Get a free AI visibility report
Search Console now reports AI Overviews impressions but not AI Overviews clicks. This guide covers what the generative AI performance report measures, why an AI-specific click-through rate cannot be calculated from it, and how to estimate click loss anyway.
For two years, measuring AI Overviews traffic meant inference and third-party estimates. Google had the data and site owners did not. That changed in June 2026, and the change is smaller than most coverage suggested. You now get impressions. You still do not get clicks.
This article works through what that actually leaves you with. The frame throughout is the post-mention journey, meaning what happens between a user seeing your link inside an AI answer and a session appearing in your analytics. Most of that path is unmeasured by design. That makes the real question less "what is my AI Overviews click-through rate" and more "how do I build a defensible estimate of what I lost and where." Tracking AI Overviews traffic properly involves Search Console, but it does not end there.
The generative AI performance report sits inside Search Console under Performance, below the Search results report. Google announced it on 3 June 2026, describing it as a dedicated view of impressions within generative AI features, with the data continuing to be tracked in the overall performance report as well.
That second half of the sentence is the part people skip. It matters more than the launch itself, and it shapes every AI Overviews traffic calculation you will attempt.
Google's help documentation for the report confirms the scope. It covers AI Overviews and AI Mode on Google Search. Generative AI features in Discover get a separate report. Data from Search Labs experiments is excluded. As of 31 August 2026 the report is live on all properties worldwide, so every site now has some AI Overviews traffic data available.
You get one metric, impressions, defined as how often links to your site were shown to a user inside a generative AI feature. You can group that metric four ways:
The export button downloads both chart and table. Values displayed as a tilde or a dash come through as zeros. That quietly breaks any AI Overviews traffic spreadsheet that treats zero as a real observation.
No clicks. No click-through rate. No average position. No queries.
There is also no split between AI Overviews and AI Mode. A citation in a summary above the blue links and a citation inside a conversational session count identically, despite being very different user journeys. The report is not currently exposed through the Search Analytics API either, so automated pipelines have to work from manual CSV exports for now.
The 1,000-row table limit from the standard Performance report applies here too. On a large site that limit alone means your page-level AI Overviews traffic view is a sample rather than a census.
Here is the detail that reframes the entire launch. It does not live in the new report's documentation at all. It sits in Google's older help page on impressions, position and clicks. There, Google states plainly that clicking a link to an external page in the AI Overview counts as a click, and the same for AI Mode.
So the clicks are not missing. They are counted, and they have been counted the whole time.
Your AI Overviews traffic is inside your Web search type total right now, blended with ordinary blue-link clicks and indistinguishable from them. The generative AI report gives you impressions for AI surfaces. The Performance report gives you clicks for everything.
You have a numerator in one table and a denominator in another, with no key between them. That is why an AI-specific click-through rate cannot be calculated from Search Console, and why anyone showing you one has either estimated it or made it up.
This is also why Cognizo treats citation share as a metric in its own right rather than a derivative of clicks. When the click cannot be attributed to the surface that caused it, measuring the citation directly is the only honest route left.
Google documents that the generative AI report draws from the Web search type in the Performance report. The AI impressions were always inside your totals. The new report isolates them.
Two practical consequences follow. Your headline impression numbers did not change on the day the report appeared, so an apparent jump in AI visibility is a reporting artefact rather than a real gain. And when you compare the two reports, you are looking at a subset against its parent, not two channels. Adding them inflates your AI Overviews traffic numbers.
Subtracting them to isolate classic search performance is only slightly better. It works as narrative. It fails as a KPI, because the newest data is preliminary and the two reports do not share every aggregation rule.
Before you can measure AI Overviews traffic changes, you need to know which existing metrics have quietly stopped meaning what they used to. Three have.
Google's documentation states that an AI Overview occupies a single position in search results, and all links inside it are assigned that same position.
Think about what that does to average position. If your page is cited inside an AI Overview at position one, your reported position is one, whether your link was listed first or sixth. If the same page also ranks organically at position eight, Search Console records the topmost position, which is the overview slot.
Average position on AI-heavy query sets therefore tends to look better than it did, while clicks fall. That combination is the most common misreading in AI-era Search Console data. Teams see position improving and conclude the SEO is working. Answer Engine Insights reports position within the AI answer itself for exactly this reason, since the Search Console figure describes the overview's slot rather than yours. Our guide to AI Overviews rank tracking covers what to watch instead.
In the generative AI report, the chart aggregates by property. If two results from your site appear in the same AI answer, they count as one impression. The table changes behaviour by dimension. Grouped by country, device or date it aggregates by property. Grouped by page it aggregates by page.
Google notes that chart totals and table totals can therefore differ. That is documented behaviour, not a bug, and it will generate a support ticket from someone on your team eventually.
