TL;DR
On July 9, 2026, Google Search Console's Search Generative AI performance reports expanded to more sites. They cover AI Overviews, AI Mode, and AI Overviews in Discover. And they expose exactly five things: impressions, pages, countries, devices, and dates.
No clicks. No CTR. No query data.
So you can now see that you appeared in an AI surface. You cannot see what was asked, whether anyone clicked, or what the AI actually said about you. Google is not going to hand you that last one — ever. Which means this is the moment SEO teams have to go collect their own ground truth, because the platform's own report is deliberately partial.
ScrapeMaster is one honest way to do that: load an AI answer surface yourself, let auto-detect find the structure in 2–4 seconds, extract it into dated rows, and build a longitudinal dataset you own. It's a personal-scale sample, not a panel study, and we'll be specific about that difference below.
What the report gives you, and what it withholds
Let's be precise, because the precision is the point.
| Dimension | In the AI performance reports? |
|---|---|
| Impressions | Yes |
| Pages | Yes |
| Countries | Yes |
| Devices | Yes |
| Dates | Yes |
| Clicks | No |
| CTR | No |
| Query data | No |
| What the AI said about you | No |
| Whether you were cited or merely crawled | No |
| Competitors cited alongside you | No |
The surfaces covered are AI Overviews, AI Mode, and AI Overviews in Discover. The expansion on July 9 widened the set of sites that see the report — it didn't widen what the report contains.
Look at the shape of that table. Everything above the line is something Google can tell you without revealing anything about user behavior or about the AI's output. Everything below the line is behavioral or generative. That's not a coincidence and it's probably not a temporary state of the roadmap. It's a boundary.
Why "no clicks" is the whole story
An impression count without a click count is a number you cannot act on.
Suppose your impressions in AI Overviews went up 40% month over month. Good news? You have no idea. It could mean:
- You're being cited more, and traffic is coming.
- You're being cited more, and the citation answers the question so completely that nobody clicks.
- AI Overviews triggered on more queries in your space, and you were swept along.
- Google expanded the surface's coverage and everyone's impressions rose.
- You're being cited more, in a context that makes you look bad.
Five explanations, one number, and no way to distinguish them from inside the report. The impression metric tells you the surface touched you. It does not tell you whether that was good.
On CTR numbers you'll see quoted
You will see confident AI Overviews CTR statistics circulating. Percentages, decline figures, precise before-and-afters.
We're not going to repeat any of them, because we couldn't verify them to a primary source — and more importantly, the entire reason they're unverifiable is the subject of this post. Google doesn't publish AI Overviews click data. It isn't in Search Console. So where would a precise industry-wide CTR figure come from? Clickstream panels of varying quality, small samples extrapolated hard, or vendor datasets nobody can audit.
Nobody actually knows. Anyone selling you a precise AI Overviews CTR figure is guessing, or measuring something narrower than they're implying. That's not cynicism — it's just what follows from the data not being public. Treat any such number as a hypothesis, not a fact.
Which is exactly why your own data, however small, is worth more than someone else's confident chart. We wrote more about this dynamic in our take on AI Overviews and SEO content.
Collect your own ground truth
Here's the honest reframe: you're not going to reconstruct Google's click data. You can't. What you can build is a longitudinal record of what the AI surfaces actually say — which Google will never give you at any price, and which is arguably the more valuable half.
Four things worth extracting on a schedule:
- AI answer surfaces you've loaded yourself. Run your target queries, look at the answer, extract the cited sources and the claims.
- Competitor content that keeps getting cited. If the same three domains show up in every AI answer in your category, that's a content brief writing itself.
- Your own ranking pages. Track which of your URLs are the ones being surfaced.
- SERP snapshots over time. The layout changes. Capture what the page looked like on a date.
The method, concretely
This is the part most "just scrape it" advice skips.
Step 1 — Fix your query set. Pick 20–50 queries you actually care about and freeze the list. A moving query set makes your time series meaningless. Write them down.
Step 2 — Sample on a schedule. Same queries, same day of week, same rough time, same location. Weekly is plenty. Consistency matters far more than frequency — a sloppy daily sample is worse than a disciplined monthly one.
Step 3 — Load the page yourself and open the side panel. ScrapeMaster opens in a Chrome side panel and reads the page you're already looking at. Auto-detect finds repeating patterns in 2–4 seconds and names the columns. No CSS selectors, no code. It reads the post-render DOM, so React/Vue/Angular surfaces are fine.
