To measure your AI citations honestly, you need to define the questions, test conditions and interpretation rule before counting responses. We recommend retaining evidence, making failures visible and presenting each result with its scope so that you can verify the finding. The HENUSSE honest measurement charter sets the rules for our measurements on ChatGPT, Claude and Gemini, from the first test to the monthly report. You can consult this charter to examine our commitments to readers: every figure is dated, a zero result is recorded as zero and the test can be repeated.1

Key takeaways

  • Before the test, we write the buyer questions and fix their wording. This step comes before reading the responses. The published rule concerns the order of work: questions are chosen before the results are known, and the selected set is then used for the measurement.1
  • Present planned trials, completed trials and interpretable responses in separate columns. Record in writing how failures will be handled in the denominator. The reader must be able to trace the path from the collection plan to the final fraction, with missing runs visible in their proper place.
  • We recommend a grid that distinguishes the presence of the name, source attribution and prominence among the choices proposed. Define mention, citation and recommendation separately. This approach gives each finding its own meaning and allows a precise discussion of what the response offers the buyer.
  • We publish the baseline measurement before starting work, after fixing the questions. Our rule thus makes public the starting point used for subsequent comparisons. The firm’s stated reason is precise: to prevent the before-and-after comparison from being rewritten after the fact.2
  • Priority depends on the importance of the question to your business and the clarity of the finding. We recommend examining requests directly related to choosing a supplier first. For each selected topic, write down the decision being considered, the evidence expected and the next check needed before starting a correction.

Contents

Which questions should you choose to measure your AI citations?

Choose questions for measuring AI citations that correspond to your buyers’ needs and decisions. We favour natural wording whose intent is recorded before the test. Organise the set by research stage, then check that each question provides distinct information for the decision you want to make. Before the test, we write the buyer questions and fix their wording. This step comes before reading the responses. The published rule concerns the order of work: questions are chosen before the results are known, and the selected set is then used for the measurement.1

Choose questions based on your buyers’ decisions. A request for a supplier, a search for expertise and a comparison of approaches address different needs. We recommend specifying the need behind each wording so that you know which decision the measurement should inform and how much weight to give it. A set of questions benefits from covering several stages of research. Make room for identifying a need, examining options and choosing a provider. This organisation helps you read responses by intent, then choose the topics that deserve particular attention.

To examine how your business is discovered, favour wording that a buyer would use without knowing your name. Reserve questions containing your brand for a separate examination of your public identity. Each group of questions thus retains a clear purpose that you can explain before looking at the responses.

How can you identify the surface actually queried?

To identify the surface actually queried, describe the interface and conditions used for each run. A collection record must allow the exact question and corresponding response to be retrieved. We recommend keeping this description close to the result so that the reader knows precisely what the finding relates to. The charter states that our instrument has been querying gpt-5.6-sol since 16 August 2026. It announces a concordance check for this pair during the November 2026 measurement round. At the stage described on this page, the Claude and Gemini surfaces also remain to be checked in the same way.1

During the night of 14 to 15 August 2026, the ChatGPT application and our instrument, then running version gpt-5.5, returned the same presence verdict on 30 out of 30 questions with fixed wording. This agreement concerns the presence verdict for that set of questions, on that date and for that pair of surfaces.1 Our agreement study explicitly leaves open the case in which a brand is actually cited. The firm specifies that this check does not yet establish that the instrument then sees the same thing as the user. The scope of this study therefore remains limited when it comes to validating the observation of an actual citation.3

Since 26 August 2026, we have measured Google AI Overviews, the generated answers at the top of the results page, every quarter. This measurement follows a separate schedule. It is published separately and remains outside the weekly count, in accordance with the scope declared by the firm.4 To compare surfaces, retain the detailed results for each. A summary becomes useful when the reader can return to the questions and responses behind it. We favour examining results by surface and intent, followed by an explicit comparison when the collection conditions actually allow it.

How can you distinguish a citation, a mention and a recommendation?

To distinguish a citation, a mention and a recommendation, separately define source attribution, the presence of the name and prominence among the choices proposed. Then review each passage using this grid. We recommend retaining the context that justifies the classification, especially when several interpretations seem possible on first reading. Our assessment also details the company’s place in the reasoning and the exact quality of the mention. These observations are separate from its presence in the response. The measurement report also distinguishes a link to a verifiable source from the absence of such a link.5

The counting rule deserves to be defined before collection begins. Specify what you will accept as an attributable citation and how you will handle ambiguous cases. A response that is difficult to classify calls for a review of its context. Keep an explicit category for observations whose interpretation remains open to discussion. Useful evidence includes the context surrounding the detected presence. Retain enough of the response to verify which company the passage refers to and how the source is used. Review the displayed links as well. A simple occurrence of the name requires this verification before it is classified in your grid.

For the agreement measurement carried out during the night of 14 to 15 August 2026, we retained 30 timestamped screenshots, one per question, and 90 raw responses with their source URLs. The complete scoring grid is also part of the evidence. The entire set is dated and provided on request.3

How do you read a citation fraction?

Read a citation fraction by examining what the numerator and denominator count. Ask for the question, the scope and the rule applied to failed trials. We recommend keeping these details beside the result so that interpretation of the count remains tied to the conditions under which it was obtained. In a fictional example, a question receives 3 usable responses and the source being sought appears in 2 of them. The row shows “2/3”, with the fictional test date and the surface chosen for the exercise. A separate column specifies whether any planned trials failed before these responses.

