Measuring AI visibility: official data, its limits and a practical tracking plan

What AI visibility means

When someone asks an AI system a question, your site can show up in three different ways, and each one needs a different data source.

• Cited. The answer links to one of your pages as a source. This is the closest thing to a ranking in AI search.
• Mentioned. The answer names your brand, product or author, but without a link. It shapes what people think of you, yet it never shows up in analytics.
• Visited. The person clicks a link in the answer and lands on your site. This is the only layer your own analytics can see directly.

A fourth layer sits on top: converted, when the visit leads to a sign-up, a lead or a sale. The layers are not interchangeable. A month with more citations can bring fewer visits, because a good answer often satisfies the question on the spot. That is why ai visibility tracking needs one number per layer instead of one blended score.

Why your pages get cited in the first place is a separate question, covered in the lesson How AI answers choose sources. This lesson is about counting what happens.

What Google officially says

Google also says that the two features behave differently: “AI Mode and AI Overviews may use different models and techniques, so the set of responses and links they show will vary.” And AI Overviews do not appear on every query: “AI Overviews are only shown when our systems determine that it is additive to classic Search, and as such, often don't trigger.” Both points matter for measurement. The same query can produce different sources in the two features, and many of your queries will never show an AI Overview at all.

Search Console: AI Overviews and AI Mode inside Web totals

Search Console is the only official source of Google impressions and clicks, and AI features are already in it. You do not need to set anything up. What you cannot do is separate them: as of October 2026, Google documents no filter, search type or search appearance that isolates AI Overviews or AI Mode. Their clicks and impressions are mixed into the Web search type together with classic results.

How AI features are counted

• Clicks. “Clicking a link to an external page in the AI Overview counts as a click.” The same rule applies in AI Mode: “Clicking a link to an external page in AI Mode counts as a click.”
• Impressions. For both features, “Standard impression rules apply.”
• Position. All links in one AI Overview share a single position. In AI Mode, position follows the same method as a normal results page.
• Follow-up questions. “If a user asks a follow-up question within AI Mode, they are essentially performing a new query.” Each follow-up creates its own query row.
• Experiments. “Search Console doesn't include data from experiments in Search Labs, as these experiments are still in active development.”

How to read AI effects in the Performance report

Because you cannot filter AI features out, you read their effect indirectly. Three patterns are worth watching:

• Impressions up, clicks flat. For informational queries, this often means your page is shown more often, sometimes inside an answer, but people get what they need without clicking.
• Average position improves while clicks fall. Since every link in an AI Overview gets the overview's position, a page cited in an overview at the top can show a strong average position with few clicks.
• Long question-style queries growing. Longer, conversational queries are typical of AI Mode and follow-up questions. Group them with a regex filter, for example queries that start with how, what, which or why, and track that group month by month.

Filters, regex and period comparisons are explained step by step in the lesson Performance report: filters, regex, comparisons. Keep one more limit in mind when you check results by hand. Google says that “Search results are specific to the time, place, device, and recent history of the person searching.” The answer you see on your laptop is one sample, not what your users saw.

GA4: the AI Assistant channel and referral data

GA4's default channel group splits AI traffic in two. Clicks from Google's own AI features go to Organic Search: “Organic Search is the channel by which users arrive at your site/app via non-ad links in organic-search results, including Google’s AI Overviews and AI Mode.” Clicks from separate assistants go to the AI Assistant channel defined above. The rule behind it is short:

That rule tells you the main weakness: the channel depends on the referrer. In practice you will see several gaps.

• Missing referrers. Clicks from desktop and mobile apps, and links people copy and paste, often arrive with no referrer and land in Direct. The AI Assistant channel undercounts real AI traffic.
• UTM tags. Links from ChatGPT search commonly carry utm_source=chatgpt.com. Check Session source / medium as well as the channel, so you see such visits even when the channel looks wrong.
• The list is Google's. A new or niche assistant may not be on it yet. Perplexity, for example, usually appears with perplexity.ai as the source; confirm where it lands in your own property.
• Google AI is invisible as a group. AI Overview and AI Mode visits sit inside Organic Search with every other Google click. GA4 cannot separate them.

To watch AI sources reliably, build an exploration or a report filtered on session source with a regex such as the one below, and add landing page and key events. “Default channel groups can't be edited in Google Analytics”, but you can create a custom channel group, or edit the primary channel group, with your own AI rule.

# Session source matches regex (GA4 exploration filter)
chatgpt|openai|perplexity|copilot|gemini|claude|deepseek|grok

Channel setup, attribution and key events are covered in the lesson GA4 for SEO; here you only need the AI view on top of it.

Bing Webmaster Tools: AI Performance

Bing is currently the only major search engine with a first-party citation report. Microsoft introduced it in February 2026 as a public preview; it “shows how publisher content appears across Microsoft Copilot, AI-generated summaries in Bing, and select partner integrations.” The report contains:

• Total Citations. “Shows the total number of citations that are displayed as sources in AI-generated answers during the selected time frame.”
• Average Cited Pages. “Shows the average number of unique pages from your site that are displayed as sources in AI-generated answers per day over the selected time range.”
• Grounding queries. “Shows the key phrases the AI used when retrieving content that was referenced in AI-generated answers.” Microsoft notes that “The data shown represents a sample of overall citation activity.”
• Page-level citations and a timeline. Microsoft is explicit: “This reflects how often pages are cited, not page importance, ranking, or placement.”

In June 2026 four preview features were added: Intents, Topics, Citation Share and Compare. Citation Share “is calculated as the percentage of citations attributed to your site out of all citations shown across all sites for that same grounding query.” Microsoft also lists what it is not: “It does not expose competitor domains, represent traffic share, or assign quality scores to content.”

