Key takeaways:
- Three sources complement each other: Search Console for exposure, a tracking tool for what the answer says, Google Analytics 4 for what happens after the click. None is sufficient alone.
- The Search Console “Generative AI features” report, launched on 3 June 2026, shows impressions only. No clicks, no CTR, no position, no queries.
- Every follow-up question in AI Mode counts as a new query, so one research session appears as several separate searches in your data.
- Share of voice aggregated across repeated runs remains the only usable indicator, given volatility measured at 9.2% URL overlap.
Measuring your presence in Google AI Mode means assembling three partial views. No single tool gives you the complete picture today, and that is precisely what most published methods leave unsaid.
On 3 June 2026, Google shipped a dedicated report for generative AI features in Search Console. It was the number one request from search professionals for two years. Its current version exposes impressions in AI Overviews and AI Mode, broken down by page, country, device and date. It exposes neither clicks, nor click-through rate, nor position, nor queries.
Let me put it plainly: this report tells you how many times you were shown, never why or to whom. The rest of the measurement work happens elsewhere, and this article details where.
Does ChatGPT recommend your brand?
Measure your presence and identify the brands cited in your place. No credit card required.
How do you measure your presence in Google AI Mode?
You measure your presence in Google AI Mode by combining three sources that each answer a different question. Search Console answers “how often was I exposed”. An AI visibility tracking tool answers “what does the answer say when it discusses my sector”. Google Analytics 4 answers “what do visitors arriving from those surfaces actually do”.
One of these three views rarely suffices. Taken alone, each produces a false conclusion. Rising impressions without clicks look like a malfunction when they describe the normal behaviour of generative surfaces. A high share of voice on queries nobody types is worthless. And growing AI traffic says nothing about your standing against the competitors cited alongside you.
The 3 measurement sources and what each one misses
The table below places the three sources side by side with their respective blind spot. This is the reading most published methods leave out.
| Source | What it measures | Its blind spot | Cost |
|---|---|---|---|
| Search Console | Impressions in AI Mode and AI Overviews, by page, country, device and date | No queries, no clicks, no position, no split between AI Mode and AI Overviews | Free |
| AI tracking tool | Share of voice, competitors cited, sentiment, sources reused in the answer | Real query volume unknown, sample chosen by you | Paid |
| Google Analytics 4 | Sessions, engagement and conversions from AI surfaces | Only sees visitors who clicked, a minority | Free |
| Manual check | Exact wording of the answer, order of brands cited | Not reproducible at scale, biased by personalisation | Time |
Search Console: exposure without context
The “Generative AI features” report sits in the Performance section of Search Console. Google states in its announcement that these reports provide dedicated views of impressions within generative AI features on Search, such as AI Overviews and AI Mode, as well as in Discover.
Two methodological caveats apply. First, the rollout happened in waves across a subset of sites, so not every property has access. Second, the report aggregates AI Overviews and AI Mode without separating them, which matters: the two surfaces cite the same URLs in only 13.7% of cases according to Ahrefs’ analysis of 540,000 query pairs.
Tracking tools: content without volume
An AI visibility tracking tool queries the engines on your behalf, across a list of queries you define, and analyses the content of the answers returned. It gives you access to what Search Console will never show: the wording used, the competing brands cited next to you, the sources reused in the answer.
Its limitation lies in the sample. You choose the queries tested, so you measure your visibility against your own representation of the market. If your 20 prompts do not match how your customers actually phrase things, your share of voice measures an imaginary universe.
In my view, the correct hierarchy runs like this: Search Console confirms the signal exists, the tracking tool explains it. Starting with the tool without checking GSC means optimising a share of voice on queries that generate no real exposure.
Google Analytics 4: after the click only
GA4 isolates sessions coming from AI surfaces through referrers. This measurement answers the one question a management team cares about: does this traffic convert. Its structural weakness is volume. Most AI Mode users do not click, as iPullRank’s readings across the product’s early months in the United States confirm.
What the Search Console report will not tell you
Three limitations of the GSC report change how you should read your numbers. Google documents them, but their practical consequences are rarely explained.
No clicks, no queries, no position
The current version of the report is limited to impressions. You know how often your pages were displayed in an AI surface, never on which questions or with what outcome. Google says it is working with publishers to determine the most useful metrics and plans to add more over time.
Direct consequence: build no performance reporting on this report alone. Watch the trend, not the absolute value. An impressions curve climbing while your clicks flatten describes normal generative surface behaviour, not a technical anomaly.
