Not necessarily. Before concluding that you have gained commercial visibility, I recommend reading the full response and the question asked. I recommend starting with the question asked. To interpret a presence, ask yourself what the person wanted to achieve: understand a concept, examine an option or find a service provider. This reading gives meaning to the passage that relates to you and helps you choose which observations are worth keeping.

Key takeaways

  • Classify each appearance according to the role assigned to your company in the response.
  • Keep the exact question and the complete answer together.
  • Attach the conditions under which the observation was recorded to the saved question and answer.
  • Ask the prospect about their journey before attributing an enquiry to an AI response.

Contents

To interpret a link to your site, first look for the information that the page supports in the AI response. I recommend keeping the relevant passage and the exact address, then checking whether the response also suggests your company for the need expressed by the person.

In a fictional example, a response about choosing a material links to an information sheet published by your company. The measurement record classifies this link as a source for the explanation. To classify a service provider recommendation, the reading separately looks for a passage suggesting your company to carry out the requested work.

In our measurement on 16 August 2026, using the Claude API with web search, the site studied appeared among the sources in 8 out of 30 responses. The name appeared in the body of 7 out of 30 responses. These counts covered the same brand and distinguished the 2 observations. These figures describe this measurement. Classifying a recommendation addressed to a buyer requires examining whether the response suggests the company for the stated need and explains that suggestion.1 To classify a named company’s appearance, record the complete sentence in which its name appears. Examine the role assigned to the company: an example, the author of a resource, a service provider to consider or a suggested choice. Keep any qualifications in the same excerpt so that the classification remains faithful to the passage.

Check that the name and link do refer to your company. A precise address, a business activity or a location can serve as checkpoints in your reading. If the identity remains uncertain, record that qualification in the measurement record before accepting the appearance as an attributable observation.

How can we recognise a recommendation addressed to a buyer?

To recognise a recommendation, look for a suggested choice linked to the buyer’s need. I would use this classification when the response suggests your company for that need and explains the relevance of the suggestion; I would also retain the conditions and qualifications that frame the choice. In a fictional example, a response to a request for a service provider names your company among the options to consider and specifies the need it meets. The measurement record retains this suggestion, its qualifications and the associated link. The classification applies to this response and this specific request.

I reserve the classification of recommendation for a suggestion that addresses the buyer’s need. Look for what the response invites the reader to choose and why. When the passage remains ambiguous, preserve that ambiguity in your interpretation rather than attributing commercial intent to a mere appearance. When a response suggests your company for a need, also examine the description of your business. Record the services attributed to it and the qualifications expressed, then compare them with what you actually offer. I recommend treating an inaccurate description as a point to verify before presenting the passage as valuable.

What evidence should we keep with an AI response?

Keep the full AI response with the exact question and the observation conditions. I recommend including the passage concerning your company, the displayed link and the classification rule so that another person can review the record and understand the decision made. For each answer, I want to be able to retrieve the exact question, the date, the engine, the model and the surface used. If the model’s identity is unavailable, state this clearly without inventing a model. Add the available settings and any specific test conditions. This record allows the observation to be reviewed with knowledge of the choices made when the answer was obtained.

You need to keep the full response to know what the engine actually did. A screenshot of the relevant passage is useful for working purposes, but I recommend also keeping the question and the context for interpreting it. This keeps the record understandable to someone encountering the observation for the first time.

Ask your service provider for the rule that distinguishes a source, a name mention and a recommendation. Then ask for a complete observation to which this rule has been applied. The discussion becomes more precise when everyone can show the selected passage and explain its classification using the same vocabulary. Before validating a classification, have the passage reviewed in context. I recommend assigning this reading to someone who can explain their decision using the response’s own words. If there is disagreement, keep the interpretations under discussion until the rule is sufficiently precise.

How should we choose useful questions for our measurements?

Choose useful questions for your measurements based on the needs your prospects are trying to address. I recommend distinguishing requests for explanations from requests for choices, then fixing the wording and interpretation criteria before seeking the responses to examine. Before taking measurements, let us write down the questions in whose responses we want to appear and how we will recognise each type of appearance. I recommend keeping this interpretation rule with the questionnaire. The report’s reader must be able to understand why a response falls into a category.

Prefer questions formulated around the reader’s need. To examine spontaneous discovery, leave your company’s name out of the question. To examine a request already about your company, state that intention. The report must keep these objectives distinct in its presentation. Classify questions by purchasing situation before reading the results. A category devoted to choosing a service provider can be examined separately from a category devoted to understanding a topic. I recommend justifying the categories through the needs you actually want to track for your company.

