In classic SEO measurement is clear: keyword, position, click, traffic. In AI answers there is no such thing as “position”. Ask the same question twice and you may get different answers; two users asking the same question in different contexts may see different sources. That uncertainty makes measuring AI visibility harder, but not impossible.
In this article, instead of searching for one magic metric, we describe four measurement layers that mean something when read together: regular question testing, tracking citations and mentions, AI-sourced traffic to your site and classic search data. An important note: there is no standard, universally accepted measurement tool in this field yet, so use the methods below as “directional records”.
What Does AI Visibility Mean?
AI visibility is the degree to which your brand or content appears in the answers produced by systems such as ChatGPT, Gemini, Perplexity, Claude, Copilot and Google AI Overviews. It shows up in several forms, each with a different value:
- Mention: your brand name appears in the answer without a link.
- Citation: the answer links to your page; the user can click through to the source.
- Recommendation: the answer directly recommends you or your service to the user.
- Accuracy: the information given about your brand is correct, incomplete or wrong.
Track the four separately. A mention indicates awareness, a citation traffic potential, a recommendation proximity to purchase intent and accuracy a reputation risk. Seeing wrong information points to content or source issues that need correcting. If measurement shows your brand isn’t in answers, we collected seven common causes and a first step for each in why your brand doesn’t appear in ChatGPT.
For the background, see what is GEO and what is AEO.
How to Do a Prompt Test
A prompt test means asking your chosen questions on AI platforms at regular intervals and recording the results. It looks simple, but done regularly and consistently it is the most useful measurement.
- Build the question set. Choose 15 to 30 questions your customers really ask: service searches (“can you recommend an accounting firm in [city]?”), comparison, price, “how to” and brand-name questions. Use questions from your sales team and support records.
- Pick the platforms. ChatGPT (with search on), Gemini, Perplexity and Google AI Overviews/AI Mode are usually enough to start.
- Fix the conditions. Ask the same question the same way, ideally logged out or in a clean session, to reduce personalization.
- Ask each question more than once. Answers can vary, so a single try isn’t reliable. Ask three times and look at the trend.
- Record the results. Keep a table of date, platform, question, mention, citation, recommended brands, answer accuracy and notes.
- Repeat monthly. For comparable data, ask the same set the same way each month.
The record table can have these columns:
| Date | Platform | Question | Mentioned? | Cited? | Competitors | Accuracy | Note |
|---|---|---|---|---|---|---|---|
| Month/Year | ChatGPT | Question text | Y/N | Y/N | Brands mentioned | Right/Incomplete/Wrong | Observation |
Presenting these results as an exact “percentage” would mislead. Make a directional reading per question, such as “this month we appeared on these questions and not on those”. Because platforms and models update often, a change within a month can be independent of your work.
How to Read Citation and Mention Data
After collecting results you need a meaningful summary. These questions structure it:
- On which questions do we appear, and on which don’t we? Most gaps point to missing content or an access problem.
- Which pages are cited? Seeing which of your pages are chosen as sources shows which content types work. Use them as models for similar pieces.
- On which questions do competitors stand out? Study the kind of page a competitor gets cited for and compare with your own.
- Is there wrong information? Errors such as an old price, a wrong address or a closed service usually come from outdated sources. Update your site and external profiles.
- Which sources get cited? Note the third-party sites often cited in answers (guide sites, news sites, forums); you can plan content or PR for them.
It’s also useful to look at the shared features of cited pages: usually pages that answer the question directly, are well structured, current and strong in trust signals. We described how to strengthen those in content optimization for AI-powered search engines and Google AI Mode.
What to Track in GA4 and Search Console
A prompt test is a sample you take; real user behavior shows in analytics tools. Read both sources together.
AI-Sourced Traffic in GA4
Visitors from AI assistants are usually reported under specific source names. In GA4 you can group them into a separate channel:
- Under Admin > Data streams > Channel groups, create a new channel group.
- Define a channel named “AI” and add a condition that the source name matches terms like
chatgpt.com,perplexity,gemini.google.com,copilot.microsoft.comandclaude.ai. - Move the channel above the default channels in the list.
