← blog / SEO & GEO

How to Read Search Console's AI Performance Report

·updated ·18 min readseogoogle-search-consolegeomeasurementcontent-strategy
Abstract Magic Share brand graphic: a field of dots with a highlighted cluster in green and teal on a near-black grid, with the Magic Share name and the label "seo".

Search Console's generative AI performance report answers a narrower question than the report you already check: not did anyone come from this post, but was this post shown as a source inside an AI answer. It reports impressions only, with no clicks, click-through rate (CTR), position or queries. Those impressions are also a slice of numbers you already had, because Google states plainly that "the generative AI performance report includes data from the Web search type in the Performance report (Search results)". The reports launched on 3 June 2026 for a subset of sites, and Google says they reached all websites worldwide on 31 August 2026. Data reaches back only to 18 May 2026, so on a twenty-post blog you are reading a short, thin series that needs its own baseline and its own rules before it means anything.

That is the whole trap in one sentence. Your AI impressions were never missing from Search Console. They were sitting inside your Web impressions the entire time, and in June 2026 Google finally drew a box around them.

In short:

  • AI impressions are a subset of your Web impressions, so never add the two. Track AI share and post-level coverage instead.
  • The series starts on 18 May 2026 and already has known logging gaps, so annotate the anomalies and export a CSV every month.
  • On twenty posts, the decision this report supports is which posts to refresh, not how to tune them.

The report answers a different question than the clicks report

The classic performance report is a distribution question: how much traffic arrived, from which queries, at what position. The generative AI report is an inclusion question: when Google assembled an AI Overview or an AI Mode answer, did a link to your page get shown to the person reading it?

Two counting rules make this concrete. First, Google defines the metric as "how many times links to your site were shown to a user in a generative AI feature on Google Search." Second, Google's counting rules for AI Overviews and AI Mode say "Standard impression rules apply." In general that means an impression is counted when the link appears in the current page of results, whether or not the user scrolled to it, as long as they did not have to click to reveal it. So an AI impression does not mean Google used my page to build an answer somewhere in its pipeline. It means your link sat in an answer a real person loaded. For a founder deciding whether a post is doing anything, that is a more useful signal than a raw citation count would be.

Now the arithmetic rule that everything else depends on: never add AI impressions to Web impressions. They are the same impressions, viewed twice. Google says the report draws from the Web search type, and John Mueller confirmed in January 2026 that when the same URL appears in an AI Overview and as a blue link for one query, Search Console counts one impression rather than two. The performance report also aggregates by property in the chart view — two results from your site on one query count as a single impression.

So "AI impressions up 30%, Web clicks down 10%" is not two independent signals to reconcile. It is one population being sliced. The only legitimate derived number is a share: AI impressions divided by Web impressions, for the identical date range and the identical page filter. Anything else is double counting.

One more aggregation quirk worth knowing before you get confused by your own spreadsheet: the chart is aggregated at property level, but adding a URL filter switches it to URL-level aggregation. Your chart total and the sum of your table rows will not match, and Google's help page says to expect exactly that.

Find the report, and understand why yours might be empty

The report sits beneath the Search results report in the Performance section. Brodie Clark documented a faster route: append /ai to your existing performance report URL, for both the Search and Discover tabs. It covers AI Overviews and AI Mode; it explicitly excludes Search Labs experiments, and Discover generative AI data lives in a separate report.

If you open it and see nothing, Google's help page gives two explanations, and only one of them is about your content.

Access is no longer one of them. The reports reached a subset of UK sites first on 3 June 2026, then expanded around 23 June 2026 to sites in the US, India and Switzerland. Google's launch post now carries the note: "As of August 31, 2026, we've rolled out these insights to all websites worldwide."

You may be under the volume gate. The help page puts it first: "This usually happens when your site hasn't received enough impressions in generative AI features on Google Search." There is a threshold, Google does not publish the number, and small sites are exactly the population most likely to hit it. An empty report at twenty posts is not evidence that you have been excluded from AI answers. It is, at most, evidence that you are below a line nobody has told you the location of.

