a crawler, and the difference decides what it can read. These products are Chromium builds driving a real browser session on the user’s behalf, so they execute JavaScript the way any browser does. The classic AI crawlers apparently do not: the only published measurement, from December 2024, concluded that none of the major AI crawlers render JavaScript, naming Applebot and Gemini as exceptions because both render through browser infrastructure.⁷ Nobody has published a replication since, and no AI vendor documents rendering either way, so the defensible sentence is that the only measurement we have found showed no execution, not that AI cannot render JavaScript in 2026.

The Tow Center at Columbia tested the consequence in October 2025 and found agentic browsers retrieved a 9,000-word subscriber-only article that the standard chat interfaces of the same two companies could not, because the publisher had blocked those crawlers but had not blocked Chrome.⁸ The same investigation found that when an agent genuinely cannot reach a page, it assembles a composite from syndicated copies, social posts and third-party coverage instead. Blocking a crawler does not stop your content being summarised. It mostly changes who gets the link.

Divergence

Do the engines behave the same way as each other?

AI answer engines do not behave the same way as each other, and the disagreement is wide enough that a single blended “AI visibility” score discards the most decision-relevant part of the data. An 11,500-query study presented at SIGIR 2026 measured URL-level Jaccard similarity of 0.11–0.18 between Google organic results, Google AI Overviews and Gemini, three surfaces from one company.⁹ An audit of 672 responses across two assistants found 355 unique domains, of which only 26% were cited by both.¹⁰ The 2026 critical survey draws the conclusion in one line: these results “refute the notion of a global GEO ranking. Visibility is indexed by engine and surface.”⁴

So “AI answer engine” names a shape, not a standard: every system on the list retrieves and then writes, and beyond that they differ in index, crawler token, source count and weighting. Report per engine or do not report.

The honest limit of this article

“Answer engine” is a category this curriculum drew, not a term the engines use consistently about themselves, and the boundary is getting blurrier rather than sharper. A chat product decides per turn whether to search, so the same product is an answer engine in one message and a memory model in the next, and the interface does not always tell you which. No engine publishes its retrieval architecture. And the 62% ghost-citation figure above comes from a 115-prompt vendor sample: enough to establish that the gap between mention and citation is real and large, nowhere near enough to support the second decimal place.

Where a product fits, and where it does not

You can do the whole of this by hand and you should try it once: open a fresh logged-out session on each engine, ask the five questions a buyer would actually type, and write down every URL that came back and whether your brand was named in prose as well as linked. Repeat it on three different days, because one run is a sample of one. Bavior automates only the tedious half of that loop: it runs a fixed prompt set across five engines on a schedule so you get a distribution instead of a screenshot, records which sources each answer cited, and where a cited source is a live discussion thread it drafts a reply for that thread on an account you control, which you approve, edit or reject before anything posts. It does nothing about what a model recalls without search, it cannot make an engine cite you, and it has no product for the interpretation stage because nothing does. 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 30 Aug 2026

Google Search Central, “AI features and your website” (first-party): developers.google.com/search/docs/appearance/ai-features
Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan, Deshpande, “GEO: Generative Engine Optimization”, KDD 2024, arXiv:2311.09735: arxiv.org/abs/2311.09735
Kim et al., “SAGEO Arena: A Realistic Environment for Evaluating Search-Augmented Generative Engine Optimization”, KDD 2026 (arXiv:2602.12187v2, 7 Aug 2026): arxiv.org/abs/2602.12187
“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
Ghost-citation study, Jun 2026: 115 prompts, 3,981 domain appearances, 4 platforms, 14 countries; ~62% of citations were links whose brand was never named in the answer text. Vendor-published; described rather than linked, per this curriculum’s sourcing rule.
Zhang, He, Yao, “From Citation Selection to Citation Absorption: A Measurement Framework for GEO Across AI Search Platforms”, 28 Apr 2026, arXiv:2604.25707 (preprint, open dataset): arxiv.org/abs/2604.25707
Zecchini, Moore, Ubl, Siddle, “The rise of the AI crawler”, Vercel, 17 Dec 2024 (infrastructure first-party log data): vercel.com/blog/the-rise-of-the-ai-crawler
Tow Center for Digital Journalism, Columbia, “How AI browsers sneak past blockers and paywalls”, 30 Oct 2025: cjr.org
Grossman et al., SIGIR 2026, 11,500 queries; URL-level Jaccard 0.11–0.18 across Google organic, AI Overviews and Gemini: arxiv.org/abs/2604.27790
Li & Sinnamon, 2024; 672 responses analysed across two assistants, 355 unique domains, 26% cited by both
Xu, Iqbal, Montgomery, “Measuring Google AI Overviews: Activation, Source Quality, Claim Fidelity, and Publisher Impact”, 13 May 2026, arXiv:2605.14021 (preprint); 55,393 trending queries, 29.8% of AI Overview source domains absent from the matching first page of results: arxiv.org/abs/2605.14021

Continue reading

Learn GEO GEO roadmap Learn GEO Next: The five-stage pipeline

FAQ

Frequently asked questions.

Is ChatGPT an AI answer engine?

ChatGPT is an answer engine on the turns where it searches, and a memory model on the turns where it does not. The product decides per message whether to retrieve, so the same interface produces both kinds of answer and does not always make the difference obvious. The reliable tell is whether source links are attached to the response. This matters for measurement: a visibility test that happened to run on a no-search turn has measured the training corpus, not the live web, and nothing you publish can move that result.

What is the difference between a mention and a citation in an AI answer?

A citation is a link the engine attaches to a claim; a mention is your brand name appearing in the answer text. They come apart constantly. A June 2026 vendor study of 115 prompts and 3,981 domain appearances across four platforms found roughly 62% of source appearances were links whose brand the answer never named in prose, and that two engines had almost inverted profiles: one mentioning brands often while linking them rarely, the other the reverse. For a brand the two outcomes are worth very different amounts, so count them separately.

Does an AI answer engine visit my site, or does it use a search index?

Both patterns exist, and which one applies changes what you can control. Google's AI Overviews and AI Mode draw on the existing Search index, so the relevant fetch was Googlebot's and may have happened weeks earlier; Google states that a page "must be indexed and eligible to be shown in Google Search with a snippet" to be eligible as a supporting link. Other engines run their own crawler plus a separate user-triggered fetcher that pulls a page at the moment someone asks about it, documented under different user-agent tokens. An agentic browser is a third case: it loads your page in a real browser session.

If I block AI crawlers, will my content stay out of AI answers?

No, and the Tow Center at Columbia demonstrated both halves of why in October 2025. Agentic browsers retrieved a subscriber-only 9,000-word article that the same companies' standard chat interfaces could not, because the publisher had blocked the crawlers but not Chrome, and an agent driving a real browser session is close to indistinguishable from a human visitor. And when an agent genuinely cannot reach a page, the same investigation found it assembles a composite from syndicated copies and third-party coverage instead. Blocking changes who gets the link far more than it changes whether you are summarised.

Why do different AI engines cite completely different sources for the same question?

Because they run different indexes, different retrieval steps and different selection rules, and the measured disagreement is large. An 11,500-query study presented at SIGIR 2026 found URL-level Jaccard similarity of 0.11–0.18 between Google organic results, AI Overviews and Gemini, three surfaces from one company. An earlier audit of 672 responses across two assistants found 355 unique domains with only 26% cited by both. The practical consequence is that a single blended visibility score averages away the only part of the data you could act on.

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