How to Measure Your Brand Visibility in ChatGPT and AI Search

AI search visibility is measured through three methods used together: a fixed set of questions run periodically and documented, referral traffic from AI products in your analytics, and bot activity in your server logs. For Google you add Search Console. No method gives an exact "position," so you report appearance rates, not rankings.

This guide describes the method we recommend, with a testing protocol, a reporting table and the limits of each data source. We will not invent benchmark numbers: there are no credible public averages for "how often you should be cited." The strategic context is in our GEO overview; measurement is also part of our generative engine optimization service.

What does "visibility" mean in an AI engine?

AI search visibility is the frequency and quality with which your brand or pages appear in generated answers to the questions that matter to your business. It has several components worth separating:

• Mention: your brand is named in the answer, with or without a link.
• Citation: the answer points to one of your pages as a source.
• Accuracy: what the answer says about you is correct.
• Context: you appear as a recommendation, a neutral example or a warning.
• Competition: who else appears next to you and how often.

A correct mention without a link has brand value; a citation with a link also has traffic potential; a wrong statement is a problem to fix. A single "visibility score" blends them all and hides what you need to do.

Method 1: documented tests on a question set

This is the most important and cheapest method. It works the same way for ChatGPT, Perplexity, Gemini, Claude or Google AI Mode.

How to build the question set

1. Write 30–50 questions real customers would ask, in their own wording. Start from emails, calls, your site chat and our question generator.
2. Split them by stage: learning ("how to," "what is"), comparing ("what is the difference"), choosing ("who do you recommend," "best X in Bucharest"), brand ("what do people think of [your brand]").
3. Include questions with and without your brand; the ones without show whether you are discovered, not only recognized.
4. Keep the set stable from month to month; add new questions at the end rather than replacing old ones, so you can compare.

If you serve the Romanian market, run the set in Romanian and, if foreign buyers matter, in English too. The two languages can produce different sources.

The run protocol

1. Use the same products and modes in each test (for example, the feature with web search switched on), in a session without a personalized account or with fixed settings.
2. Run each question 3 times, in fresh sessions, and record every result.
3. Save the date, product, mode, approximate country, exact question text and the answer (screenshot or text).
4. Record: brand mentioned (yes/no), page cited (URL), accuracy (correct/wrong/incomplete), which competitors appear.
5. Calculate the rate: in how many runs you appear, out of all runs.

Because answers vary, appearing in 1 of 3 runs is different information from appearing in 3 of 3. Report the rate, not the best case.

Reporting table (template)

The numbers in the table are a formatting example, not a performance benchmark.

Method 2: referral traffic in analytics

When a user clicks a link inside an AI answer, the browser may send a referrer (for example, the product's domain), and analytics can classify it. Steps:

1. In your traffic acquisition reports, filter for sources containing AI product domains (chatgpt.com, perplexity.ai, claude.ai, gemini.google.com, copilot.microsoft.com and similar).
2. For repeatable reporting, define a custom channel group with a regular expression such as:

chatgpt\.com|chat\.openai\.com|perplexity|claude\.ai|gemini\.google\.com|copilot\.microsoft\.com

1. Check which landing pages receive visits and what visitors do (time, conversions).
2. Compare monthly, not daily: volumes are usually small and noisy.

Two points to verify in your own account. First: some sources report that ChatGPT appends a parameter like utm_source=chatgpt.com to links; if you see it in your landing URLs, you can use it for segmentation. Second: some sources report that Google Analytics 4 added a default "AI Assistant" channel in 2026; check whether it exists in your property and which sources it covers, because the full list may not be published and coverage may not include every product. The channel relies on the referrer, so visits without one still land in Direct.

What you cannot see here: visits without a referrer land in "direct"; users who see your brand in an answer and then search for you on Google show up as organic traffic; and clicks from Google's AI Overviews are counted as Google traffic, not as a separate referral. For the last case you have Search Console (below). For the basics of KPIs and reports, see SEO KPIs and the monthly report.

