The Search Console Generative AI Report Gives You Impressions and No Clicks

Google announced the generative AI performance reports on the Search Central blog in June 2026, and they have been quietly landing on properties ever since. The first reaction is always the same: where are the clicks?

The interesting answer is not "Google withheld them." It is that Google already counts them, in a different report, with nothing you can join on. That distinction changes what you should build, and it is the part every write-up of this launch has skipped.

What the report actually contains

First correction, because most coverage gets it wrong: this is two reports, not one. Search and Discover are separate, they carry different dimensions, and neither breaks the AI surfaces apart.

Two things follow immediately. Discover has no device dimension, so any cross-surface dashboard you build has a ragged edge. And within the Search report there is no surface breakdown — you cannot tell an AI Overviews impression from an AI Mode one, even though those are very different user journeys, as the AI Overviews vs AI Mode citation comparison lays out.

On the Search Labs exclusion, Google is explicit in the Performance report documentation: "Search Console doesn't include data from experiments in Search Labs, as these experiments are still in active development." So whatever you are seeing, it is production surfaces only.

The clicks are not missing. The join key is.

Here is the line that reframes the whole launch, and it is not in the new report's documentation — it is in the old Performance report help page, where it has been sitting all along:

Read that twice. Google does count AI Overview and AI Mode clicks. It has been counting them the entire time. They are in your Web search type total right now, blended in with ordinary blue-link clicks, indistinguishable from them.

So a generative-AI CTR is not a missing measurement. It is a numerator in one table and a denominator in another, with no key between them. The gen-AI report hands you impressions for the AI surfaces. The main Performance report holds the corresponding clicks, but only inside a total that also contains everything else. Same site, same dates, same pages — and still unjoinable, because the two reports are cut on different axes.

That is worth being precise about, because "Google didn't give us clicks" and "Google gave us clicks but not separably" imply completely different workarounds. The first says wait for a feature. The second says the split will never arrive from an arithmetic trick, no matter how you slice your existing exports.

Assembled as a table, the asymmetry is stark:

Exactly one row is green. Everything you would actually want to report on — did the AI answer cost me clicks, at what rate, on which queries — sits on a red row. If you have been trying to reconcile a rank tracker against Search Console and failing, the rank tracker vs Search Console mismatch breakdown covers the older half of that problem; this is a new instance of the same disease.

Generative AI impressions are a subset, not a new channel

This is the mistake I expect to see in the most dashboards, because the intuitive reading is exactly backwards. Google's help page states it plainly:

Generative AI impressions were already in your totals before the report existed. The report is a view over data you already had, not a new stream arriving alongside it. Which means:

• Never add gen-AI impressions to Web impressions. That double-counts every one of them.
• Never treat a rise in gen-AI impressions as incremental reach. It is a reclassification of visibility you already had.
• The only defensible calculation is a share, on the same URL scope and the same date window.

# WRONG — double counts
total_visibility = web_impressions + genai_impressions

# RIGHT — a share of an existing total
ai_share = genai_impressions / web_impressions      # same property
                                                     # same URL scope
                                                     # same date range

# Illustrative, NOT measured:
#   web_impressions   = 120_000
#   genai_impressions =  18_000
#   ai_share          =  0.15   → 15% of Web impressions came via an AI surface

Those figures are illustrative. I have not published measured impression volumes here, and you should be suspicious of anyone who published them in the first week — see the honest limits section for why.

If you are folding this into a reporting pipeline, the share is the number that belongs on the dashboard, and it belongs next to the traffic-loss framing in the great decoupling measurement guide rather than in its own isolated tile. Getting that plumbing right is ordinary work — the patterns in automating SEO reports and the SERP API SEO dashboard both apply unchanged.

Three traps in the one number you can compute

You get exactly one honest metric out of this report: AI impression share. It has three ways to mislead you, and none of them is documented in one place.

