AI search · Citations by AI systems

Why can nobody guarantee that you will be cited by ChatGPT, Claude or Gemini?

An AI citation cannot be bought or reserved. It is decided for each answer by a model that chooses whether to search, what to read and whom to name. This page explains that mechanism, what a provider can honestly promise instead, and how to recognise a promise that nobody can keep. It applies to us just as it applies to any provider.

The fundamental question

Why can an agency not promise a citation?

Because there is nothing to reserve. An advertising banner occupies a position that exists before you arrive, that someone sells and whose nature the buyer knows. An AI answer does not work that way. It is produced when the engine receives a question, by a model that decides, according to the question asked, whether to search the web, which pages to retain and whom to name. No inventory is for sale, so no provider can sell you a share of it.

This is not our interpretation: the providers themselves describe it in the technical documentation read by developers and available to everyone. OpenAI’s documentation says that the model can choose whether or not to search the web according to the content of the question. Anthropic describes the same mechanism for Claude: the model decides when to search based on the question, then returns a final answer with cited sources. Writing about AI Overviews and AI Mode in its search engine, Google says that there are no additional requirements for appearing in them. Even a page that complies with all its recommendations is not guaranteed to be crawled, indexed or served. This documentation does not concern the Gemini application, whose internal operation Google does not publish.

Three providers’ documents lead to the same conclusion for a business leader: the decision to search and cite belongs to the model and is made answer by answer. These three documents are written for developers and do not describe the inside of consumer applications, which no provider publishes. A provider that guarantees you a citation is promising an action by a third party that it does not control.

What the readings show

On the same day, the same company is cited for one question and absent from the next.

The variation can be measured, and we publish it. On the same website, in the reading of 12 August 2026, one buyer question was cited 7 times out of 8 while a question asked much earlier in the buying journey remained at 0 times out of 8. That is why our readings are written as fractions, “cited this many times out of this many trials”, and never as yes or no. The details of each question, with its totals, are published in the Barometer.

This finding does not come from us alone. A guide hosted by France Num, the public portal for the digital transformation of businesses, says the same thing. It is not signed by the administration: its author is the managing director for France of a search optimisation agency. In the “Expert advice” section he writes that “answers sometimes remain unstable” and that “visibility measurement is still imperfect”. We cite it for what it is: the observation of a market professional, not a position taken by the state. The same guide recommends regularly testing the questions your customers might ask, then checking whether your company is cited in the answers.

A guarantee would therefore have to hold over something that changes from one question to the next and from one day to the next. This is what makes the promise impossible to keep, and repeated measurement indispensable. We apply this protocol to our own name and publish the readings, with their dates and zeros, in the HENUSSE Barometer.

A dated case

Can a citation disappear after it has been obtained?

Yes, and we have published a dated case. For one buyer question, in the diagnostic of 7 August 2026, our measurement instrument found that a website was cited as a source in all three passes. In the next reading, on the night of 14 to 15 August 2026, none of the tested surfaces cited it. We do not claim that the website was identical on both dates: it continued to publish between the two measurements. What the dated reading establishes is that the engine’s answer to the same frozen question had changed.

The two measurements are one week apart. No commercial guarantee survives this gap because a citation obtained one day is not permanently acquired the next. The details, the method and the limitations of this reading appear in the correlation study, with the dates.

This is also what separates a measurement from a screenshot. A screenshot shows one selected moment and says nothing about what happened the following week. A repeated, dated and retained reading does.

What can still be promised

What can honestly be promised to you, then?

Two things, and they are enough to judge a provider. First, the work: pages that genuinely answer your buyers’ questions, a website that the engines’ robots can read, and a presence in the third-party sources those engines consult. Second, the measurement must be verifiable point by point: the questions are written before the first test, the engine tested is named, the count is shown, and the answers are retained so that you can recount them.

These commitments have one property that a guarantee does not: they can be verified one by one, before and after signing. This is exactly what our measurement chartersets out in nine written rules. You can apply to it the only question that matters when faced with any commitment: how will I know if it has not been kept?

A provider can promise a third thing: to tell you to stop. A public stopping rule, backed by the measurements the provider produces, turns open-ended spending into a dated decision. Our own stopping rule, two consecutive quarters without progress after full deployment, is written in the questions to ask before signing.

In practice

How can you recognise an impossible promise in a quotation?

Four phrases should stop you because none can be honoured by anyone who does not control the model.

A guaranteed number of citations or a guaranteed position

The engine decides whether to cite you each time it produces an answer. Promising a number means promising the behaviour of a third party, even though our own readings vary from one date to the next on identical questions.

Fees paid according to the measured result

Someone who is paid according to the figure has an interest in choosing the questions that suit them. The conflict is structural, and it affects the measurement before it affects the work.

A score with no date, no question and no denominator

A visibility figure that does not say which questions were asked, which engine was tested or how many times cannot be challenged or repeated. It therefore provides no information.

