Short answer: Visibility in AI answer engines is measured with four metrics: mention rate (in what share of answers your brand appears), citation rate (how many of those answers link to your content), share of voice against competitors, and accuracy of what is said about you. Three data sources feed them: the Generative AI performance report launched in Google Search Console on 3 June 2026, the AI assistant channel in GA4, and a manual prompt panel run once a month.
Why the old metrics stop working
For twenty years SEO was measured with three numbers: position, impressions and clicks. In a world where a large share of searches is answered inside an answer box, the first two keep rising while the third stays flat. That is exactly the pattern many sites see today: average position and impressions improving, click-through rate eroding.
The explanation is not that titles got worse. The user simply got the answer without clicking. So the right question is no longer “where do we rank” but “are we inside the answer, and who is cited in it”.
The four metrics that do work
| Metric | How to calculate | What it tells you |
|---|---|---|
| Mention rate | Answers mentioning the brand divided by relevant prompts | Whether the engine knows you exist in the category at all |
| Citation rate | Answers linking to your domain divided by total answers | Whether your content is the source, or people merely talk about you |
| Share of voice | Your mentions divided by all competitor mentions in the same panel | Your relative standing in the category |
| Accuracy and sentiment | Manual classification of every mention: correct or wrong, positive, neutral or negative | Whether the engine tells a true story about you. Easy to ignore, expensive to discover late |
The three data sources
1. The Generative AI report in Search Console
On 3 June 2026 Google launched dedicated Search Console performance reports that break out impressions inside generative AI features on Search: AI Overviews, AI Mode, and generative AI features in Discover. The breakdown is by page, country, device and date.
Two caveats matter. First, at this stage the report shows impressions only — no clicks, no CTR, no query data. It tells you that you were seen, not what it was worth. Second, the rollout is gradual to a subset of sites, so the report may not have appeared in your property yet. These impressions were always included in the overall performance totals, so this is a breakout of existing data rather than new data. Alongside it, Google added a control that lets a site block its content from appearing in those generative features.
2. The AI assistant channel in GA4
GA4 now recognises some assistants as a separate channel, but coverage is partial: traffic arriving without a referrer falls into Direct, and several platforms add no parameters at all. ChatGPT appends utm_source=chatgpt.com to citation links, while Perplexity, Claude and Gemini generally do not. The practical recommendation is to build a custom channel group with a regular expression on session source:
chatgpt.com|chat.openai.com|perplexity.ai|claude.ai|gemini.google.com|copilot.microsoft.com|you.com
Then look at the landing-page-by-source report. That is the view that shows which of your pages are actually being cited and which are not.
3. A manual prompt panel
This is the only source that measures mentions and sentiment rather than just traffic. Build a fixed list of 25 to 40 prompts representing the questions customers really ask, run each one three to five times on every engine, and log the results. Five runs rather than one, because language models are not deterministic and a single answer is an anecdote, not a measurement.
How to build a prompt panel that produces a reliable number
- Three prompt categories. Generic (“how do I choose an X provider”), comparative (“X versus Y”), and branded (“is [brand] good for Z”). Each measures a different stage of the journey.
- Lock the wording. Changing one word changes the answer. The panel must be identical month to month, otherwise you are measuring the prompt rather than your visibility.
- Clean sessions. No history, not logged in, and ideally from the country your customers ask from. Otherwise you are measuring personalisation.
- A fixed sheet. Date, engine, prompt, run number, mentioned, cited, link, sentiment, factual error. Nine columns is enough.
- Monthly cadence. More often produces noise. Less often misses model changes.
What each source gives you
| Source | What you get | What is missing |
|---|---|---|
| Search Console, Generative AI report | Impressions in AI Overviews and AI Mode, by page and country | Clicks, CTR and queries |
| GA4 with a custom channel group | Sessions and conversions by engine and landing page | Everything arriving without a referrer |
| Server logs | Which bots crawl, which pages, how often | No direct link between a crawl and a citation |
| Manual prompt panel | Mentions, citations, share of voice, sentiment, accuracy | Manual effort, and a sample rather than a population |
What to do with the numbers
Three scenarios cover almost every case. If mention rate is low, the problem is absence from the corpus: no dedicated pages on the topic, no presence in the third-party sources engines trust, no clearly defined entities. If mention rate is high but citation rate is low, the problem is structural: your content is not written so an answer can be extracted from it, so the engine summarises you without linking. If accuracy is low, there is outdated or contradictory information on the web that needs correcting at the source.
The structural lever is the cheapest to fix: a direct answer in the opening paragraph, headings phrased as questions, tables instead of paragraphs, explicit numbers, and an FAQ section marked up with FAQPage schema. Pages updated within the last year tend to earn more citations than older ones, which is why consistently refreshing existing pages usually pays better than writing a new one.
Frequently asked questions
Does Search Console show clicks from AI Overviews?
Not at this stage. The Generative AI report launched in June 2026 shows impressions only, broken down by page, country, device and date, with no clicks, CTR or query data.
How many prompts does a panel need?
25 to 40 prompts is enough for most businesses, split between generic, comparative and branded. Keeping the wording fixed matters more than making the list long.
How many times should each prompt be run?
Three to five runs. Language models are not deterministic, so a single answer is an anecdote rather than a measurement.
Why does traffic from Perplexity and Claude not show up in GA4?
Because they usually do not append campaign parameters, and when the referrer is lost the session falls into Direct. The fix is a custom channel group with a regular expression on session source.
Does falling CTR always mean the titles are bad?
No. If position and impressions rise while CTR falls specifically on broad informational queries, and narrow commercial queries keep a healthy CTR, that is the signature of an answer box absorbing the click rather than a weak snippet.