How an Assistant Decides Whom to Cite, and Where Your Page Gets Eliminated

You ask ChatGPT which company to hire for something your company does. The answer comes back with three names, and none of them is yours.

The first impulse is to think the model does not know you. Sometimes that is exactly it. Most of the time, something else happened, and it started before any search.

The question you ask is not the question it answers

Before searching for anything, the system rewrites what you asked for.

A Google patent application describes the mechanism. The passage reads: “Additional/alternative queries may be, for instance, alternative query suggestions, supplemental queries, rewritten versions…”. ^{1 2}

Your question becomes several. Each one goes searching in its own way, and each one comes back with different candidates.

One observation about the document, and it matters. The application does not use the expression “query fan-out”. The name showed up later, in product communication. ³

Google confirmed the term on May 20, 2025, in the I/O announcement: “Under the hood, AI Mode uses our query fan-out technique”. ⁴ It confirmed the term in Portuguese as well, on September 8, 2025, when AI Mode reached Brazil. ⁵

This is called query fan-out.

Three memories in the same head

A model has three ways of knowing something, and they work differently.

The first is what it studied. Knowledge that went into training and stayed. The second is what it looks up on the spot, going after documents while it answers. The third is what it noted about you during the conversation.

Splitting this into three is mine, and it is didactic. The three mechanisms are real and distinct, and putting them together in one figure is my own organizing, to make the reading easier. ⁶

The second memory has a name and a paper. Patrick Lewis and others published “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks” at NeurIPS 2020, peer reviewed. ⁷

Out of this comes the most practical consequence in the article. What you publish today does not enter what the model already studied. It enters, at best, what the model looks up on the spot.

Three robots, three badges

Whose robots

OpenAI documents its crawlers on an official page. GPTBot collects public content that may be used to improve and train the models. OAI-SearchBot has a distinct function. ⁸

Perplexity does the same. In its documentation, PerplexityBot is described as “designed to surface and link websites in search results”, and the quotation was checked on August 1, 2026. ⁹

Google keeps the list of its own crawlers in Search Central, checked on the same date. ¹⁰

What this means for your robots.txt

Three companies, three sets of documentation, three groups of robots with separate functions.

The reading that each system has its own corpus and its own access policy is mine, and it rests on those three independent documents. None of them says this about the others. ¹¹

In practice, blocking a training robot and blocking a search robot are two different decisions. Whoever treats them as one usually vanishes from a place they wanted to be in.

Four gates in a row

I call gates the four steps between your page and the citation. The organization into four is mine. The steps themselves are known engineering: indexing, candidate retrieval, reranking, and generation with attribution. ¹²

Gate 1: being in that system’s corpus

The first filter has nothing to do with quality. It has to do with presence.

If that system’s crawler never came through your site, or came and was blocked, you are not a candidate. Symptom: the AI describes your market competently and never mentions your company, not even wrongly.

Gate 2: having a passage that answers on its own

The system does not retrieve your site. It retrieves a slice of text.

That slice has to make sense outside its context. If the paragraph starts with “this happens because”, it depends on the one before, and on its own it answers nothing. Symptom: the site has authority on the subject and the citations go to weaker pages with more direct answers.

Gate 3: surviving the reranking

The retrieved candidates go through a second queue, which puts them in order again.

Symptom: you show up in some answers and vanish from others, on the same subject, with no visible pattern.

Gate 4: being named in the attribution

The last gate is the cruelest, because by then you have already been read.

The system used your content to compose the answer and credited another source. It happens when the passage does not carry inside it the name of whoever is speaking. Symptom: the answer contains information that exists only on your site, and it cites another address.

The timeline, for anyone who wants to know since when

The patent application that describes the fan-out was filed on February 27, 2024. Publication came out on August 29 of the same year. ¹

A note on the ruler: that is an application, not a granted patent. An application shows intent to protect an idea. A grant shows that the office examined it and accepted it. Those are different rungs of evidence, and treating one as the other is the most common mistake made in citing Google patents.

AI Mode became generally available in the United States on May 20, 2025. It reached Brazilian Portuguese on September 8, 2025, three and a half months later. ¹³

The technical documentation on AI features and websites was updated on December 10, 2025. ¹⁴

Your Monday exercise

Ten minutes, three tabs.

Open ChatGPT, Gemini and Perplexity. In each one, ask for a recommendation of what your company sells, without naming the company. Write the way a customer would write.

Note down who got cited in your place. This is not about comparing yourself to them on quality. It is about finding out which gate you stopped at, and the symptoms of each gate are in the section above.

Save the three screenshots with the date. Without them, three months from now you will not be able to say whether anything moved.

Frequently asked questions

Can you reverse-engineer ChatGPT?

Not the way it usually gets sold. You can know the public steps, the documented robots, and what the companies declare officially. The split into three memories in this article is my own didactic organizing, not a description of the system’s internals. ⁶

Does blocking GPTBot take my company out of ChatGPT?

They are separate things. GPTBot collects content that may be used for training, and OAI-SearchBot has another function. Blocking one is not the same decision as blocking the other. ^{8 11}

Is “query fan-out” an official Google term?

It is an official product communication term, used in the announcement of May 20, 2025. The patent application that describes the mechanism does not use that expression. ^{3 4}

Does publishing today put me in the answer tomorrow?

Not in what the model already studied. Training has a cutoff. The possible route is what the model looks up at answer time, and that depends on the four gates in this article. ^{6 7}

Notes and sources

1. Patent application US 2024/0289407 A1, “Search with stateful chat”, Google LLC. Filed on February 27, 2024, published on August 29, 2024, priority of February 28, 2023. An application, not a granted patent.
2. Literal passage from the application in note 1, reproduced in the body.
3. The text of the application in note 1 does not use the expression “query fan-out”. The name comes from official product communication.
4. Google, “AI Mode in Google Search: Updates from Google I/O 2025”, blog.google, May 20, 2025. Literal quotation reproduced in the body.
5. Google Brasil, “Busca do Google: agora o Brasil já pode usar o Modo IA em português” (Google Search: Brazil can now use AI Mode in Portuguese), blog.google/intl/pt-br, September 8, 2025.
6. The three memories: the author’s didactic split. The three mechanisms are real and distinct, and the stitching between them is the author’s organizing.
7. Patrick Lewis et al., “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks”, NeurIPS 2020. Peer-reviewed paper.
8. OpenAI, official documentation of bots and user agents, platform.openai.com/docs/bots. GPTBot collects public content that may be used to improve and train the models; OAI-SearchBot has a distinct function.
9. Perplexity, “Bots”, docs.perplexity.ai/guides/bots. Literal quotation verified on August 1, 2026.
10. Google Search Central, “Google crawlers”. Literal quotations verified on August 1, 2026.
11. Corpus and access policy per system: the author’s inference, supported by notes 8, 9 and 10. Three official sets of documentation from three independent companies, none of them making claims about the others.
12. The four gates: the author’s organizing. The steps described, indexing, candidate retrieval, reranking, and generation with attribution, are known engineering.
13. Timeline: AI Mode generally available in the United States on May 20, 2025; in Brazilian Portuguese from September 8, 2025.
14. Google Search Central, “AI features and your website”, updated on December 10, 2025.