You may have opted out, possibly without knowing. The help page's second explanation is that you may have excluded your site from Search generative AI features. Search Console has a Search generative AI control under Settings. For the first UK properties, Clark reported it took effect from 17 June 2026, and Google's control help page says it reached all websites worldwide on 31 August 2026. The default is Include. The trap is inheritance: an exclusion carried down from a parent property is enough to empty the report. Check the setting before you conclude anything about your content.

If your report is still empty, there is no workaround: the data is not in the API, so there is nothing to script. Do the two useful things instead. Confirm your generative AI control is set to Include, and start a clean clicks baseline in the classic report now, because that is the series you will pair the AI data against once your site clears the threshold.

What's in it, and what isn't

The full inventory is short. Four dimensions — Pages (grouped by final URL after redirects, canonical preferred), Countries, Devices, and Dates in Pacific Time — and one metric, impressions. A search-type filter also splits Web searches into text-based and multimodal, the latter covering image-led searches such as Google Lens and Circle to Search.

What is absent matters more:

  • No clicks, no CTR, no position, no queries. TDMP's field write-up puts the consequence bluntly: it is "impossible, based on the GSC reports alone, to say whether your influence in Google's AI search features is leading to website visits and driving revenue."
  • No non-Google surfaces. ChatGPT, Claude and Perplexity appear nowhere in this report.
  • No API. The Search Analytics API's type field still accepts only web, image, video, news, discover and googleNews. There is no generative AI value, no BigQuery export, no automated pull. Your options are the dashboard and the export button, which covers both chart and table data.
  • No long history. Data starts 18 May 2026. Year-over-year comparison is not possible until May 2027.

With queries absent from the report, a practical workaround is a query fan‑out keyword research workflow that surfaces likely search prompts for your pages; see query fan‑out keyword research for a step‑by‑step guide.

On the missing metrics, Google's launch post says: "We're continuing to work with website owners to understand what insights and data would be most helpful to inform their strategies, such as adding additional metrics over time." No timeline is attached, and as of October 2026 impressions are still the only metric.

On granularity, the launch post lists hourly, daily, weekly and monthly views. The help page lists days, weeks and months and defers to the main performance report, where hourly data appears only in the 24-hour view. For a monthly ritual the difference does not matter.

There is a small consolation in being small here. The two limits that cripple large-site Search Console analysis barely touch you. The 1,000-row limit is irrelevant when you have twenty pages. And query anonymization cannot bite a report that has no query dimension at all. In Ahrefs' study of 22 billion clicks across 887,534 properties, anonymized queries made up 46.77% of traffic in April 2025, and the most common per-site rate fell between 45% and 80%. You can read every row of your site's entire AI visibility on one screen. What you lose instead is the volume gate, and the statistical fragility of small counts.

Set a baseline before you interpret anything

A ten-week series with known holes in it will happily tell you a story that isn't there. Before you read a single trend, write down what the holes are.

Start date: 18 May 2026. Everything before that is blank, not zero.

The eleven-month impression bug. Google's data anomalies page records, verbatim: "A logging error prevented Search Console from accurately reporting impressions from May 13, 2025 until April 27, 2026." The same entry says impressions, CTR and average position were affected but clicks were not, that it affected data logging only, and that you may notice a decrease in impressions now it is resolved. Practical consequence: if you compare AI-era impressions to your 2025 impression baseline, you are comparing against inflated numbers. Clicks are the cleaner historical series. Use them for anything that crosses that boundary.

24 June 2026. The anomalies page documents a logging error that decreased clicks and impressions in the Discover performance report that day and, for properties with access to generative AI features in Discover, also decreased reported impressions. The same thing happened again on 13 August 2026.

13–17 August 2026. A logging error decreased impressions in the generative AI report for Search itself. Google updated the entry on 21 August to say the missing data had been restored. If you exported a CSV covering that week before then, export it again. The generative AI reports already contain several known bad days inside their short history.

7 May 2026. FAQ rich results stopped appearing in Google Search, causing a permanent impression drop for anyone who had them. Discover logging errors also hit 7–8 May and 21 May 2026.

Country launches reset to zero. AI Overviews and AI Mode went live in France on 22 July 2026. French impressions begin from zero on that date. Read the Countries dimension as a set of separate start lines, not a single growth curve.

