To measure AI visibility is to check whether your brand shows up when customers ask generative AI assistants such as ChatGPT, Gemini, and Google’s AI features. Measurement is always a sample: answers vary, and the API is not the same as the app.
The blog’s guide on how to show up in ChatGPT covers what to change on the site; this article is about finding out whether the work paid off. Google’s guide to optimizing for generative AI features reminds site owners that indexing and serving are not assured.
The next sections cover the difference between the API and the apps, how to build questions by funnel stage, the metrics, Google’s and Bing’s official reports, the spreadsheet method, automated tracking, what to do with the results, and the limits.
Key Takeaways
- Measuring AI visibility means working with a sample: the same questions, in the same assistants, repeated over time.
- Search Console and Bing Webmaster Tools show impressions and citations for your own site, but they do not show what ChatGPT says about the brand.
- Measurement only pays off when it turns into action: being present in the sources the AI cites and fixing wrong facts at the origin.
Does your business show up when customers ask an AI?
ConverZAP’s free AI visibility X-ray asks an AI with web search 5 typical customer questions and shows whether your brand appears in the answers, with a score and what to fix on your site. No sign-up required.
Run the free X-rayTest the AI with my websiteWhat It Means to Measure AI Visibility
Before choosing a tool, define what counts as “showing up.” An AI answer can mention the brand, use the site as a source, or recommend the business, and each case calls for a different entry in your tracking.
Mention, Citation, and Recommendation
A mention is when the brand name appears in the text of the answer. A citation is when a page from the site is used as a source, with a link. The two do not always go together: the AI can cite the site without recommending the business.
A recommendation is the most valuable case: the brand appears as an option for the customer’s problem, often alongside competitors. For local service businesses, this is the signal that comes closest to a referral from someone you know.
In practice, tracking records the three signals separately. A brand cited as a source but never recommended has useful content and little reputation; a brand recommended without citations depends on third-party portals it does not control.
Why Measure With a Sample of Questions
There is no single dashboard that shows every conversation about the brand across assistants. Anyone measuring picks a set of representative questions and watches the answers, like an opinion poll rather than a census.
Language models with web search also break a question into several searches. Google calls this query fan-out: a question about a lawn full of weeds turns into queries about herbicides and prevention, and each one can bring in different sources.
Why Answers Change: API, Apps, and Variation
Two factors explain why two measurements of the same question rarely match. The first is the channel used, API or app; the second is the natural variation of the models, which exists even when nothing changed on the site.
The API vs the Assistants’ Apps
Through the API, the business controls the question, the model, and the search tool. OpenAI’s web search documentation says the model may or may not choose to search the web depending on the question, and Claude’s documentation says the same.
The apps add factors the API does not reproduce. Google’s Help Center says Gemini Apps can personalize the experience with memory of past conversations, connected apps, and preferences. The same question, asked by two people, can get different answers.
Natural Variation in Answers
OpenAI’s documentation warns that outputs are non-deterministic by default in the chat API: the model’s output can change from one call to the next. With web search, variation grows, because the set of pages found also changes.
That is why a single result proves nothing. A mention today and none tomorrow may be just noise. What matters is the trend over weeks, with the same questions, in the same assistants, and with the same recording method.
It is also worth noting the model used in each round. When the vendor changes the model version, answers can shift from one week to the next for reasons unrelated to the site, and the history needs to show that change.
How to Build the Question Set by Funnel Stage
The quality of the measurement depends on the questions chosen. They should sound like the customer talks, cover the stages of the purchase, and leave the brand name out, to reveal whether the AI remembers the business without being prompted.
Discovery, Comparison, and Decision Questions
A simple split, inspired by funnel-based digital marketing strategies, sorts the questions into three groups:
- Discovery: questions about the problem, such as “how to reduce appointment no-shows.”
- Comparison: requests for options, such as “best med spas in Denver.”
- Decision: questions about price, timing, and trust, close to the purchase.
Comparison questions tend to reveal the most competitors. Discovery questions show which portals and guides the AI uses as references, even when no brand appears, and point to where useful content is worth publishing.
Sample Size and Writing Rules
There is no official number of questions. A practical rule is to start with 15 to 30 questions and grow gradually. Good prompting techniques help: full sentences, city or industry context, and no instruction that forces the answer.
The questions that come up during WhatsApp lead qualification are a good source of real questions. Google warns that creating pages for every search variation, just to manipulate answers, violates its spam policy. The list is for measuring only.
AI Visibility Metrics
To measure AI visibility from the recorded answers, five metrics sum up the picture. They follow the logic of customer service metrics: few, well defined, and always calculated the same way so weeks can be compared.
Mention Rate and Citation as a Source
The mention rate is the share of answers in which the brand appears: mentions divided by total valid answers, per assistant. Splitting by assistant matters, because each one searches and writes in its own way.
