There is a question I hear increasingly often in conversations about online visibility: “If someone asks ChatGPT to recommend a product or service like mine, will my business appear in the answer?”
It is a reasonable question. If people are starting to use AI assistants to compare products, evaluate suppliers and narrow down their options, understanding how a business is represented in those systems becomes relevant.
The problem is how we usually try to find out.
Someone opens ChatGPT, types “What are the best companies in this industry?” and checks whether their brand appears.
If it does, everything looks fine. If it doesn’t, apparently they need GEO.
Sometimes the test is repeated on Gemini and Perplexity, three screenshots are collected, and the result is presented as an AI visibility audit.
I do not find this particularly useful.
A response generated at one moment does not necessarily represent what users will see tomorrow. It certainly cannot establish how frequently a brand is recommended.
The goal is not to ask ChatGPT once whether it knows your company. It is to develop a method for observing when, how and in which contexts your company appears in AI-generated answers.
In this article, I want to explain how I would structure that analysis, which data I would collect and, perhaps most importantly, which conclusions I would avoid drawing.
First, distinguish brand recognition from recommendations
Before building a measurement framework, we need to separate three situations.
The first is when an AI assistant recognizes a brand. Ask “What is company X?” and it may be perfectly capable of describing its products, services and market.
The second is when the company is mentioned without being explicitly named in the question. For example, ask “Which property management systems should I consider for a small hotel?” and company X appears among the suggested options.
The third is when the assistant cites the company’s website as a source, perhaps using an article to explain a technical topic without recommending any of its products.
These are different outcomes.
A company may be recognized but rarely appear in commercial recommendations. It may be cited as an informational source without being presented as a supplier. Or it may be recommended without receiving a direct link to its website.
Research published by Victorious in July 2026 highlighted this distinction. In a study covering brands from several industries, 96% of the brands eligible for its recognition analysis were described accurately, while 89% of those in its mention analysis did not appear in generic category-research answers.
These results apply to specific samples and methods, rather than all businesses. The research was also published as sponsored content, which is worth considering when interpreting the findings.
Nevertheless, the distinction is valuable: knowing a brand and recommending that brand are not the same thing.
That is why I would never consider asking ChatGPT to describe a company a sufficient GEO audit.
Start with the questions real customers would ask
The first part of the work has little to do with ChatGPT itself.
We need to understand what the business sells, which problems it solves, which customers it serves, which alternatives people consider before purchasing and when a recommendation would actually be appropriate.
Consider a company developing property management software for independent hotels.
Rather than testing a single question, I would build a prompt set organized by search intent.
| Intent | Example question |
|---|---|
| Exploratory | How can a small hotel simplify reservation management? |
| Informational | What is the difference between a PMS and a channel manager? |
| Comparative | Which property management systems are suitable for independent hotels? |
| Decision-making | Which PMS should I consider for a 25-room hotel? |
| Requirements-based | I need a PMS with a booking engine, rate management and OTA integrations |
| Branded | What are the strengths and limitations of software X? |
Not every question should result in a commercial recommendation.
If someone asks about the difference between a PMS and a channel manager, a straightforward explanation without brand names might be the most useful answer.
A comparative question, on the other hand, offers a much better opportunity to observe which suppliers the system considers relevant.
I would not treat the absence of a brand as a negative result when mentioning that brand would not be useful to the user.
For this reason, I would distinguish informational prompts from prompts where recommending a supplier is relevant.
How many prompts do you need?
There is no universal answer.
For an initial audit of a small or medium-sized business, I would start with approximately 30 carefully selected questions across the main search intents.
These should not be 30 random prompts, nor a list generated automatically without human review.
I would draw on the website’s content, Search Console queries, Google Ads search terms, sales conversations, frequently asked questions and competitor research.
The aim is to cover a credible range of purchase situations.
This does not make the sample representative of every question people ask AI assistants. It creates a controlled observation panel that can remain consistent over time.
And that is already much more useful than testing a single question.
One prompt is not enough: wording matters
There is another issue that is arguably even more important.
Two people looking for the same product may describe their requirements in completely different ways.
One might ask:
“What are the best property management systems for small hotels?”
Another might ask:
“I manage a 25-room property and want to replace our PMS. My main priority is reducing manual work around reservations. Which suppliers should I evaluate?”