The newest data is preliminary. Search Console shows it with a dotted line and warns it is "still being collected and might change in the next few hours." Do not react to the right-hand edge of the chart.

Then build the archive you actually control. Because there is no API and Google demonstrably restates history, the only durable record is a manual export. Once a month, export the table to CSV, drop it in a dated folder, and keep a plain-text annotations file next to it: rollout dates, anomalies, the day you shipped a redesign, the day you changed a title. It takes two minutes and it cannot be reconstructed later.

That is the same argument this site makes about publishing itself: a blog compounds when it is tended on a schedule rather than in bursts. If your drafting cadence already runs on a timer — Magic Share's agent wakes at 06:00 UTC and drafts what is due, about one post a week on the paid plans — putting the measurement ritual on the same calendar costs you nothing extra and gives every future comparison a clean edge to sit against. Related reading: how often a small blog should actually publish.

The four numbers a twenty-post blog should track

At your scale, most of the dashboard's charting is decoration. Four numbers do the work.

1. Coverage — how many posts appeared at all

With no queries and no clicks, the metric that is both available and stable at small volumes is binary per post: did this URL record at least one AI impression this month? On twenty posts, "13 of 20 posts had at least one AI impression, up from 9 last month" is a genuinely readable signal. It survives noise, it maps to a decision (which posts to look at), and it does not pretend to precision it lacks.

2. AI share of Web impressions

The one legitimate ratio: AI impressions ÷ Web impressions, same date range, same page filter. The best-known public reference point is Clark's worked example on a large UK site in early June 2026: 813,000 AI impressions against 8.4 million total Search impressions over seven days, roughly 10%. That sat on a cumulative base of 1.79 million AI impressions since 18 May, about 120,000 a day. Clark cautions that share is "likely on the lower end" for that site because branded traffic dominates it. Treat it as one data point from one large site, not a benchmark you are failing to hit. Your share will differ for reasons that have nothing to do with quality.

3. The per-post pairing

For each post, put its AI impressions next to its Web clicks from the classic report, same dates. This is the only way to reconstruct the missing half of the picture, and it produces the four states in the next section.

4. The noise floor

This one is arithmetic, not a finding. Impression counts are arrival counts, and arrival counts wobble by roughly their own square root. At about 40 impressions a month, a swing of ±6 is ordinary. At about 400, ±20 is ordinary. So the rule for a small blog is: ignore any month-over-month move smaller than roughly the square root of the smaller number, and compare three consecutive 28-day windows rather than two. Two points make a line whether or not anything happened. Three points, on a report whose newest data is explicitly preliminary, is the cheapest protection against rewriting your content strategy over rounding.

Reading the four states

Pair each post's AI impressions with its classic Web clicks and you get four cells, each with a different action.

Web clicks present Web clicks flat or near zero
AI impressions present Working. Leave it alone; note what it does structurally and reuse the pattern. The interesting case — cited, not clicked.
No AI impressions Ranking conventionally but not entering answers. Check structure and freshness. Not competing yet. Decide whether the topic is worth another pass at all.

Cited but not clicked is the state the classic report alone would have shown as failure, and the evidence says it is now common. Ahrefs found the correlation between daily impressions and daily clicks on its own property moved from +0.425 in the second half of 2024 to −0.352 in the first half of 2025; across a 300,000-keyword sample, AI Overviews were associated with a 34.5% reduction in click-through rate. Ryan Law's summary: "Clicks and impressions have become decoupled from one another — one can grow while the other declines."

The strongest causal evidence points the same direction. Saharsh Agarwal (Indian School of Business) and Ananya Sen (Carnegie Mellon) ran a pre-registered randomized field experiment with 1,065 US desktop Chrome users recruited via Prolific, two weeks each between January and February 2026. AI Overviews reduced organic clicks by 38% on queries where they appeared, and zero-click searches rose from 54% to 72%. AI Overviews showed on 42% of queries in that sample. Scope matters: US, desktop Chrome, two weeks per participant, and the paper is an unreviewed draft.

