I start by asking for the measurement protocol and the responses that support the report. A figure alone is insufficient to compare two sets of measurements if the questions, engines or conditions have changed. I want a list of the exact questions, their language, the test date, the engine and model when known, and the surface used. Ask for this list in a form you can review and retain. It must make it possible to understand precisely which request each observation addresses.

A readable version of the protocol can bring together the shared conditions, followed by the changes specific to each period. Ask for a revision history rather than a method rewritten without a record of changes. I want to be able to retrieve the rules applied at the time of measurement, even when the provider has since changed how it works. Our charter stipulates fees that are independent of the measured figure. We justify this rule by the risk that payment based on results could encourage the provider to choose questions that suit it. The charter considers this incentive incompatible with honest measurement.1

Key takeaways

  • Keep this requirement in mind when reading a summary: each indicator must remain linked to its questions, its definition and the responses examined. Ask for the report to allow this return to the observations. You must be able to choose a passage to review and then retrieve the conditions under which it was obtained.
  • If the measurement comes from an API, it must say so. I do not present it as proof of what all users see in the consumer application. Ask the provider to describe the surface used and limit its conclusion to that surface, including the test conditions.
  • The definition of presence must be written before counting begins. Ask what the provider classifies as a citation, mention and recommendation, then have these definitions applied to the retained responses. I prefer a rule everyone can review to an appealing label whose meaning changes with the passage examined.
  • The report must give the number of tests performed, the responses received and the tests that could not be completed. A collection failure does not become an absence of citation. I ask for a separate status for each situation so that the reader can follow the progression from planned tests to responses actually examined.
  • When deciding what to do next, tie the provider’s recommendation to a specific observation. I ask which question warrants work, which response motivates that choice and how the next measurement will be interpreted. The proposed action must remain understandable, along with its rationale and the conditions for examining it.

Contents

How should you choose the questions to compare?

I choose the questions to compare according to the decision they need to inform, then ask for their exact wording to be retained. The selection must be explained before the test and distinguish the reader’s intentions so that each set of responses remains linked to the need that prompted the measurement. Our charter requires buyer questions to be written, dated and fixed before the first measurement. We therefore set their wording before knowing the answers. The rule excludes selecting questions after the test to favour the findings: that choice would make the scope depend on the results obtained.1

In our presentation of AI search optimisation, we distinguish a category of questions that arises early in the buyer’s thinking. We define it by its intent: understanding a problem. Choosing a supplier comes later; this first category corresponds to the stage when the person is still seeking to understand.2 In a fictional example, one report studies discovery questions, while another selects requests about choosing a supplier. Before comparing their results, ask for the questions to be classified by intent. Then compare the sets addressing the same need, keeping the questions specific to each report visible.

To compare two periods, I keep the same set of questions and flag changes to the protocol. Show questions that have been added, removed or reworded alongside the measurement record. Readers must be able to distinguish what falls within the retained scope and what belongs to the new scope.

What should you ask about the surface used for the measurement?

I ask for the exact name of the surface used for the measurement and a description of the session conditions. The provider must specify its access method, the engine queried and the model when known, then explain under what conditions a comparison with the other report can be considered. Our charter stipulates that the report must name the engine actually queried, ChatGPT, Claude or Gemini, and explain how it was queried. It requires this identification in the report and prohibits presenting an intermediary tool as though it were itself the engine to which the response is attributed.1

In our charter, we describe the selected surface as using a neutral session, without history or personalisation. The stated conditions must allow the test to be repeated under the same conditions and our counts to be checked as of the measurement date. This is the justification we give for choosing that surface.1 If the measurement comes from an API, it must say so. I do not present it as proof of what all users see in the consumer application. Ask the provider to describe the surface used and limit its conclusion to that surface, including the test conditions.

In our concordance study, we specify that the check on 15 August 2026 compared gpt-5.5 with the application. We explicitly limit its validity to that configuration. The page states that this historical check no longer describes the configuration in service: its scope remains tied to the version compared at the time.3

What supporting material should you request to verify the count?

To verify the count, I ask for the complete responses, the cited URLs and the classification rule applied to each observation. The report must make it possible to link this material to the exact question and the test conditions so that you can review a case and discuss its classification. We state that the raw responses from each measurement are retained and dated. They are provided on request with the report so that you can check our counts. The stated practice thus provides access to the responses supporting the count, alongside the report to which they relate.4

I also ask for the complete responses and cited URLs so that I can review the observations. A selected excerpt must remain linked to its complete response. You must be able to examine the wording, context and selected source before accepting the provider’s proposed classification of that response. Review deserves a written place in the protocol. Ask who examines the passages identified, how ambiguous cases are handled and how a classification error can be corrected. I prefer disagreements to remain visible until they are examined, together with the response concerned and the rule to be applied.

The presentation of findings must provide a way to discuss the observations. Ask how to flag a misclassified response and receive a correction to the report. I favour a discussion centred on an exact question, a retained response and a written definition: everyone can then examine the same material and explain their judgement.

The public off-site authority report is presented as a working document produced by HENUSSE for Real French. The page specifies that Real French is the family business used as a testing ground. HENUSSE states there that it tests its method in-house before offering it to others.5 To examine a reference presented by a provider, ask about the document’s provenance and any ties to the company observed. Work carried out on the provider’s own ground should be identified as such. I also want to know what kind of material is being presented: a protocol, a working document or a record of responses.

How should you ask for a definition of presence in a response?