The commercial intent is similar, but the level of detail is different. The answers may be different as well.
Research published in July 2026 by researchers associated with Peec AI examined the impact of prompt wording on brand visibility, supporting the argument that a single canonical prompt is not a sufficient basis for measurement.
For this reason, I would organize monitoring around topics rather than individual sentences.
A topic such as “property management software for small hotels” could include a general question, a comparative question, one with technical requirements and another built around an operational problem.
I would still distinguish wording variations from changes in intent. Rephrasing a question is not the same as introducing new purchasing requirements.
This makes it possible to observe whether a company appears consistently across a topic, or only when a question is phrased in a very particular way.
How I would run the tests
Once the question set has been defined, we need a testing procedure.
For an initial manual audit, I would use the available search-enabled experiences in ChatGPT, Gemini and Perplexity, recording the conditions in which each response is generated.
I would not mix answers produced without web retrieval with answers produced using live external sources. These are different environments.
Where possible, I would use fresh conversations and limit personalization. I would document language, country, date, platform and product mode.
I would also repeat prompts across several sessions because identical questions can produce different answers.
A starting configuration might look like this:
- 30 representative prompts.
- Three wording or intent variants for priority topics, included in the total sample or tracked separately.
- Three platforms: ChatGPT, Gemini and Perplexity.
- Three repetitions per prompt and platform.
With 30 prompts in total, this produces up to 270 responses per measurement cycle.
That is manageable with suitable collection tools, although substantial for a fully manual process.
Three repetitions are not a scientifically established threshold. They are simply a practical starting point for observing variability without making the project unnecessarily expensive.
Larger samples and explicit uncertainty estimates would be needed for more robust conclusions.
A note on automation
Automating these tests can be helpful, but it introduces complications.
An API or third-party monitoring tool does not necessarily produce the same output a user would see in a platform’s consumer interface.
Models, retrieval tools, settings, citation behavior and product features may differ.
I would therefore document the testing environment and avoid combining results from different interfaces without accounting for these differences.
Any automated collection should also comply with the relevant service terms and permitted access methods.
What to collect: an AI visibility matrix
I would record more than whether the company name appears.
For each response, I would capture at least:
- Prompt identifier and topic.
- Platform, product mode and test date.
- Whether the brand appears in the answer text.
- Whether the official website receives a citation.
- The brand’s position in any list of alternatives.
- Other brands mentioned.
- Cited sources and their URLs.
- Accuracy of the information provided.
- Nature of the mention: neutral, positive, negative or conditional.
- The complete response for later review.
Classification rules need to remain stable.
For instance, I would only count a brand mention when the brand is explicitly identifiable in the answer text, not when the system merely cites a document on its website.
Likewise, I would only count a recommendation when the company is presented as a relevant option to consider, rather than mentioned incidentally.
These distinctions may seem less impressive than a single visibility score, but without them, comparisons over time quickly become unreliable.
The four metrics I would use
For an initial dashboard, I would avoid compressing everything into an unexplained “GEO score”.
I would use four separate indicators.
1. Brand Mention Rate
This measures the proportion of valid responses in which a brand is explicitly named.
Brand Mention Rate = responses mentioning the brand / valid responses × 100
Suppose we collect 90 valid responses to unbranded commercial prompts and the company appears in 18 of them.
The Mention Rate is 20%.
That means the brand appeared in 20% of the responses in our sample, not in 20% of all conversations users have with ChatGPT.
This distinction belongs in every professional report.
2. Citation Rate
This measures the proportion of responses containing at least one link to the company’s official domain.
Citation Rate = responses citing the domain / valid responses × 100
I would track citations separately from mentions because a website may be used as a source without the company itself being named.
A Semrush study published in June 2026 examined this behavior, sometimes described as ghost citations: a domain receives a citation, but the brand is not explicitly mentioned in the answer.
Within its sample, approximately 62% of domain citation appearances were not accompanied by an explicit brand mention.
This illustrates why the two signals should not be treated as equivalent.
3. Recommendation Rate
For prompts where suggesting a supplier makes sense, I would measure the proportion of answers explicitly presenting the company as an option worth considering.
Simply appearing in a list or a secondary sentence is not necessarily enough.