Choose the questions to track according to the decisions they need to inform. A question linked to your priority business activity warrants a careful reading of the need expressed and the role assigned to your company. I recommend defining this priority before taking measurements, then retaining it in the commentary.

What value does a measurement obtained through a measurement tool have?

A measurement obtained through a measurement tool is useful when its evidence and scope allow a precise interpretation. I would comment on it within the scope of the system queried, including its conditions of use; to examine another surface, I would request an observation specific to that surface.

Our published method involves asking each question several times on ChatGPT, Claude and Gemini. We then publish the number of responses that cite the site out of the number of responses obtained. This presentation provides both the numerator and the denominator of the count.2 A preprint revised on 2 April 2025 examines the repeatability of models configured to be deterministic. The authors subjected them to repeated tasks, with or without examples provided in the instructions. Under this protocol, none of the models studied consistently maintains the same accuracy across all tasks, much less identical output texts. The finding applies to the models, tasks and settings examined in this research.3

I prefer a clear count, including the number of attempts and missing responses, to a percentage with no defined scope. Keep the relevant questions readily available to read and specify the counting rule. Someone outside the measurement process must be able to establish what was counted and on what basis. A measurement that could not be made remains reported separately from a response that does not cite us. I recommend providing a category for unusable attempts and stating the reason. The reader will then know which responses were examined and which attempts still require a usable observation.

For the brand studied in our measurement on 16 August 2026, results ranged from citations in responses to 2 out of 10 questions to 3 out of 10, depending on the surface. However, the questions yielding positive results were not all the same. Similar totals therefore still reflected differences between the responses.1 Ask for a question-by-question reading when the totals seem close. To inform your decision, look for the situations in which your company appears and those that remain to be examined. A useful table provides access to the relevant passages, with their classifications, so that the meaning of each observation can be discussed.

To compare measurements, keep the questionnaire and interpretation rules accessible with each period. If the scope changes, explain the change before commenting on the difference. I recommend distinguishing retained questions from added questions so that you can choose a comparison with an explicit basis.

To examine the link between an AI response and a sales enquiry, ask the prospect to describe how they found you. I recommend keeping their attribution and the available records, then classifying the connection according to what this evidence can actually document. To discuss a sales enquiry, let us ask the prospect how they found us and keep their attribution. Prefer an open question, then ask for details about the journey they describe. I recommend keeping the prospect’s own words to distinguish their account from your team’s interpretation of it.

In its analysis, the Pew Research Center examined the addresses participants visited on a tracked device from 1 to 31 March 2025. These data made it possible to trace the searches and subsequent actions. For the Google searches in the dataset, a third-party service then repeated the same searches and collected the text on the first page, the links and the AI summaries displayed. This collection of results took place from 7 to 17 April 2025. The observed browsing and the responses collected therefore correspond to different periods.4

In the report, separate observed responses, documented visits and enquiries received. For each proposed connection, ask what record links the events. I recommend keeping uncertain attribution labelled as such so that the sales discussion reflects the quality of the available evidence.

Frequently asked questions

How should we classify the appearance of our company’s name?

Classify the appearance of your company’s name according to the role the sentence assigns to it. I recommend recording the complete wording, then specifying whether the company serves as an example, provides information or is included in a suggested choice, while retaining the qualifications that limit this interpretation.

How should we present an attempt at a measurement that could not be made?

Present an attempt at a measurement that could not be made in a dedicated category, with the reason that prevents interpretation. I recommend keeping a record of the attempt and specifying how it was treated in the count so that everyone knows which responses could actually be examined.

How should we compare measurements across several periods?

Compare measurements across periods using the retained questions and interpretation rules. I recommend identifying changes in scope before commenting on differences, then examining the relevant passages to distinguish an observed difference from a change in the counting method.

What should we ask the service provider reporting on our presence in AI responses?

Ask the service provider for the questions tested, the responses retained and the rule used to classify your presence in AI responses. I recommend examining a complete record with them, then checking how their report presents missing responses, unusable attempts and the limits of interpretation.

Sources

  1. HENUSSE, Une marque est-elle citée de la même façon selon la surface interrogée ? (2026)
  2. HENUSSE
  3. Atil et al. (arXiv preprint; published version at Eval4NLP 2025, ACL Anthology), Non-Determinism of “Deterministic” LLM Settings
  4. Pew Research Center, Google users are less likely to click on links when an AI summary appears in the results

Théo Hénusse, Founder of HENUSSE