1. The date-window trap

The gen-AI reports have a much shorter history than your Web data. If you compare a 60-day gen-AI window against a 90-day Web window, the share is mechanically understated by roughly a third — not because your AI visibility fell, but because you divided a short numerator by a long denominator. Pin both windows to the same start and end date, and start that window no earlier than the day gen-AI data actually appears for your property.

2. The chart total does not equal the sum of the page rows

This one is buried in a single sentence of Google's documentation and I have not seen it written up anywhere:

Page-level rows count separately; the chart total deduplicates same-site appearances within one AI feature. So your page rows will sum to more than the chart total, and the gap is not a bug — it is the number of times an AI answer cited you twice at once.

The practical consequence is that "AI impression share" has two legitimate denominators and they give different answers. Pick one, write it down, and never mix them across reports. If you want the property-level story, use the chart total. If you want per-URL attribution, use the page rows and accept that they over-count multi-citation events. Quietly switching between them is how a dashboard starts lying.

3. Impressions require the link to be seen

Impressions in AI features are not "the answer rendered." For AI Overviews, Google's rule is that "the link must be scrolled or expanded into view." In Discover, "only one impression is counted per result per session; if a user scrolls past a card and then scrolls back, only one impression is recorded."

So a citation that sits behind a collapsed AI Overview and never gets expanded is not an impression. Your gen-AI impression count is a measure of seen citations, not of earned citations. Those are different quantities, and only the second one tells you whether your content is winning — which is precisely what an independent AI Overview exposure audit is for.

What AI Overviews already did to average position

A corollary nobody seems to have drawn. Google documents how position works for these surfaces:

Every link inside one AI Overview inherits one shared slot. That is a structurally different object from an ordinary organic rank, where position 3 means one result in one place. And because AI Overview impressions live inside your Web search type totals, your average position in the main Performance report is already an average over two different kinds of position.

Now add the new report's shape: it has no position dimension at all. So you cannot subtract the AI-surface contribution back out, and average position becomes a metric you can watch but not decompose. Anyone still treating it as a clean ranking signal is measuring a blend and calling it a rank — the same category error the num=100 removal forced on rank trackers last year, arriving from a different direction. If you need position to mean one thing, it has to come from observing the SERP, not from a report that pools surfaces. The overlap between AI Overview citations and classic rankings is a useful sanity check on how far apart the two have drifted.

No API value. No BigQuery export.

I checked both programmatic surfaces at source, because "it's not in the API" is asserted constantly and rarely verified. Both are closed, and both close the same way.

The Search Console API's searchanalytics.query method accepts exactly six values for type:

web · image · video · news · googleNews · discover

The bulk data export to BigQuery exposes exactly the same six in its search_type field:

web · image · video · news · googleNews · discover

There is no aiOverview, no aiMode, no generativeAi in either list. The report is a UI surface only. Every automated pipeline you own — scheduled exports, warehouse loads, client dashboards — is blind to this data until Google adds a type value, and there is no announced plan that it will.

That also means the practical answer to "how do I trend this?" is manual export, dated and stored yourself, on a cadence you own. Unglamorous, but it is the only thing that works today.

What impressions alone can and cannot prove

Impressions are a real signal. They are just a narrow one, and it is worth being blunt about where the line falls before someone builds a strategy on the wrong side of it.

The last two rows are the ones that hurt, and they are structural: Search Console is a first-party tool. It will never tell you what the AI answer said, which competitors it cited, or which phrasing triggered it. That has always been true; the AI surfaces just make it expensive. The framing in AI search visibility metrics and the strategy split in GEO vs AEO vs SEO both assume you have a way to see the answer itself, and that assumption is doing a lot of work.

Closing the query-side gap

There is exactly one way to recover the query dimension: observe the AI answer directly for the queries you care about, and hold your own history of it. That is a SERP-observation job, not a Search Console job, and the two are complements — Search Console tells you that you were seen; observing the SERP tells you for what, alongside whom, and with what wording.