Source names can change over time, so check the report every few months. The volume of this traffic is small on most sites at first; the more valuable indicators are conversion rate and session quality.
Search Console
Search Console doesn’t label the effect of AI features directly, but gives indirect signs:
- Rising impressions with falling clicks can be a sign answers are being consumed on the results page.
- A rise in brand-name searches can be an indirect effect of mentions in AI answers.
- Changes in position and clicks for question-style queries.
For the general frame on reading Search Console data, see Google Search Console and SEO reporting. You can also read our 12-month Search Console data review, which shows with our own data how impressions can mislead.
Conversions
The last link is the business result: form submissions, phone calls, quote requests and sales. Even if AI-sourced traffic volume is small, a high conversion rate means this channel is gaining importance.
A Measurement Plan and Realistic Expectations
A monthly routine bringing all these layers together:
| Layer | Frequency | Output |
|---|---|---|
| Prompt test | Monthly | Per-question record of mention, citation and accuracy |
| Citation and competitor analysis | Monthly | Which pages are chosen as sources, questions where rivals are strong |
| GA4 AI channel | Monthly | Sessions, engagement and conversions |
| Search Console | Monthly | Trend on brand and question queries |
| Content and correction list | Monthly | Pages to update and wrong information to fix |
Set expectations right from the start:
- Results show a trend, not an exact ranking.
- A change within a month may come from a platform’s model update.
- The effect of work to increase AI visibility is often delayed.
- No agency or tool can guarantee appearing in a particular answer.
To set up this plan and run the work together, see the process and scope on our GEO consultancy page. We can also share our own tracking template with your site.
Common Measurement Mistakes
- Trusting a single try. Looking at a question once and saying “we appear” or “we don’t” misleads. Answers vary; several tries are needed.
- Changing the question set every month. Use the same questions to compare. New ones can be added but don’t remove the old ones.
- Only asking with your brand name. Appearing for “What is [your brand]?” is easy; the real value is appearing for general questions that don’t name you (“a reliable SEO agency in [city]”).
- Forgetting personalization. In a logged-in account your past questions can affect results. Check in a clean session or on different devices if possible.
- Reducing the result to one number. An exact “40% visibility” hides the sample’s limits. Report the trend and per-question status.
- Ignoring mentions without citations. A mention without a link can still contribute to awareness, but it won’t bring direct traffic.
- Not tracking conversions. Traffic may be small, but AI-sourced visitors can convert at high quality; if you don’t measure, you can’t see the channel’s value.
What a Sample Monthly Report Looks Like
A monthly summary for management or a client should consist of headings that drive decisions, not a pile of numbers:
- Overall status: number of questions tested this month, and the list where the brand was mentioned and cited.
- Change: questions newly appearing and disappearing compared with last month.
- Competitors: questions where rivals stand out and the page types that get cited.
- Accuracy issues: answers with wrong or old information, and corrections to make.
- Traffic and conversions: sessions, engagement and conversions from the AI channel in GA4.
- Action list: pages to update next month, content to write and technical fixes.
The aim of the report is not to say “we’re doing well” but to clarify what to do next month. That logic matches the prioritization approach in SEO audit work.
Frequently Asked Questions
How is AI visibility measured?
By regularly testing set questions on the platforms, recording mentions and citations, tracking AI-sourced traffic in GA4 and reading Search Console data together. There is no single standard tool.
Is prompt testing reliable?
Single tries aren’t, because answers vary. Asking the same question several times and fixing conditions improves reliability, but results are still a sample.
What is the difference between a mention and a citation?
In a mention your brand name appears without a link; in a citation your page is linked. A citation increases the chance of the user visiting your site directly.
How do I see traffic from ChatGPT in GA4?
The source is reported under values like chatgpt.com. Create a custom channel group covering these sources to track AI traffic separately.
Which tools exist for AI visibility?
Various paid tracking tools are emerging, but their methods and coverage differ. For small and medium businesses, regular manual testing and GA4 and Search Console tracking are often enough to start.
How often should I measure?
Monthly is enough for most businesses. After a major content or technical change, an extra test within a few weeks is also useful.