In AI Mode, a follow-up question is treated as a new query. All impression, position and click data in the new response is counted against that new query.
For measurement this means AI Mode impressions inflate relative to session count. One user working through a five-turn conversation can generate five impressions. Since the report does not split the two surfaces, any property with heavy AI Mode exposure has an impression curve that is partly a conversation-depth curve. Our guide to tracking your brand in Google AI Mode covers that surface on its own terms.
No single view answers this. The method below builds an AI Overviews traffic estimate from parts you can actually verify, which is the honest version of the question.
Pull 12 to 16 months of Performance report data at page level: clicks, impressions, CTR and position. Export it before you start. Comparison windows inside the interface will not hold the baseline you need.
You are looking for the period before AI Overviews became common on your query set, which varies by market and vertical. AI features rolled out to different countries at different times. A property selling into several markets therefore has several AI Overviews traffic baselines, not one.
This is the core of the method, and it is where most AI Overviews traffic analysis actually happens. For each priority page, put two columns side by side over an identical date range:
Pages with high AI impressions and flat or falling clicks are source material that sends no traffic. That is your click loss candidate list.
Crucially, this is not automatically bad. A glossary page cited without a click may be exactly what you intended. The point is that it should be a deliberate choice rather than a discovery you make a year later. Joining the two exports by hand gets tedious past a few hundred URLs, which is what our Google Search Console integration exists to remove.
A page can lose clicks for two completely different reasons, and the remedy differs.
Ranking loss means you fell in the results and fewer people saw you. Impressions drop, clicks drop, position worsens. This is an ordinary SEO problem with an ordinary SEO fix.
Click loss means you are still being seen, and increasingly seen inside an AI answer, but fewer people click through. Impressions hold or rise, position holds or improves, clicks fall. Anyone assessing AI Overviews SEO impact on website traffic should treat this pattern, not a ranking decline, as the signature.
Sorting your pages into those two buckets before assigning work is the highest-value hour in the entire process. Cognizo runs the split against your tracked prompt set daily, so the buckets stay current rather than aging into a stale spreadsheet.
Query-level data still exists in the standard Performance report, even though the generative AI report has no query dimension. Use it.
Split your query set into informational, comparative and transactional groups. Informational queries are where AI Overviews resolves the question in place most often, so that is where click loss concentrates. Transactional and navigational queries hold up better, because the user has to reach a specific destination to finish the task.
If your click loss is concentrated in informational queries and your transactional traffic is stable, the correct response is a content-mix decision, not an emergency. Google's query report only tells you what people typed into Google. Cognizo's Prompt Volumes, built on billions of real-world signals, tells you what they are asking AI. That is a different question set, and it is where your click loss is coming from.
This is the part almost nobody instruments, and it is where the post-mention journey actually resolves.
A user sees your brand cited in an AI Overview, does not click, then searches your brand name directly two days later. Search Console records that as a branded query with a click. Nothing connects it to the AI Overview that caused it.
So track branded impression and click volume as a trend line alongside your AI impressions. A rising branded curve running against falling non-branded clicks is the clearest available evidence that AI citations are producing delayed, unattributed demand. Our piece on the AI-driven customer journey covers the full path this creates.
Pair it with a "How did you hear about us?" field on your forms, carrying an explicit AI option. That single question recovers more attribution than any amount of UTM engineering, because the journey you are measuring carries no UTM at all.
Search Console and your analytics tool disagree, always, and the disagreement is informative rather than a bug to fix.
Search Console counts a click when someone leaves Google for your page. Your analytics counts a session when the page loads and the tracking fires. Bounces before load, consent refusals, ad blockers and prefetching all sit in the gap. That gap tends to widen on AI surfaces, since the traffic skews toward users who arrived already half answered.
Compare landing-page sessions in analytics against Performance report clicks for the same pages and dates. A widening divergence on high-AI-impression pages is worth investigating, not reconciling away. Our guide to AI traffic analytics in GA4 covers the setup, including the referral sources worth isolating.
One thing to avoid: do not build a channel grouping called "AI Overviews" in your analytics. Referrals from Google AI surfaces arrive as ordinary organic traffic. Any such grouping is a guess dressed as a dimension.
Most of the damage in this area is done in reporting rather than in measurement. Three rules keep an AI Overviews traffic report defensible.
If a figure came from a subtraction, a ratio across two reports, or a reconstructed baseline, say so on the slide. An estimate presented as a measurement becomes a target within a quarter. Nobody hits a target built on an assumption.
The useful output is not "we lost X percent of traffic." It is a table of pages showing AI impressions, clicks, direction of travel and the bucket each page falls into. That format survives questions. A single percentage does not.
Absolute AI impression counts mean little on their own. They depend on how often AI features trigger for your query set, which you do not control. Direction over time is the signal. Set the comparison window to match a release or a campaign, and annotate what changed.