An answer stating that “software X is a recognized player in the industry” is different from one saying “for a property with these requirements, software X would be worth evaluating”.
This classification inevitably involves some interpretation.
I would use documented rules, examples and a second review of ambiguous responses to maintain consistency.
4. Competitor Presence
Finally, I would measure how often competitors appear within the same prompt sample.
This is often more informative than looking at the target brand in isolation.
If our company appears in 15% of answers, it matters whether our main competitor appears in 70% or 12%.
I would not automatically call this market share or an industry-wide share of voice. We are comparing responses from a defined test sample, not measuring every user’s behavior.
Instead, I would report comparative brand presence, together with the underlying response counts.
An example report
Imagine running 90 valid tests with unbranded commercial prompts on one platform.
The results might look like this:
| Brand | Responses mentioning the brand | Mention Rate |
|---|---|---|
| Target company | 18 | 20% |
| Competitor A | 39 | 43.3% |
| Competitor B | 25 | 27.8% |
| Competitor C | 12 | 13.3% |
Entirely hypothetical data, used to illustrate the method.
The percentages do not add up to 100%, because a response can mention several companies.
The table shows relative frequency, but the more interesting work begins afterward.
Which topics cause competitor A to appear most frequently? Which questions bring competitor B into the answer? Which sources are cited when our company is absent?
And, perhaps most importantly, are competitors providing useful information that our website lacks, presents ambiguously or fails to organize clearly?
A useful report should attempt to answer those questions as well.
Source analysis can be more useful than counting mentions
When an AI assistant retrieves information from the web, some interfaces display links to supporting sources.
Analyzing those sources helps reveal what types of material appear in our sample: official websites, software comparison platforms, directories, forums, trade publications, technical documentation or reviews.
For example, if responses about hotel management software frequently reference specific comparison platforms, it may be worth checking whether our company is included and whether the information is accurate.
This does not mean attempting to place the company on every cited website.
It means identifying sources that appear relevant in the observed research journeys and considering whether they justify further attention.
There is an important distinction here: the most frequently cited website is not necessarily the most frequently recommended brand.
A Semrush study published in July 2026, covering more than 50,000 brands and 1,094 topic areas, highlighted the difference between citation frequency and explicit brand presence in AI-generated answers.
That has practical implications.
A blog can provide information that AI systems cite regularly without its owner being recommended as a supplier.
Conversely, a company may appear among recommended alternatives without receiving a direct link to its own website.
This is another reason to keep editorial visibility and commercial visibility separate.
Connecting the analysis to GA4 and Search Console
So far, we have focused on what happens within AI-generated answers.
But assessing business impact also means examining what happens when people reach the website.
In GA4, we can analyze referral traffic from domains associated with AI assistants, where the referral information is actually passed through.
This should not be treated as a complete measure of AI visibility.
Someone might discover a company through ChatGPT and then search for its name on Google. They may type the address directly, use another device or never visit the website at all.
In these situations, traditional attribution reports may not reveal the contribution of the original AI interaction.
A significant Search Console update in 2026
On June 3, 2026, Google began introducing dedicated performance reports for generative AI features in Search Console.
The initial rollout was limited, but Google expanded availability worldwide by August 31, 2026.
This provides a dedicated view of impressions in generative Search experiences, including AI Overviews and AI Mode.
It is a useful development, but these metrics should not be confused with brand mentions or recommendations in ChatGPT and other independent assistants.
The platforms operate differently and expose different kinds of data.
I would bring these metrics together in a dashboard, but I would not combine them into one headline number without clearly explaining what each represents.
How can we tell whether GEO improvements are working?
Suppose the initial audit identifies several areas for improvement.
We update service descriptions, add more specific technical information, clarify use cases and correct inaccurate or outdated content.
Several weeks later, we repeat the monitoring exercise.
Our Mention Rate rises from 20% to 27%.
Can we claim that our GEO work caused the improvement?
Not yet.
First, we need to confirm that the samples are comparable: the same topics, the same distribution of intents, and equivalent platforms and settings.
We should also examine response variability and determine whether the observed difference is larger than the fluctuations normally seen during testing.
Other factors may have changed: model updates, new sources, competitor activity or changes in the search environment.