Citation as a source counts the answers that used a page from the site as a reference. This number shows whether your own content is being read. A brand can be mentioned without being cited, when the AI takes the name from portals and reviews.
Position counts too. Being the first name in a list of five options is worth more than appearing at the end. Recording the order in which the brand appears helps show whether it is gaining ground even when the mention rate stays flat.
Share of Voice, Cited Domains, and Accuracy
The other three metrics look at the brand’s surroundings. The table sums up what each one measures and a simple way to calculate it:
| Metric | What it measures | How to calculate |
|---|---|---|
| Share of voice | The brand’s weight among competitors | Brand mentions ÷ mentions of all tracked brands |
| Cited domains | Which sites the AI uses as sources | Count of citations by domain, split into your own, competitors, and third parties |
| Factual accuracy | Whether the AI describes the business correctly | Answers with correct price, address, services, or hours ÷ answers that mention this data |
Bing uses a similar idea in its citation share report, calculated over all the citations for a query. Sentiment can be added as a simple score: positive, neutral, or negative.
The Generative AI Report in Search Console
On June 3, 2026, Google launched generative AI performance reports in Search Console. They first reached some sites and, since August 31, 2026, all sites worldwide, according to the announcement.
What the Report Shows
According to Search Console Help, the report shows the site’s impressions in AI Overviews and AI Mode. An impression is each time a link to the site appears to someone in one of these features. Discover has its own report.
The data can be grouped by page, country, date, and device, and exported. This shows which URLs appear most in AI features and whether a rewritten page gained impressions in the following weeks, with the same logic as the Performance report.
The most practical use is to cross-reference the report with the question list. If the pages that answer comparison questions gain impressions in AI features, Google is reading the content, even without data from other assistants.
What the Report Does Not Show
The Help page describes impressions and dimensions but does not list the queries behind each appearance. The report also does not cover Search Labs experiments and keeps the 1,000-row table limit, like the standard Performance report.
There is also a prerequisite: the report may not appear if the site has few impressions or is excluded from Search generative AI features in Search Console. And it measures Google only: it says nothing about what ChatGPT or Claude answer.
AI Performance in Bing Webmaster Tools
Bing Webmaster Tools shows how many times the site’s pages appear as sources in AI answers. The feature is in public preview and covers Copilot, Bing’s AI summaries, and some partner integrations.
Citations, Cited Pages, and Grounding Queries
In its February 10, 2026 announcement, Bing describes five data points: total citations, average cited pages per day, grounding queries, citations by page, and the trend over time. The queries are a sample.
Bing warns that citation metrics do not indicate ranking, authority, or the page’s role within an answer. The same post recommends clear headings, tables, FAQ sections, and sourced data so the AI uses the content accurately.
Intents, Topics, Citation Share, and Compare
On June 16, 2026, Bing added four features in preview: intents, which classify queries; topics, which group themes; citation share, which measures the site’s portion of all citations for a query; and comparison between periods.
Bing’s post makes an important caveat: citation share is an observational metric, not a scoreboard. It does not show competing domains, does not represent share of traffic, and does not grade content quality.
To measure competitors, the business still needs its own questions. Bing’s report shows the performance of the registered site, and comparison with specific names in the market still depends on the sample the business builds.
The Manual Method: A Spreadsheet and a Weekly Round
For small businesses, a spreadsheet handles much of the work of measuring AI visibility. The manual method is slow, but it teaches you to read the answers and shows which questions are worth it before any investment in a tool.
How to Set Up the Spreadsheet
A simple spreadsheet, with one row per question and assistant, can follow these steps:
- List the questions in column A, with the funnel stage in column B.
- Create columns for date, assistant, brand cited, position, and competitors cited.
- Record the source domains and whether the answer included any wrong fact.
- Calculate mention rate, share of voice, and accuracy in a summary tab.
- Compare week by week and note what changed on the site.
Before starting, check how an AI understands your own site. The free website test shows a summary of the business and how an agent would answer customer questions using only the text of the page you enter.
A Weekly Routine Across Several Assistants
The round should happen on the same day of the week, with the same questions, in every tracked assistant. An incognito window, with no account signed in, reduces the effect of memory and personal preferences on the answers.
Anyone who compares the leading AI assistants soon notices very different answers to the same question. That is why the spreadsheet splits results by assistant: adding everything into a single number hides where the brand is strong and where it does not appear at all.
The limit of the manual method is time. With 30 questions and three assistants, each round adds up to 90 answers to read and record, and the risk of recording things differently each week grows with the list.
Automated AI Visibility Tracking
Once the list goes beyond a few dozen questions, manual work gets heavy. Automation, common in business AI projects, sends the questions through the APIs, stores the answers, and calculates the metrics, always with the same setup in each round.