The follow-up question trap
AI Mode works through successive exchanges. Google treats each follow-up question as a new search, and attributes all impressions, clicks and positions to that new query. Positions restart at one.
A session where a user asks four chained questions therefore produces four distinct queries in your data, with no marker indicating they belonged to the same journey. Any attempt to reconstruct a research funnel from these figures fails.
One impression, not two
When the same URL appears simultaneously inside an AI response and among the classic links on the same results page, Google counts a single impression. This rule avoids double counting but makes it impossible to calculate a differential between organic and generative exposure on the same query.
The 5 KPIs that actually matter in AI Mode
Five indicators withstand the system’s volatility. The others produce noise.
- Aggregated share of voice. The percentage of runs where your brand appears, across all readings performed. The only robust indicator against volatility measured at 9.2% URL overlap between three runs of the same keyword.
- Generative impressions. The GSC report curve, read as a four-week rolling trend, never as a point value.
- Competitors cited in your place. The list of brands appearing when you are absent. This indicator drives your action plan directly.
- Associated sentiment. The tone of the description when the answer mentions you. A brand cited but poorly described has a different problem from an absent brand.
- Sessions and conversions from AI surfaces. The only KPI that speaks to a management team.
The 6-step measurement protocol
Here is the sequence I apply, in this order. Each step conditions the next.
- Check access to the GSC report. Open Search Console, Performance section, and look for the view dedicated to generative AI features. If it is missing, your property is not yet in the rollout scope.
- Export your generative impressions by page. Identify the 20 URLs concentrating the most AI impressions. These are your pages already recognised by the system.
- Build your prompt list from those pages. Write 20 to 40 questions matching the topics actually covered by those URLs, in your customers’ language and phrasing.
- Run each prompt at least five times. Record your brand’s presence, the competitors cited and the tone used. A single reading is not usable.
- Calculate share of voice by topic. Group prompts by subject rather than producing one global figure. An average score hides the gaps that matter.
- Connect everything to GA4. Cross share of voice by topic with sessions from AI surfaces on the matching pages.
How often should you measure your presence in AI Mode?
Weekly measurement is the right rhythm for operational tracking. Below that, you confuse volatility with trend. Above it, you produce noise without gaining precision.
In my view, the classic mistake is measuring daily for the first few weeks out of curiosity, then abandoning it after a month for lack of time. A weekly reading sustained for six months beats a daily reading sustained for three weeks. The value of this measurement comes from the series, not the point.
That is the logic behind Cockpyt AI: one prompt variation measured each week across ChatGPT, Gemini and Perplexity, completed by a monthly GEO audit with a prioritised action plan and a GA4 connection linking visibility to real traffic.
FAQ
Does Search Console measure AI Mode separately from AI Overviews?
No. The report dedicated to generative AI features aggregates both surfaces without distinguishing them. This limitation weighs all the more since the two systems cite the same URLs in only 13.7% of cases.
Why are my impressions rising while my clicks fall?
This decoupling characterises how generative surfaces work. Your page is displayed as the source of an answer that satisfies the user without a click. The measuring instrument works, the results page changed.
How many prompts do you need for reliable measurement?
Between 20 and 40 prompts covering your priority topics, each run at least five times. Reliability comes from repeated runs as much as from the number of distinct queries.
Can you measure your AI Mode presence for free?
Partly. The Search Console report and GA4 are free and cover exposure and post-click behaviour. Manually checking answers remains possible but becomes time-consuming beyond a dozen prompts tracked each week.
Should you opt out of appearing in AI responses?
Google offers a setting to remove your content from generative surfaces without affecting indexing or organic ranking. This choice means surrendering AI visibility with no ranking benefit in return, which rarely makes sense outside the case of press publishers.
Is the GSC report data retroactive?
The report arrived with limited history and without full reconstruction of earlier periods. Export your data regularly to build your own long series.
Sources
- Google Search Central, Introducing Search Generative AI performance reports in Search Console, 3 June 2026. developers.google.com
- Google Search Console Help, What are impressions, position, and clicks?, AI Mode section. support.google.com
- Ahrefs, AI Overviews vs AI Mode, December 2025, analysis of 540,000 query pairs. ahrefs.com
- SE Ranking, AI Mode Research: Sources, Volatility, and Differences between AIO and Organic Search, August 2025, 10,000-keyword study. seranking.com
