How Do You Measure Your Presence in AI Overviews?

Measuring your presence in AI Overviews rests on two instruments that chatbot tracking doesn’t use: Google Search Console for the indirect signal, and repeated manual checks for the direct signal. AI Overviews live in the SERP, not in a chat window, and that changes the whole method.

One trap dominates the current period. AI Overviews are rolling out in France progressively since 22 July 2026. Two users see different results on the same query. Any baseline taken this week will stay unstable until the rollout completes.

I’m Florian Zorgnotti, an SEO and GEO consultant. What follows gives you the method specific to AI Overviews: the scissor signal in GSC, the number of checks needed for a reliable figure, and the right denominator to use.

Why measuring an AI Overview has nothing to do with measuring ChatGPT

An AI Overview displays inside Google’s results page, not in a conversational interface. That difference in location dictates the whole method. For ChatGPT or Perplexity, you build a prompt basket and query the engine. For AI Overviews, you have an instrument chatbots don’t offer: your own site’s performance data in Google Search Console.

You gain one signal and lose another. The gained signal is indirect but free and exhaustive: GSC covers all your real queries. The lost signal is the simplicity of the direct test: you can’t “ask” an AI Overview to generate cleanly the way you do with a chatbot prompt.

In my view, this is the most common confusion right now. People ask me to “track AI Overviews the way we track ChatGPT”. The two measurements share neither the same tools, nor the same denominator, nor the same unit of analysis. Treating them identically produces false numbers.

How do you read the scissor signal in Google Search Console?

The scissor signal is the pattern where your impressions stay flat or rise while your clicks and CTR fall. That crossover signals a panel capturing attention ahead of your organic link. It’s the most reliable and cheapest indicator of an AI Overview on your queries.

The method takes four moves in the GSC Performance report.

1. Select a rolling four-week window, then compare it to the previous period.
2. Enable the Clicks, Impressions and CTR metrics together.
3. Sort by CTR decline, then spot the rows where impressions aren’t dropping.
4. Isolate those queries: they’re your candidates for an active AI Overview.

One caution. The scissor proves a panel is capturing clicks; it doesn’t prove you’re cited inside it. You could be the cited source losing the click, or a page displaced by third-party sources. The GSC signal identifies queries to verify; it doesn’t replace visual verification.

The scale of the effect justifies the effort. According to Ahrefs, the average position-one CTR on queries triggering an AI Overview fell to 0.016 in December 2025, against 0.073 two years earlier. The click loss attributed to AI Overviews reaches 58% on affected queries.

How many checks do you need for a reliable figure?

A single check is worthless, because AI Overviews aren’t deterministic. Ahrefs’ research establishes that 45.5% of an AI Overview’s citations change when the panel regenerates. Almost one citation in two disappears or appears from one run to the next. Measuring once means photographing noise.

The consequence is mathematical. To estimate a real presence rate within a few points, one check per query produces an unmanageable margin of error. You have to repeat, across several days, then reason on the average.

Two logics coexist here, and the gap between them deserves honesty. Strict statistics, a proportion at 95% confidence, would demand dozens of checks per query for a ten-point margin. That’s mathematically exact and operationally unusable. In practice, you accept looser confidence in exchange for a sustainable load, and you compensate by tracking the trend over time rather than the precision of a single snapshot.

Here are the pragmatic orders of magnitude I apply, for an estimated presence around 30%. They correspond to roughly 80% confidence, owned as such.

These numbers climb fast. Twenty queries tracked with ten checks each is two hundred verifications per cycle, by hand, in private browsing. For one language and one country. That’s exactly where manual checking hits its limit and a tool becomes worthwhile.

The interactive tool below calculates the number of checks for your estimated presence and target precision, and shows both confidence levels so you choose knowingly.

How do you establish a baseline during a progressive rollout?

Establish your baseline now, but timestamp every check and treat the figure as provisional until things stabilise. The progressive French rollout means AI Overview coverage scales up week after week. A presence rate of 20% recorded today could reach 35% in September without you changing anything.

Good practice means separating two curves. The first tracks the trigger rate: across your strategic queries, how many display an AI Overview. That curve reflects Google’s rollout, not your work. The second tracks your conditional presence rate: among queries that trigger an AIO, on how many are you cited. That curve reflects your real visibility.

Never mix the two. If you track a raw presence rate across all your queries, the rollout’s rise will inflate or mask your results and you’ll draw false conclusions. By isolating the conditional rate, you neutralise the rollout effect and measure what depends on you.

This transitional period is an opportunity. Your competitors are still watching without measuring. A clean baseline set today, even provisional, will give you three months from now a depth of history no one else has.

What’s the right denominator for your presence rate?

The right denominator is genuinely citable queries, not all your queries. Two filters remove the queries that structurally can’t cite you.

First filter, AI Overviews that cite no source. Ahrefs measures that 11% of AI Overview responses display no source. On those queries, nobody is cited: including them in your denominator artificially crushes your rate.

Second filter, AI Overviews that name no brand. Also per Ahrefs, 59.41% of AI Overview responses mention neither brand nor person. On a purely factual query with no brand dimension, your absence carries no competitive meaning.

The right fraction is therefore:

• Numerator: the number of queries where your brand is cited or named.
• Denominator: the number of queries triggering an AIO that cites at least one brand in your sector.