I ask for a written definition of presence in a response, with separate rules for citation, mention and recommendation. Ask for the passage justifying each classification and the handling of ambiguous cases to be specified so that you can apply the same definition to the responses in the reports being compared. The definition of presence must be written before counting begins. Ask what the provider classifies as a citation, mention and recommendation, then have these definitions applied to the retained responses. I prefer a rule everyone can review to an appealing label whose meaning changes with the passage examined.

Our page on visibility scores distinguishes 4 observations when a response cites a company. The framework covers its presence in the response, its place in the reasoning, the exact quality of the mention and the presence or absence of a link to a verifiable source. Each observation is kept separate.6

How should you handle tests from which no response could be collected?

I ask for tests from which no response was collected to be treated as collection incidents, with the reason given when known. The report must distinguish them from the responses examined and explain how they are handled in the calculation so that the reader can trace the planned tests, those performed and the available observations. The report must give the number of tests performed, the responses received and the tests that could not be completed. A collection failure does not become an absence of citation. I ask for a separate status for each situation so that the reader can follow the progression from planned tests to responses actually examined.

In a fictional example, a measurement exercise plans 12 tests, receives 9 responses and encounters 3 collection failures. Ask for the report to retain these categories and explain which total is used in the calculation. The 3 tests that could not be completed must remain identifiable when you examine the 9 available responses and their classification. Our charter requires a date for every measurement. If the company appears nowhere, the report records the zero result as it stands. A measurement problem must also be reported as such: the rule prohibits replacing it with a convenient figure in the dossier delivered to the reader.1

Is an overall score enough to compare two reports?

An overall score alone is not enough to compare two reports using our assessment framework. This framework distinguishes the company’s presence, its place in the reasoning, the quality of the mention and the presence or absence of a verifiable link. Ask for the questions, the details of these observations and the calculation rules, then reread the answers to examine the differences. For each indicator, ask what is included in the total and what is excluded. The rules must specify how incomplete responses, tests that could not be completed and repetitions are handled. I want to be able to reconstruct the calculation from the retained observations without having to guess what the displayed denominator means.

Google’s help documentation defines how AI Overviews are counted in Search Console: clicking a link to an external page counts as a click. For position, an AI Overview occupies a single place in the results. All links in that overview receive the same position.7 When a report brings together several engines, also ask for a separate breakdown by engine. I want to see how each set of responses contributes to the total. Keep the categories clear and ask which grouping rules were chosen so that you can examine an individual observation without relying solely on the summary.

I expect a conclusion limited to what was observed. If the conditions do not allow a comparison, we say so. Ask the provider to separate the available observations, its interpretation and the questions still open. A useful conclusion must let you examine the path from responses to judgement.

Our AI Citation Diagnostic is presented question by question, showing the presence or absence of a citation of your company by the engines queried. The page states that a dated dossier and a human-written summary are provided. The verdict may be to proceed, make corrections first, wait or refrain from buying.8 When deciding what to do next, tie the provider’s recommendation to a specific observation. I ask which question warrants work, which response motivates that choice and how the next measurement will be interpreted. The proposed action must remain understandable, along with its rationale and the conditions for examining it.

Frequently asked questions

What protocol should you request to compare two AI visibility reports?

I ask for a written protocol specifying the questions, test conditions and counting rules for the AI visibility reports. The document must make it possible to retrieve the method applied to each measurement, with changes flagged before drawing a conclusion about how the measurements have evolved. A readable version of the protocol can bring together the shared conditions, followed by the changes specific to each period. Ask for a revision history rather than a method rewritten without a record of changes. I want to be able to retrieve the rules applied at the time of measurement, even when the provider has since changed how it works.

Can you compare measurements whose questions have changed?

I consider it acceptable to compare measurements whose questions have changed within a clearly identified common scope because it allows the same requests to be examined. I present new or reworded questions separately and ask for the conclusion to specify the limits of the comparison between periods. To compare two periods, I keep the same set of questions and flag changes to the protocol. Show questions that have been added, removed or reworded alongside the measurement record. Readers must be able to distinguish what falls within the retained scope and what belongs to the new scope.

What is a report worth when its complete responses are unavailable?

I consider a report without complete responses insufficient for verifying its classifications because reviewing the context is part of my examination. I ask for the available supporting material and the limits of its retention before using the summary, then reserve judgement on observations that remain impossible to review. We state that the raw responses from each measurement are retained and dated. They are provided on request with the report so that you can check our counts. The stated practice thus provides access to the responses supporting the count, alongside the report to which they relate.4

What conclusion should you request when measurement conditions differ?

When measurement conditions differ, I ask for a conclusion that describes the protocol differences and specifies what remains comparable. The provider must explain its judgement using the retained observations, then leave open the questions that the available measurements cannot resolve. I expect a conclusion limited to what was observed. If the conditions do not allow a comparison, we say so. Ask the provider to separate the available observations, its interpretation and the questions still open. A useful conclusion must let you examine the path from responses to judgement.

Sources

  1. La charte de mesure honnête
  2. Référencement IA
  3. HENUSSE, Étude de concordance : l’application ChatGPT et notre instrument (2026)
  4. Les questions à poser avant de signer avec un prestataire de référencement IA
  5. Exemple : un rapport d’autorité hors-site
  6. Pourquoi un score sur 100 mesure-t-il mal la visibilité dans les réponses IA ?
  7. What are impressions, position, and clicks?, Search Console Help
  8. Le Diagnostic de citation IA

Théo Hénusse, Founder of HENUSSE