A file or trick presented as the key

No file placed on a website can force an engine to cite anything. An offer that gives the leading role to a technical trick moves the conversation away from what can be verified.

These signals can be identified by reading a commercial proposal, before any meeting. The question “can you guarantee me a citation?” is therefore settled, and the honest answer will always be no. The useful question comes next: what should I require in writing, then? We have published the questions to ask, with our own answers: the questions to ask before signing with an AI search provider.

The public mechanisms

What makes a citation more likely?

A guarantee is impossible, but probability can be worked on. Five mechanisms are public and depend on no proprietary method. Google itself writes, about AI Overviews and AI Mode in its search engine, that no special optimisation is required to appear in them. What follows is therefore not a recipe, only what makes a page useful to someone looking for material with which to answer.

A page that engines can read. If the providers’ search robots cannot read your page, the question goes no further. This access cannot be guessed. It must be tested, and a block can go unnoticed for months. We explain the families of robots and what they change in the questions to ask before signing.
A direct answer to a real question. An engine looks for material with which to answer. A page that answers in its first lines, using a buyer’s words, is useful to it. A page that begins by introducing itself is not.
Verifiable, dated facts. A buyer looks for prices, figures, places, dates and conditions. A dated, quantified fact confirmed elsewhere can be cross-checked. An unverifiable claim cannot.
A named author who takes responsibility. A signed page, with an identifiable person and their real expertise, is distinct from a text with no author.
Third-party sources that confirm what you say. What other websites say about you matters as much as what you say about yourself. Work that stops at the boundaries of your website leaves out this half.

None of these five mechanisms is a lever on the machine. They make your company more useful to an answer, and that is all serious work can aim to do. What happens next remains the engine’s decision, which measurement observes instead of promising.

Frequently asked questions

Five questions to ask before signing

Can a provider guarantee me first place in ChatGPT?

No, and first place does not even exist in that form. An AI answer is not a ranked list: it is text produced on demand, which sometimes cites sources and sometimes cites none. Nobody sells a position in a text that has not yet been written.

What is the point of AI search work, then?

It makes your company usable by an engine looking for material with which to answer, and tells you where you stand. The work acts on what depends on you: your pages, your evidence and your presence in third-party sources. The measurement tells you what the engines do with them on a given date.

Why do my results change from one week to the next?

Because the engine decides for each answer whether to search and which pages to retain, and because our own readings vary from one date to the next on identical questions. That is why every question is asked several times and every figure carries its date. A reading is a snapshot, not something permanently acquired.

What if the measurement is zero?

A zero is written as a zero, and it becomes the starting point. An honest baseline is the only way to know later whether anything has moved. Our own Barometer publishes zeros alongside the same protocol applied to a website that is already indexed.

Can I check all of this myself?

Yes, and it is the simplest test to set a provider. Ask ChatGPT, Claude and Gemini a buying question, note who is cited, repeat the exercise the next day and compare. You will see the variation described here with your own eyes.

Sources

These pages from our own house are cited above:

The Honest Measurement Charter, the nine rules we impose on ourselves.
The HENUSSE Barometer, our dated readings, including the zeros, and the same protocol applied to a website with established exposure.
The correlation study between the ChatGPT application and our instrument, including its method and limitations.
The questions to ask before signing, with our own answers.

These providers’ documents are public and were accessed on 16 August 2026. Each quotation is followed by its translation:

Google Search Central, AI features and your website: “There are no additional requirements to appear in AI Overviews or AI Mode” (there are no additional requirements to appear in AI Overviews or AI Mode) and “Indexing and serving isn’t guaranteed” (indexing and serving are not guaranteed). The page itself states its scope: “This guide covers how AI features like AI Overviews and AI Mode work in Google Search” (this guide explains how AI features such as AI Overviews and AI Mode work in Google Search). It does not mention the Gemini application.
Anthropic, Web search tool: “Claude determines when to search based on the prompt” (Claude determines when to search based on the question), then “Claude provides a final response with cited sources” (Claude returns a final answer with its sources cited).
OpenAI, Web search: “the model can choose to search the web or not based on the content of the input prompt” (the model can choose whether or not to search the web according to the content of the question asked).

This public French source was accessed on 16 August 2026:

France Num, the government initiative for the digital transformation of very small, small and medium-sized businesses, led by the Directorate General for Enterprise, hosts the guide Optimisation for generative engines (GEO), published on 9 February 2026 and updated on 24 April 2026. This guide was produced by Jérémy Lacoste, managing director for France of the agency Eskimoz, which is listed as a France Num Activator; it therefore does not commit the administration. “Expert advice” section: “answers sometimes remain unstable, and visibility measurement is still imperfect”. “Evaluating your visibility in generative engines” section: “regularly test the questions your customers might ask” and “check whether your company, website or content is cited in the answers”.

Page published on 16 August 2026. The cited documents may change; their access date is written above. If one of them changes its position, this page will be corrected and the correction dated.