Being in the citation set is still worth something. Seer Interactive reported in September 2025 that brands cited in AI Overviews earned 35% more organic clicks and 91% more paid clicks than non-cited competitors — a figure worth quoting carefully, since the secondary write-up does not disclose the sample size. The direction is what matters for the four-state read: a cited-but-not-clicked post is doing a job your clicks report was never built to show. We worked through what that job is worth in why a blog still earns its keep when most searches end without a click.

What to change when a post never appears

Suppose coverage says five of your twenty posts have recorded no AI impressions in three consecutive windows. Google's own AI optimization guide is unusually direct about this, and the most useful half is the list of things that do not help:

  • llms.txt files. Google Search does not use them.
  • Deliberately chunking content into small pieces for AI comprehension.
  • Rewriting content specifically for generative AI search. The systems handle synonyms and general meaning without special phrasing.
  • Chasing inauthentic mentions.
  • Over-relying on structured data, which the guide states is not required for generative AI search.

What the guide does recommend is unglamorous: content with a genuine first-hand perspective rather than commodity coverage, technical soundness so pages can be crawled and indexed, semantic HTML, good page experience, and less duplication. Its guiding line is "Focus on what your visitors would enjoy, find helpful, and feel satisfied with after visiting your website." Our own breakdown of the on-page patterns that get small blogs cited by ChatGPT and AI Overviews covers the paragraph-level craft that sits underneath that advice.

Then there is freshness, which is the most actionable finding for anyone with exactly twenty posts. Seer Interactive studied 7,683 pages carrying 47,097 citations between March and June 2026 across four brands and three engines, and found 75% of cited pages had been updated within the last year and 88% within two. Among the 4,124 pages where both dates were readable, 72% looked fresh by last-update date but only 42% had actually been published within the last year. Their conclusion: "Publish and forget loses. Publish and maintain wins."

One caveat: that study covers ChatGPT, Gemini and Perplexity, not Google's AI Overviews and AI Mode, so it is a directional signal rather than a measurement of the surface this report tracks. But the implication for a twenty-post blog is hard to ignore: updating the five posts that already appear may beat publishing five more. Coverage data tells you which five. A topic ledger that records everything already written or planned is what stops the next draft from becoming a near-duplicate of the post you should have refreshed instead.

Last thing: resist attributing an AI-drafted post's performance to the fact that it was AI-drafted. Nothing in this report distinguishes drafting method. It tells you whether this page was used as a source, which is a question about the page's usefulness, structure and freshness — exactly what Google's guide says matters.

What the report can never tell you

It is worth stating the objections plainly, because they are good ones.

It is not actionable for optimisation. Clark, one of the most careful readers of Search Console reporting, calls the current data "nowhere near actionable enough to be used in a practical sense," missing queries and clicks — "effectively the most important metrics." Drew Garrett's version is sharper: an impression you cannot tie to a click or a query "is a scoreboard with the score taped over". The fair rebuttal is a narrow one: it is not actionable for tuning, but it is actionable for one decision — whether a post entered the answer set at all. That is worth ten minutes a month and no more.

The tool has been unstable. TDMP reported the report crashing when they adjusted a date range, then showing impressions that had not been there before the crash, and recommends taking the reports "with a grain of salt" for now. Google's anomalies page has since logged a five-day gap in the Search report itself, later restored. If it breaks on you, it is not just you.

The evidence about harm is contested. Google's position, from Liz Reid on 6 August 2025: "total organic click volume from Google Search to websites has been relatively stable year-over-year", and average click quality has increased. No supporting dataset was published. The Agarwal and Sen experiment measured a 38% drop and concluded AI Overviews "divert traffic away from publishers without delivering measurable improvements in user experience" — with user satisfaction ratings nearly identical between groups. Google's dataset is larger; the experiment's method is stronger. Both are true at once.

AI traffic volume is small — but it converts. Ahrefs' own analytics put AI search at roughly 0.5% of visits but about 12.1% of sign-ups, with ChatGPT at 0.24% of total traffic share; Buffer reported LLM-driven traffic converting at 20.15% against 7.06% for organic search. These are individual SaaS companies' own numbers rather than controlled studies, and the same article notes e-commerce sites often see the opposite.