Method 3: server logs and bots

Logs show which AI bots access your pages. How to use them correctly:

• Search bots (for example OAI-SearchBot, PerplexityBot, Claude-SearchBot) indicate that your pages are being considered for search results, not that you will be cited.
• User-triggered bots (for example ChatGPT-User, Perplexity-User, Claude-User) indicate that someone asked the product to access your page. This is a concrete signal of interest, especially useful for specific pages. Note that OpenAI and Perplexity say in their documentation that these user-initiated fetchers may not follow robots.txt rules.
• Verify IPs. OpenAI, Anthropic and Perplexity publish official IP ranges; compare before drawing conclusions, because a user agent can be spoofed.

Access setup is explained in our guide to AI crawlers and robots.txt. Without access, there is no point measuring citations.

Method 4: Search Console for Google's AI features

Google's documentation says traffic from AI Overviews and AI Mode is counted in the Performance report, under the "Web" search type. In addition, Google introduced a dedicated "Generative AI performance" report that shows impressions (not clicks) in AI Overviews and AI Mode, grouped by page, country, date and device; according to the help page, it had been rolled out to all websites worldwide as of August 31, 2026. If you do not see the report, the cause may be too few impressions in these features, or that you have opted out of them. Details in our guide to Google AI Overviews and our Search Console guide.

Method 5: third-party tools

Some services automatically run questions in AI products and compute "scores." We do not recommend names, because the market changes fast, but you can evaluate any tool with these criteria:

• Can you define the questions, or are they imposed on you?
• Does it run each question several times and show the variation?
• Does it state clearly which product, mode and country it simulates?
• Can you export the raw data?
• Does it avoid promising access to providers' "internal data"? Google explicitly warns that no third-party tool has access to its internal ranking or AI systems.

A good tool automates the protocol above; it does not replace it with an opaque score.

Comparison: what each method shows

How to interpret results without fooling yourself

Imagine that two months apart you have the same choosing question. The first time you appear in 1 of 3 runs, the second time in 2 of 3. Is that an improvement? Possibly, but a difference that small on a single question can be ordinary variation. You need a pattern: the same direction across many questions, in several products, over several months.

A few practical rules of interpretation:

• Look at aggregates by stage (learning, comparing, choosing), not isolated questions.
• Mark every major site change in the table (rewritten pages, new content, bot policy changes) so you can correlate later.
• Treat factual errors as a separate priority: a missing or unclear page can be the cause of a wrong answer about you.
• If a competitor shows up consistently, read the pages the product cites and compare them with yours; the answer to "why them?" is usually in clarity, originality or coverage.

Common mistakes

• One test, one time. Variation makes conclusions unreliable.
• Questions worded to make you look good. Use your customers' language, not yours.
• Mixing conditions. A personalized account today, a clean session tomorrow: you are comparing apples to oranges.
• Treating referral as the only measure. It undercounts influence.
• Blind attribution. If you changed five things in a month, you do not know which one mattered.
• Trusting scores with no method. Always ask for the raw data.

Limits, and when measurement tells you little

• There is no stable "position." The answer is composed anew each time.
• There is no public volume data for questions asked in these products.
• Products change. A new model or feature can shift results without you doing anything.
• For very local or tiny-niche businesses, referral may be close to zero; manual tests remain the only signal.
• Correlation is not causation. If you start appearing after publishing an llms.txt file, you have not proven it caused the change.

A five-step implementation plan

1. Build the set of 30–50 questions and split it by stage.
2. Run the baseline (3 runs per question) and save everything in a table.
3. Configure the AI channel group in analytics and, if you need tagged links for your own materials, use the UTM link generator.
4. Turn on bot tracking in your logs and verify IP ranges.
5. Report monthly: mention and citation rates, accuracy, referral, Search Console impressions, plus a short action list for next month.

For a quick first read on your site's readiness, you can also use the GEO score.

Conclusion: measure to decide, not to impress

The most useful measurement is simple and repeatable: the same questions, the same conditions, honestly recorded results. Combine it with referral, logs and Search Console, and do not buy scores that hide their method. Next steps: write the questions, run the baseline test, set up channels in analytics. If you want help with setup and interpretation, see our GEO service page.