Use this report as your one official citation series, but do not treat it as traffic. Citations are not clicks, the data is a sample, and it covers Microsoft surfaces and partners only, not ChatGPT, Perplexity or Google.

What you cannot measure officially

Put the three official sources side by side and the gaps become clear.

As of October 2026, neither OpenAI nor Perplexity offers site owners a citation report comparable to Bing's. For those surfaces, your own sampling is the measurement.

Prompt panel tracking: the method

A prompt panel is a fixed list of questions you ask AI systems on a schedule, recording who gets cited and mentioned each time. It works like rank tracking for a keyword set, with one difference: the results are much less stable, so you need repeat runs.

1. Build a fixed prompt set

Start with 30 to 100 prompts taken from real demand: question queries from Search Console, questions your sales team hears, support tickets and on-site search. Write them as people actually ask, in full sentences. Tag each prompt by intent (learn, compare, buy, fix) and as branded or non-branded. Non-branded prompts, such as best accounting software for a small agency, tell you most about visibility; branded prompts, such as is Example CRM good for agencies, show what AI systems say about you.

2. Fix locations, language and setup

Choose the countries and languages you sell in and keep them constant. Run prompts logged out or from a clean profile, without personal memory or custom instructions, so earlier conversations do not shape the answer. Note the device and app or web version.

3. Choose surfaces

Typical surfaces are Google AI Overviews, Google AI Mode, ChatGPT, Perplexity and Microsoft Copilot. Track only those your audience uses; five surfaces times 50 prompts times three runs is already 750 answers a month.

4. Repeat each run

Ask every prompt several times per month on each surface, at least three runs, ideally on different days. A single run tells you almost nothing about a system that changes its sources from one answer to the next.

5. Record citations, mentions and competitors

For each answer, log the date, surface, location, run number, which of your URLs were cited, whether your brand was mentioned, which competitors were cited or mentioned, and a short note on accuracy. Keep the raw answer text or a screenshot so you can recheck later.

6. Keep the panel stable

Keep the same prompts for at least a quarter. If you change half the panel every month, you cannot tell whether a change came from your work or from the new prompts. Add new prompts as a separate group and fold them in at the next quarter.

Third-party AI visibility tools

A growing category of tools runs prompt panels for you at scale. They send your prompts to several AI systems, store the answers, extract cited domains and brand names, and compute share-of-voice figures against competitors you name. They save a lot of manual work, but their numbers are only as good as their method. Before you rely on one, check:

• Where the prompts come from. Your own list, or prompts generated by the tool? Generated prompts may not match real demand.
• How answers are collected. Some tools query models through an API rather than the consumer app. API answers can differ from what people see in the app, especially for search features.
• Locations, language and login state. These should match your audience and stay fixed.
• Runs and sample size. How many runs per prompt, and how often? A single daily run per prompt is still one sample.
• Raw data export. You should be able to read the actual answers, not only a score.

Spot-check a tool against a few prompts you run by hand. If the two disagree often, trust neither until you understand why.

Metrics that matter

Six metrics cover the stack from citation to conversion. Define them once and calculate them the same way every month.

Branded search lift is the indirect measure for mentions. If AI answers name you more often, some people will search for you by name later. Track it with a regex that covers your brand and its common misspellings, for example example|exampel|example crm, and compare it with the same period a year earlier to remove seasonality.

Sampling and volatility

AI answers are generated, not looked up. The same prompt can cite different sources a minute later, in another city or the next day. Microsoft says the same about its own data: “Citation patterns can shift due to changes in user behavior, evolving models, freshness signals, partner refresh cycles, and broader changes across the web itself.”

Treat every panel result as a sample and follow a few rules:

• Report rates over many answers, never a single answer. With 10 answers, one extra citation moves your citation share by 10 points.
• Look at trends across months rather than one month against the previous one.
• Accept a change as real when it holds across runs, across many prompts and for more than one month.
• Note model and product changes. When a surface changes its model or interface, mark it in the log; a jump that week may have nothing to do with your site.

A monthly AI visibility report template

A report that follows the same outline every month lets you and your stakeholders compare like with like.

Behind section 5 keep one example row per surface, for example: ChatGPT, UK, 50 prompts × 3 runs: cited in 42 of 150 answers, mentioned in 61 (illustrative). Absolute counts next to percentages stop readers from over-reading small samples.

One technical point belongs in every report. Google only uses pages that are indexed and eligible for a snippet, so a page that drops out of the index also drops out of AI Overviews and AI Mode. Check indexability of your most-cited pages before you look for any other cause.

Common mistakes

• ✕Looking for AI Overviews in a separate Search Console filterGoogle documents none; AI features are part of Web totals.
• ✕Treating the AI Assistant channel as all AI trafficGoogle AI clicks are in Organic Search, and visits without a referrer are in Direct.
• ✕Reporting a single prompt runOne answer is one sample; report rates over repeated runs.
• ✕Reading Bing citations as trafficCitations are not clicks, and the data is sampled.
• ✕Changing the prompt set every monthTrends become impossible to compare.
• ✕Ignoring unlinked and wrong mentionsA mention without a link still shapes opinion, and a wrong one needs fixing.
• ✕Trusting a tool score without the raw answersIf you cannot read the answers, you cannot check the method.

FAQ

Sources

• Google Search Central — AI features and your website
• Search Console Help — What are impressions, position, and clicks?
• Search Console Help — Performance report (Search results)
• Google Analytics Help — Default channel group
• Microsoft Bing Blog — Introducing AI Performance in Bing Webmaster Tools Public Preview
• Microsoft Bing Blog — New AI Visibility Insights in Bing Webmaster Tools

Quotes checked against the official pages on 30 October 2026.