A failed test remains visible in our report. The same rule applies to a test that could not be carried out: its status is recorded as it stands. The report thus preserves a record of these runs, with their status as failed or not performed.1 Present planned trials, completed trials and interpretable responses in separate columns. Record in writing how failures will be handled in the denominator. The reader must be able to trace the path from the collection plan to the final fraction, with missing runs visible in their proper place.

When a HENUSSE dossier includes an aggregate score, we publish its formula, its weights and every data point used alongside it. The charter makes this a condition for releasing the dossier. The summary figure is therefore accompanied by the calculation details that the firm states it makes public.1

How can you compare successive measurement reports?

Compare successive measurements using questions and conditions that have remained identical. Keep a register of protocol changes, then examine separately any series whose scope has changed. We recommend presenting differences with the corresponding evidence and recording the limits of comparison before proposing an explanation. We publish the baseline measurement before starting work, after fixing the questions. Our rule thus makes public the starting point used for subsequent comparisons. The firm’s stated reason is precise: to prevent the before-and-after comparison from being rewritten after the fact.2

New wording deserves its own version in the register. Keep the previous series for comparison and explain why a question was added, removed or corrected. This discipline allows you to track changes to the protocol while retaining a clear point of comparison for wording that has remained identical.

To review a change over time, keep your website change log separate from your observation log. You will be able to cautiously relate events to one another and examine what deserves another check. The explanation you adopt must remain proportionate to the available evidence, with other changes in conditions recorded. The firm presents the risk that a slow or difficult-to-crawl website may be retrieved less effectively as a working hypothesis. This qualification applies to both search engines and answer engines. The hypothetical nature of this relationship remains explicitly stated in our presentation.6

How should you decide what to do after a measurement?

Decide what to do after a measurement according to the importance of the questions to your business and the quality of the available evidence. A correction deserves a precise objective and a planned check; waiting also deserves a written reason. We recommend stating the intended decision before turning the measurement report into a programme of work. The AI Citation Diagnostic presents the findings question by question. We provide a dated dossier and a summary written by a human. The stated verdict may be to proceed, make corrections first, wait or refrain from buying: these possibilities are part of the deliverable described.7

Priority depends on the importance of the question to your business and the clarity of the finding. We recommend examining requests directly related to choosing a supplier first. For each selected topic, write down the decision being considered, the evidence expected and the next check needed before starting a correction.

To assess a measurement service, ask for an example of a dossier that can be reviewed. Examine the precision of the questions, the counting rule, the evidence retained and the handling of failures. Compare service categories according to this verification work and the decision the deliverable helps you make. We commit to the work, the method and the measurement. We refuse performance-based fees. The published reason concerns methodology: payment tied to results would encourage us to select questions that suit us rather than those of your buyers. A citation remains outside our commitment.7

A report for your management team benefits from linking each finding to a possible decision. Keep the wording restrained: what was observed, the limits of interpretation and the check to plan. Detailed evidence must remain accessible to allow a precise discussion when a finding is disputed.

Frequently asked questions

Why repeat measurement runs?

Repeating measurement runs allows you to examine the consistency of observations for the same question. We recommend fixing the repetition conditions before collection, then retaining each response with its status. You then have the evidence needed to discuss differences and decide what deserves an additional check.

How should failed tests be presented?

Present failed tests with an explicit status and their place in the collection plan. Also specify the rule used to handle them in the totals. We recommend keeping execution failures distinct from usable responses so that the reader can verify how the final result was constructed. A failed test remains visible in our report. The same rule applies to a test that could not be carried out: its status is recorded as it stands. The report thus preserves a record of these runs, with their status as failed or not performed.1

What evidence should you retain to review a result?

Retain the exact question, the complete response, the displayed links and the conditions of the run to review a result. The classification grid must remain linked to this evidence. We recommend an arrangement that lets you retrieve the evidence for each row directly, then understand the reason for the verdict reached. For the agreement measurement carried out during the night of 14 to 15 August 2026, we retained 30 timestamped screenshots, one per question, and 90 raw responses with their source URLs. The complete scoring grid is also part of the evidence. The entire set is dated and provided on request.3

What value does a zero result have in a measurement report?

A zero result serves as a useful starting point when it comes with a clear scope and evidence that can be reviewed. Its value lies in helping you choose which questions deserve further verification. We recommend examining the importance of these requests to your buyers before deciding on a correction or another collection round. The HENUSSE honest measurement charter sets the rules for our measurements on ChatGPT, Claude and Gemini, from the first test to the monthly report. You can consult this charter to examine our commitments to readers: every figure is dated, a zero result is recorded as zero and the test can be repeated.1

Sources

  1. HENUSSE (2026), « La charte de mesure honnête »
  2. Les questions à poser avant de signer avec un prestataire de référencement IA
  3. HENUSSE, Étude de concordance : l’application ChatGPT et notre instrument (2026)
  4. HENUSSE (2026), « Pourquoi ne pas mesurer Perplexity, Mistral ou Grok ? »
  5. Pourquoi un score sur 100 mesure-t-il mal la visibilité dans les réponses IA ?
  6. La vitesse des sites du marché, mesurée
  7. HENUSSE (2026), « Le Diagnostic de citation IA »

Théo Hénusse, Fondateur de HENUSSE