Where practical, I would use control questions or unchanged pages as reference points, while recognizing that these do not eliminate every possible source of interference.
A before-and-after comparison is not useless. It is evidence to interpret, rather than automatic proof of causation.
What I would deliver after an initial audit
A project like this should result in something more substantial than a colorful dashboard.
I would structure the report around five areas.
First, the baseline: how many responses were analyzed, on which platforms, using which questions and under what conditions.
Second, brand visibility: mentions, recommendations and citations, keeping branded and unbranded prompts separate.
Third, competitive positioning: which companies appear, in what contexts and how consistently.
Fourth, source and accuracy analysis: which resources are cited, which information is inaccurate or outdated, and where content could be improved.
Fifth, an action plan: a limited set of prioritized interventions, with an explanation of why they matter and how their effects will be evaluated.
I would not automatically recommend publishing dozens of new articles.
In many cases, it could be more effective to improve existing service pages, correct inaccurate third-party descriptions or make product documentation easier to understand.
The aim is not to manufacture content simply because there is a new acronym to include in a proposal.
It is to address problems the analysis has actually revealed.
How reliable is AI visibility measurement?
That depends on what we expect it to tell us.
If our goal is to know precisely how many people worldwide saw our company recommended by ChatGPT in a given month, a prompt-based monitoring exercise cannot provide that answer.
We do not have access to the complete distribution of user questions, and response variability makes broad generalizations difficult.
If, instead, we want to monitor commercially relevant topics, compare competitor presence and identify inaccurate information, we can build a useful and repeatable process.
Existing research supports the need for caution.
A 2025 AirOps study examining more than 45,000 citations highlighted how brand visibility can fluctuate between sessions.
A June 2026 research paper analyzing more than 100,000 responses identified further differences in AI visibility across brands of different sizes and levels of recognition.
I therefore see little value in pursuing a number-one position in a single AI response as though it were a traditional search results page.
It makes more sense to monitor a topic over time and assess whether a company appears consistently in conversations where mentioning it would be relevant.
Conclusion: measure before you optimize
GEO remains an area where it is easy to sell certainty we do not actually have.
Producing a screenshot in which ChatGPT mentions a business is relatively simple. Demonstrating that the mention is frequent, representative and connected to commercial outcomes is considerably harder.
That does not mean abandoning measurement.
It means defining the questions carefully, separating the metrics, repeating the tests and being transparent about the limitations.
The principle is the same one that should guide any digital marketing activity: do not confuse what we can measure with what we would like to prove.
When assessing visibility, I would rather present a less impressive but verifiable result than an apparently precise score built on unclear assumptions.
Ultimately, the value of a GEO audit is not in telling us whether ChatGPT knows our name today. It is in helping us understand where our business appears, where it is missing and what we can realistically do to improve the quality of its online presence.
And if that work eventually contributes to measurable commercial opportunities, we will have achieved something considerably more valuable than a passing mention in an AI-generated answer.
References and further reading
- Google Search Central, Introducing Search Generative AI performance reports in Search Console, June 3, 2026, updated August 31, 2026. https://developers.google.com/search/blog/2026/06/gen-ai-performance-reports
- Semrush, AI visibility is a topic-level game: A study of 50,000 brands in ChatGPT, July 20, 2026. https://www.semrush.com/blog/chatgpt-topic-authority-study/
- Semrush, Why 62% of AI citations don’t lead to brand mentions, June 9, 2026. https://www.semrush.com/blog/the-ghost-citations-study/
- Victorious / Search Engine Journal, AI Recognizes 96% of Brands But Mentions Almost None, July 29, 2026. https://www.searchenginejournal.com/ai-brand-mention-study-victorious-spa/582765/
- Ehrlinspiel J., Landwehr M., Rudzki T., Prompt Tracking Works, But Not as a One-Prompt Measurement System, July 2026. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6914539
- Kumar P., Generative Engine Optimization at Scale: Measuring Brand Visibility Across AI Search Engines, June 2026. https://arxiv.org/abs/2606.20065
- AirOps, Staying Seen In AI Search: How Citations & Mentions Impact Brand Visibility, September 23, 2025. https://www.airops.com/report/how-citations-mentions-impact-visibility-in-ai-search
Leave a reply