Measurement without a response is a nicer-looking decline. Three moves matter, in this order.
If a page is cited constantly and clicked rarely, the click was never the value. Decide what the page now does for you: earning the citation, shaping how the model describes your category, or feeding brand recall.
Then judge it on that. Keeping a click target on a page whose job has changed guarantees a quarterly conversation about failure that nobody can fix.
Click loss concentrates on questions an AI can fully answer in a paragraph. The pages that survive offer something a summary cannot replace: original data, calculators and tools, situation-specific pricing, comparisons requiring judgment, and first-hand testing.
This is not a call to abandon informational content, which still earns the citations that build authority. It is a call to stop giving that content click targets it can no longer hit. Cognizo's Content Optimization prioritises which pages to rework first by tying the recommendation back to the visibility data that justified it.
Being the cited source in an AI Overview shapes the answer, whether or not a click follows. Google's systems pick those sources. The competition for the slot is real even when the traffic is not.
Our analysis of what sources Google AI Overviews cite most breaks down which domain types win those slots. A large share of citations point at third-party domains, so a strategy confined to your own site addresses only part of the problem.
Search Console is first-party, free and authoritative. It is also scoped to Google, and that scope is the limitation.
You cannot see which question produced an AI impression. You get the page and the country, and you can guess from your query report, but the join does not exist.
That matters because the phrasing of a prompt changes the answer. Two ways of asking the same thing can return completely different source sets. An impression count cannot show you that.
This is the gap Prompt Volumes is built for, mapping the prompts buyers actually use so coverage is measured against the real question space rather than the slice Google chose to report.
AI Overviews and AI Mode are two surfaces. Your buyers also use ChatGPT, Gemini, Perplexity, Copilot, Claude, Meta AI, Grok and DeepSeek. None of them report anything to you.
If your reporting covers only the Google surface, you are measuring the part that happens to be instrumented rather than the part that matters. Cognizo's Answer Engine Insights tracks six metrics across every supported platform instead: Visibility Score, share of voice, citation share, source mention rate, sentiment and positioning accuracy, broken down by model, topic, prompt and region.
Search Console reports Googlebot activity. It says nothing about the other crawlers, or about what happens after they visit.
Connecting crawler behaviour to sessions and conversions per platform is what Cognizo's AI Traffic Analytics handles. It covers GPTBot, ClaudeBot, OAI-SearchBot and other major bots alongside the human referral traffic that follows. That puts the Google surface and everything else in one view rather than two.
Being the first source listed in an AI Overview and being the sixth are not the same commercial outcome. Search Console cannot tell them apart.
The report gives you an impression when your link was shown. It says nothing about where in the answer you appeared, how prominently you were attributed, or which passage of your page the system used. Position does not help here either, since every link in an overview inherits the overview's own position.
This is the gap that separates knowing you were cited from knowing whether the citation was worth anything. A brand named in the opening sentence of an answer carries far more weight than one listed in a source tray at the bottom, and both look identical in your impression count.
Capturing that requires reading the rendered answer as a user sees it rather than counting entries in a report. Cognizo does this through UI scraping, which captures the answer as displayed instead of sampling an API response, so placement and framing survive into the data.
An impression confirms presence. It says nothing about whether the description attached to your brand was accurate.
Positioning accuracy is a separate failure mode from absence. A model can cite you consistently while describing your product as something it is not, and no impression count will ever surface that. Sentiment works the same way. Both are tracked per prompt in Answer Engine Insights, because an aggregate score hides the single prompt that is describing you wrongly.
AI answers are not stable. The same query returns different sources across days as index freshness, model updates and competitor publishing shift the candidate set.
Monthly sampling gives you a number with no way to separate trend from noise. It also leaves a sudden drop undetected for weeks. Continuous tracking is what lets you attribute a change to a specific release or a competitor move. That is the difference between reporting a decline and explaining one.
Everything above establishes the same conclusion four times over. Search Console tells you that you appeared. It will not tell you what it was worth, who asked, what they were told about you, or what happened on the nine surfaces Google does not own.
Cognizo is built for that half of the problem. Here is specifically what it does that Search Console cannot.
Cognizo tracks ChatGPT, Google AI Overviews, Google AI Mode, Gemini, Perplexity, Microsoft Copilot, Meta AI, Claude, Grok and DeepSeek. AI Overviews and AI Mode are tracked as separate surfaces rather than blended into one impression curve. That is exactly the split Search Console refuses to give you.
It also captures those answers through UI scraping, meaning the response as a real user sees it on screen. An API sample can differ from the live surface. Reading the rendered answer is how placement and framing survive into your data, and it closes the gap identified two sections above.
Answer Engine Insights reports Visibility Score, share of voice, citation share, source mention rate, sentiment and positioning accuracy, broken down by model, topic, prompt and region.