Serpent API's /api/search returns the AI Overview block alongside organic results, so you can track the citation set for a keyword list on your own schedule. We have written the mechanics up in extracting Google AI Overviews via API, the vendor landscape in which SERP APIs return AI Overviews, and the AI Mode side in tracking AI Mode citations in Python. If you want the model-side view as well — what several assistants say about a topic, not just what Google shows — that is what AI Rank is for.

Pricing is flat per call: $0.60 per 1,000 web searches on the default tier, $0.06 per 1,000 on Growth and $0.03 per 1,000 on Scale. Full details on the pricing page.

Honest caveat, because it matters for how you read your numbers: SERP observation and Search Console will never agree exactly. They sample different things — one is a first-party log of real user sessions, the other is a point-in-time look at a result page — and the query fan-out behaviour described in AI Mode query fan-out means a single user prompt can expand into several underlying searches you never typed. Use each for what it is good at. Search Console owns "did real users see me"; SERP observation owns "what did the answer look like, and who else was in it." Understanding how AI search selects citations is what turns the second into something you can act on.

Honest limits of this post

Three things I am not claiming.

No impression volumes. I have deliberately published no measured gen-AI impression numbers. The reports are days old on most properties, and a single-property figure from week one would be an anecdote dressed as a benchmark. Any "average AI impression share" you see quoted right now is built on the same thin base — treat it accordingly.

No documented start date or retention. Google's help pages for both reports state no data start date, no backfill policy and no retention window. Practitioners have reported data beginning in mid-2026, but that is observation, not documentation. Do not assume the standard 16-month retention applies here until Google says so.

No rollout announcement. Google announced the reports in June 2026. It has not announced that the rollout finished; properties simply started showing the reports. If you do not have them yet, that is not evidence of anything about your site.

Everything else in this post is quoted from Google's own documentation, checked on 13 August 2026, and linked in the references below so you can verify each line yourself.

FAQ

Why does the Search Console generative AI report show no clicks?

Google designed the report as an impressions-only view. The clicks are not unmeasured — Google's Performance report documentation states that clicking an external link in an AI Overview or in AI Mode counts as a click. Those clicks are recorded inside your Web search type total, blended with ordinary organic clicks, and the generative AI report offers no dimension you can use to separate them out.

Are AI Overview clicks counted in Search Console at all?

Yes. Google's Performance report help page says explicitly that clicking a link to an external page in an AI Overview counts as a click, and the same rule applies to AI Mode. Those clicks have been flowing into your Web search type totals all along. What you cannot do is tell which of your clicks came from an AI surface, because no report splits them apart.

Should I add generative AI impressions to my total Search Console impressions?

No, and this is the most common mistake. Google's documentation states that the generative AI performance report includes data from the Web search type in the Performance report. Generative AI impressions are a subset of impressions you already had, not a new channel arriving alongside them. Adding the two double-counts every generative AI impression. Calculate a share instead, using the same URL scope and the same date window on both sides.

Can I get generative AI report data from the Search Console API?

No. The Search Console API's searchanalytics.query method accepts six values for the type parameter: web, image, video, news, googleNews and discover. The BigQuery bulk data export exposes the same six values in its search_type field. Neither list contains a value for AI Overviews, AI Mode or generative AI, so the report is a UI surface only and automated pipelines cannot read it today.

Why do the page rows not add up to the chart total?

Because the chart deduplicates and the table does not. Google documents that if two results from the same site appeared in one generative AI feature, they count as a single impression in the chart total, while page-level rows count each result separately. The sum of your page rows will therefore exceed the chart total, and the gap is the number of times an AI answer cited your site more than once at the same moment.

Does the generative AI report cover AI Mode as well as AI Overviews?

Yes, but combined. The generative AI performance report for Search covers both AI Overviews and AI Mode, and there is no dimension that separates one from the other, so you cannot tell which surface produced an impression. The report also excludes Search Labs experiments. Discover has its own separate generative AI report, which carries Pages, Countries and Dates but no device dimension.