With this framing, a rate of 40% means something: on queries where a spot was up for grabs, you hold four out of ten. A raw rate calculated across all your queries means nothing.

How do you tell citation, mention and recommendation apart?

Three levels of presence coexist in an AI Overview and aren’t equal. Confusing them distorts your reading as much as a bad denominator does.

A citation is a source link displayed beside the answer. It brings you a potential click, but not necessarily a mention of your name in the text. A mention is your brand appearing in the body of the answer, named explicitly. It builds reputation, independent of the click. A recommendation is the case where the AI Overview actively presents you as a solution, an advised choice.

A common case shows the gap. Your page can be the cited source of an answer that recommends a competitor by name. You win the potential click, you lose the reputation battle. If you only measure citations, you miss that drop. This citation-mention-recommendation triptych is the frame I use in every audit.

The complete measurement protocol, step by step

Here is the sequence I’ve been applying with clients since the French rollout. It combines the instruments above into a sustainable tracking routine.

1. Frame the scope. Twenty to thirty strategic queries, the ones that weigh in your acquisition. Not five hundred generic queries.
2. Filter the denominator. On each query, first check whether an AIO triggers and whether it cites at least one brand. Set the others aside.
3. Check repeatedly. Eight to ten verifications per query, in private browsing, spread over several days, for usable monthly tracking.
4. Record the three levels. For each check, distinguish citation, mention and recommendation, and note the brands cited in your place.
5. Cross-reference with GSC. Link your scissor queries to your direct checks to prioritise real click losses.
6. Historise and timestamp. Separate trigger rate from conditional presence rate, and timestamp everything while the rollout isn’t stabilised.

This protocol is heavy by hand beyond a dozen queries. That’s the logic I built Cockpyt AI on with Laurent Séjourné: weekly measurement across ChatGPT, Gemini and Perplexity, share of voice against competitors, a Google Analytics 4 connection linking visibility to real traffic, and a monthly GEO audit producing a prioritised action plan. Flat price at 29 euros a month billed annually, 14-day trial with no credit card.

The mistakes that distort your measurement

• Measuring once. With 45.5% of citations changing between two generations, a single check is a random draw.
• Taking an undated baseline during the rollout. Your figure will move on its own and you’ll wrongly attribute it to your actions.
• Calculating across all your queries. AIOs with no source and no brand crush your rate for no reason.
• Confusing citation and recommendation. Being the source of an answer that recommends a competitor is not a win.
• Stopping at the GSC signal. The scissor locates queries to verify; it doesn’t say whether you’re in the panel.

FAQ

Does Google Search Console display AI Overviews directly?

No, not as a dedicated metric to date. GSC doesn’t label AI Overview impressions separately. You read their presence indirectly, via the scissor signal, flat or rising impressions with falling clicks and CTR. It’s a reliable indicator for spotting affected queries, to be confirmed by visual verification.

How many checks do you need to measure AI Overview presence?

Count three to five checks per query for a rough trend, eight to ten for usable monthly tracking, fifteen to twenty for a fine competitive benchmark. These pragmatic orders of magnitude target roughly 80% confidence: strict 95% statistics would demand far more, at an unusable load. The need to repeat comes from the variance Ahrefs measured, 45.5% of citations changing between generations. Below three checks, you measure noise.

Can I measure my AI Overviews without a paid tool?

Yes, on a small scope. GSC is free for the indirect signal, and manual checking in private browsing only costs time. The limit arrives fast: beyond a dozen queries tracked with repeated checks, several competitors and a dated history, the manual load becomes unmanageable and a dedicated tool becomes more worthwhile than the time it saves.

Why does my AI Overview presence change without me modifying anything?

Two causes combine. First the panel’s intrinsic variance, regenerating nearly one citation in two from run to run. Second, in France in 2026, the progressive rollout raising coverage week after week. To tell the two apart, separate your trigger rate from your conditional presence rate.

Should you measure AI Overview and AI Mode presence separately?

Yes. The two surfaces cite the same URLs only 13.7% of the time, per Ahrefs. A single metric aggregating them averages two nearly independent realities and produces a misleading figure. Each surface deserves its own check, with its own denominator.

What presence rate should you target in AI Overviews?

There’s no universal target, because everything depends on competition on your citable queries. The right reference isn’t an absolute threshold but your relative share of voice against competitors measured on the same basket. A 40% presence rate on hotly contested queries beats 70% on queries nobody wants.

Is AI Overview traffic visible in Google Analytics 4?

Only partly. A click from an AI Overview source link arrives as standard Google organic traffic, with no reliable distinctive label to date. GA4 shows you overall organic traffic but doesn’t natively isolate the share from an AI panel. It’s a structural limit that makes direct checking all the more necessary.

Sources

• Despina Gavoyannis and Xibeijia Guan, “Are AI Mode and AI Overviews Just Different Versions of the Same Answer?”, Ahrefs, 15 December 2025 — ahrefs.com
• Ryan Law and Xibeijia Guan, “Update: AI Overviews Reduce Clicks by 58%”, Ahrefs, 4 February 2026 — ahrefs.com
• Louise Linehan, “Update: 38% of AI Overview Citations Pull From Top 10 Pages”, Ahrefs, 2026 — ahrefs.com
• Johan Sellitto, “Google lance officiellement les AI Overviews en France !”, Abondance, 22 July 2026 — abondance.com