Should a small blog opt out? Google's wording is unambiguous: opting out means "you won't receive any traffic or impressions from these features", though "this control isn't used as a ranking or inclusion signal affecting other parts of Search." It propagates in a day or two, and it does not stop your content being used for model training — Google points to Google-Extended for that. The control exists because the UK's Competition and Markets Authority required it, and roughly a third of SEO professionals said in early research that they would block; Clark doubts even 1% of sites actually will. For a small SaaS blog seeking distribution rather than ad revenue, the citation-lift evidence argues for staying in. It is a real decision — but Include is the sensible default.

The ten-minute monthly checklist

  1. Open Performance → append /ai. Confirm the generative AI control is still set to Include.
  2. Export the table to CSV. Save it in a dated folder. Add a line to your annotations file.
  3. Count coverage: how many of your twenty posts recorded at least one AI impression.
  4. Compute AI share: AI impressions ÷ Web impressions, same range, same filter.
  5. Pair the five posts you care about most with their Web clicks from the classic report.
  6. Ignore any change smaller than the square root of the smaller number. Wait for the third window.
  7. Check Google's data anomalies page before you explain any dip with a story about your content.
  8. Pick one post from the no-impressions column and refresh it rather than writing a new one.

FAQ

Why is my Search Console AI performance report empty?

Google's help page gives two reasons: your site is below an unpublished impression threshold, or it is excluded from Search generative AI features, possibly through a setting inherited from a parent property. Access is no longer a cause, since the report reached all websites worldwide on 31 August 2026.

Do AI impressions add to my Web impressions?

No. The generative AI report draws on the Web search type, so these impressions were already counted, and the same URL in an AI Overview and a blue link counts once. Track AI impressions as a share of Web impressions instead.

Can I pull AI performance data through the Search Console API?

Not as of October 2026: the Search Analytics API has no generative AI search type and there is no BigQuery export. Use the report's export button and keep a monthly CSV in a dated folder.

How many AI impressions should a twenty-post blog expect?

There is no published benchmark, and one large UK site's roughly 10% AI share is not a target for a small blog. Track coverage and your own share over three consecutive 28-day windows instead.

Does the report include ChatGPT, Claude or Perplexity?

No. It covers AI Overviews and AI Mode in Google Search (Discover has its own report), so other AI assistants have to be tracked separately.

The short version

The generative AI report will not tell you whether AI search is sending you customers. It tells you whether your pages are entering the answers — and on a twenty-post blog that single binary, tracked as coverage over three settled windows against an annotated baseline, is enough to decide what to refresh next month. Export the CSV, write down the anomalies, and check it once a month with the classic clicks report open beside it.

For a small site, structuring content into topic clusters and tightening internal links can make this coverage signal more actionable; see topic clusters and internal linking for a 30-post blog for a practical, step‑by‑step plan.

The measuring is ten minutes. The publishing is the part that eats afternoons. If you would rather the research, drafting and link-checking ran on a schedule while you keep the approve-or-reject decision, Magic Share drafts posts with every cited URL verified and holds them until you say yes — your first three posts are free.

More on SEO & GEO →

by email

Get new posts in your inbox

A short digest when new posts go up. Confirm once, unsubscribe from any email.

We use your address only for this digest. Privacy

Keep reading

  1. ·16 min read

    How to Write a Fair SaaS Competitor Comparison Page

    How to write a competitor comparison page for SaaS that names rivals fairly, sources every claim, and converts — with the evidence and update cadence.

    content-strategyseosaas-marketingcomparison-pagesgeo

  2. ·16 min read

    Query Fan-Out Keyword Research: One Search, Ten Queries

    Query fan-out keyword research: AI Mode fires about ten searches per prompt and 95% have zero volume. How to map sub-questions one post can answer.

    seogeoai-modekeyword-researchcontent-strategy

  3. ·21 min read

    Topic Clusters for a 30-Post Blog: Internal Linking

    How to shape a 30-post blog into topic clusters with bidirectional internal links, and why anchor variety beats page count for earning AI citations.

    seointernal-linkingcontent-strategycontent-planninggeo

Want posts like this for your site?

Magic Share researches, writes and fact-checks posts like this for any site — point it at your URL and review your first draft today.

Plant your first post