Impressions answer whether you showed up. These answer whether showing up did anything. They also catch the failure mode where a model cites you constantly and describes you wrongly.
Tracking runs daily on every tier. AI answers change day to day. A monthly sample cannot separate a real decline from ordinary variance, and it leaves a drop undetected for weeks.
Content Optimization turns a visibility gap into a prioritised brief and a draft, tied back to the data behind it. Prompt Volumes tells you which prompts to cover next. AI Traffic Analytics connects crawler behaviour to sessions and conversions per platform.
MCP and API access ship on every paid tier rather than sitting behind an enterprise contract. The data goes into your own reporting, your agent or your warehouse without a sales conversation first.
Pricing is $499 a month on Growth and $999 on Pro. Enterprise runs on custom terms, covering all platforms and multiple workspaces. Both paid tiers include the content agent with auto-publishing, so measurement and the response to it sit in one place.
Keep using Search Console. It is the authoritative first-party record of the Google surface and nothing replaces it there. Just stop asking it questions it was never built to answer.
Treating the generative AI report as a new traffic channel. It is a breakout of impressions you already had. Nothing was added on launch day.
Calculating an AI CTR. You cannot, and a number that looks like one is an estimate wearing a disguise. Label it as such or drop it.
Reading improving average position as good news. On AI-heavy queries it usually means you are cited inside an overview rather than ranking better beneath it.
Citing click-loss percentages from studies without checking the methodology. Most circulating figures come from vendor research with varying definitions of a loss, different query samples and different markets. Interest in phrases like "AI Overviews traffic stealing SEO risks 2025" tells you how heated the conversation became, not what your own numbers are. We keep a running view of the credible figures in our Google AI Overviews statistics roundup, but your own property is the only authority on your own AI Overviews traffic.
Comparing across the rollout boundary. AI features launched in different countries on different dates. A year-over-year comparison straddling a market's launch date measures the rollout, not your performance.
No. The generative AI performance report contains impressions only, with no clicks, click-through rate, position or query data. Google does count AI Overviews traffic, since clicking a link to an external page in either surface counts as a click. Those clicks land in your overall Performance report, pooled inside the Web search type, and no filter separates them. Google has indicated it will add metrics over time, though no timeline has been published for clicks.
Open Search Console, select your property, and look under Performance in the left-hand menu. The Generative AI entry sits below Search results, with a separate entry for Discover underneath. If it is not showing, the likely reasons are that your site has not received enough impressions in AI features, or that your property was excluded from those features through the exclusion control. The report reached all properties worldwide on 31 August 2026.
It depends heavily on your query mix, and your own data is the only reliable answer. Click loss concentrates on informational queries where a summary resolves the question in place. Transactional and navigational queries hold up better. To measure Google AI Overviews impact on SEO traffic, use the impression-to-click gap method: compare generative AI impressions against clicks for the same pages over identical dates, then separate pages that lost visibility from pages that kept visibility and lost clicks.
There is no difference in the report, which is the problem. Google groups both surfaces into one impressions figure with no per-feature breakdown, so a citation above the blue links counts the same as a citation inside a conversational AI Mode session. Those are very different journeys with different click behaviour, and blending them makes AI Overviews traffic harder to isolate. AI Mode also counts each follow-up as a new query, inflating impressions relative to sessions.
Because an AI Overview occupies a single position and every link inside it inherits that position. If you are cited in an overview at the top of the page, your reported position reflects the overview's slot rather than where your link sat within it. Search Console also records your topmost position per query, so an AI citation can override a weaker organic ranking. Improving position alongside falling clicks is the signature pattern of AI Overviews traffic loss.
Not currently. The report is available in the Search Console interface with an export button, but the data is not exposed through the Search Analytics API. Automated AI Overviews traffic pipelines therefore work from manual CSV exports. Cognizo ships MCP and API access on every paid tier, so the cross-platform half of this reporting can be queried programmatically even while the Google half cannot. Note that tilde and dash values export as zeros, which corrupts any calculation treating zero as observed. Filter those rows out first.
The history is short, because the report is new. That constrains year-over-year AI Overviews traffic analysis for at least the first full year, which is why the baseline step matters so much. Export your pre-AI Performance report data at page and query level now, and store it outside Search Console, since the interface retains a limited window. Teams that skipped this in 2025 are reconstructing baselines from archived reports, which is not a pleasant process.
Search Console does offer an exclusion control for generative AI features, but the trade-off is severe and hard to reverse in practice. Excluding your content removes the citations that shape how AI describes your brand. It does not return those clicks to classic results either, since the overview still appears, built from other sources. You also lose the impression data that makes AI Overviews traffic measurable at all. For most sites the better response is changing what specific pages are for.
.svg)
.svg)