What the APIs Return for Measurement
APIs with web search return more than text. Grounding with Google Search in the Gemini API returns the queries made and citations linked to passages of the answer; OpenAI lists the sources consulted, often more numerous than the citations displayed.
In the Claude API, the web search tool always includes citations, with URL, title, and the cited passage. These fields make it possible to count cited domains without copying answers by hand, and also to check where a wrong fact came from.
ConverZAP’s Free X-Ray and Weekly Module
ConverZAP’s AI visibility X-ray asks Gemini five questions with web search, through the API. The result shows mentions, cited sources, and technical, content, and authority checks, with scores.
For weekly tracking, the AI Visibility module sends up to 50 questions to ChatGPT, Gemini, and Claude, only to the assistants with a registered API key. Gemini can run on ConverZAP credits; ChatGPT and Claude require your own key.
The module calculates the brand’s share against up to 15 competitors and audits the site. It is part of ConverZAP’s marketing solutions and is included in its plans. Google notes that no third-party tool has access to its internal ranking or AI systems.
How to Act on the Results
Measuring without acting only produces reports. The data points to two practical paths: being present in the sources the AI uses to answer and fixing, at the origin, the facts it repeats incorrectly.
Third-Party Sources Where the Brand Needs to Be
The list of cited domains shows where it pays to be well described: industry guides, local portals, review sites, and communities. Google says its AI features show what people say in blogs, videos, and forums.
The same guide warns that chasing inauthentic mentions does not help. The path is real presence: customer reviews, which a customer satisfaction survey helps encourage, and useful content on your own site, following the blog’s GEO guide.
Priority goes to sources that repeat. A portal cited across several questions and in more than one assistant weighs more than a site that appeared only once, and it deserves a complete, up-to-date profile.
Wrong Facts: Where to Fix Them
When the AI gets a price, address, or service wrong, the fix starts at the cited source. The Business Profile Help Center says businesses with complete and accurate information are more likely to show up in local search.
That is why the Business Profile comes early on the list of fixes, which usually follows this order:
- Your own site: prices, services, service area, and contact details on the right page.
- Business Profile: address, hours, category, and replies to reviews.
- Third-party portals: a request to update the site that published the old data.
On the site, the About page and Organization structured data help Google understand and disambiguate the organization. After the fix, digital marketing automation can remind the team to review the answers in the following rounds.
Measurement Limits and Next Steps
Every AI visibility measurement has blind spots, however careful the method. Knowing these limits prevents hasty conclusions and promises no company or tool can keep, such as assuring a citation in the answers.
What No Measurement Can Promise
No sample covers every question asked to the assistants, and the API does not reproduce each user’s history. Official reports only measure their own search engine. Improving the site raises the odds, but it does not ensure a citation, as Google makes clear.
AI agents, which already make bookings and compare products on people’s behalf, according to Google, open a front that the reports described here do not break out. AI trends suggest the method will change, which is why the historical record is so valuable.
There is also the attribution limit. Even when the brand is recommended, it is not always possible to know how many customers arrived because of an AI answer, since many type the name into a search engine or message the business directly on WhatsApp.
Ongoing Measurement as a Marketing Routine
Measurement works best as a marketing automation routine: the same list, the same day, and a monthly review of actions. That way, every change on the site or new mention can be compared with previous rounds.
When the AI recommends the business, customers tend to reach out right away, often on WhatsApp. ConverZAP combines tracking of AI answers with customer service on the official WhatsApp Business API, with AI agents that reply and record where each contact came from. The AI visibility solution shows how the weekly tracking works.
Frequently Asked Questions
You can start with a spreadsheet, 15 to 30 questions, and a weekly round in the assistants, plus the free reports in Search Console and Bing Webmaster Tools. ConverZAP’s free X-ray runs a first diagnosis with five questions to Gemini.
One round per week is usually enough to see trends without piling up noise. Measuring every day tends to show only the natural variation of the answers. What matters most is keeping the same day, the same questions, and the same assistants.
There is no official number. A list of 15 to 30 questions, spread across discovery, comparison, and decision, already shows patterns. Over time the list can grow, as long as the older questions stay the same so rounds can be compared.
No. The generative AI report in Search Console shows the site’s impressions in Google’s AI Overviews and AI Mode. What ChatGPT answers only shows up in tests with the questions, manual or automated.
In the apps, the answer can take into account the memory, location, and preferences of the person asking. Through the API, those variables are left out. In addition, models are not deterministic, so even two identical calls can return different text and sources.
The first step is to see which source the answer cited and fix the data there: on the site, in the Google Business Profile, or on the third-party portal. The following rounds then show whether the answer changed.