Step 4 — Rename your columns to something your future self understands. Auto-detect gives you sensible names; you should still change them. source_domain, claim_text, position beat whatever the DOM called them.
Step 5 — Keep the date column. This is the whole thing. Every row needs the date it was collected. A dataset without a date column isn't a time series, it's a snapshot you'll misread in three months. If there's one instruction to take from this post, it's this one.
Step 6 — Go one level deeper with Follow detail. Follow detail opens each cited link in a background tab, pulls extra fields, and merges them back into the row. Use it to grab the title, publish date, or author of every source the AI cited. Now you can ask "does the AI prefer recent content in my category?" and answer it with data.
Step 7 — Handle pagination if you're doing SERPs. Next-page buttons, load-more, numbered pagination, and infinite scroll are all handled. Set an extraction delay. Be polite.
Step 8 — Export and accumulate. CSV, XLSX, JSON, or straight to the clipboard for Google Sheets. Data lives locally in IndexedDB until you export. Append to one master sheet, don't overwrite. Our guide on exporting website data to Google Sheets covers the clipboard path in detail.
Step 9 — Triangulate against GSC. Now you have two datasets: Google's impressions (real, authoritative, incomplete) and your citation record (partial, yours, complete on the dimension Google withholds). Line them up by date and page.
What triangulation actually buys you
| Your data says | GSC says | Reasonable read |
|---|---|---|
| Cited often | Impressions up | Consistent. You're in the surface. |
| Cited often | Impressions flat | Your sample may not match real query distribution |
| Rarely cited | Impressions up | The surface is triggering; you're being seen, not quoted |
| Cited, unflatteringly | Impressions up | Visibility isn't the win here. Fix the source content. |
| Not cited | Impressions zero | Clear. You're absent. Start there. |
None of these are proof. All of them are better than staring at an impression count alone and guessing. And the fourth row is one you can only reach by reading the actual AI output — GSC will never surface it, because "impressions" doesn't have a sentiment.
Being honest about what this method is not
If we sold you this as a measurement solution, we'd be doing the same thing as the people quoting unverifiable CTR figures.
A personal-scale sample is not a representative panel study. You're sampling your queries, from your machine, in your location. That's a convenience sample with an n in the dozens. It is not a random sample of the query distribution and it cannot be extrapolated to "AI Overviews cite X% of the time" for anyone but you.
Results are personalized and vary by location. AI surfaces personalize. Your logged-in session, your location, your history, and the specific data centre you hit all move the output. Two people running the same query at the same second can get different answers. Any conclusion you draw is conditioned on all of that.
AI outputs are non-deterministic. Run the same query twice, get different phrasing and sometimes different sources. If you see a change between two samples, that might be a trend or it might be a coin flip.
You should say all of this in your own reporting. When you present this to a client or a CMO, label it: "convenience sample, n=40 queries, single location, weekly, non-deterministic surface." That framing makes your work more credible, not less. The people who present a scraped sample as a market study are the reason nobody trusts SEO data.
What you can honestly say: "Across 40 tracked queries in our category, over 12 weeks, competitor X was cited in 60% of AI answers and we were cited in 15%. Here's the raw data. Here's the date column." That's a real, defensible, useful finding. It just isn't a population statistic.
What we send, and what we don't
Because it's relevant when you're pointing a tool at your own search data:
- Extracted data never leaves your browser. It's stored locally in IndexedDB and exports to CSV, XLSX, JSON, or the clipboard. We never see your rows.
- The one network request is during auto-detect. The page's HTML structure — not its content — goes to our analysis API so it can suggest columns and selectors. Structure, not text.
- No account. No sign-up, no email, no trial. Free is a feature, not a funnel.
That's the full disclosure. If it mattered less to us we'd have buried it in a privacy policy instead of a blog post.
Where this sits against other tools
| Tool | Runs where | Setup | Sends your data where |
|---|---|---|---|
| ScrapeMaster | Your browser, side panel | Auto-detect, 2–4s | Nowhere — local IndexedDB (structure-only to analysis API) |
| Octoparse | Their cloud | Workflow builder | Their servers |
| ParseHub | Their cloud | Project setup | Their servers |
| Import.io | Their cloud | Enterprise onboarding | Their servers |
| Instant Data Scraper | Your browser | Auto-detect | Local |
| Simplescraper | Your browser + cloud | Recipe builder | Local or their cloud, depending |
| Web Scraper.io | Your browser + cloud | Sitemap definition | Local or their cloud |
| Thunderbit | Your browser | AI field suggestion | Depends on the AI step |
For sampling AI surfaces specifically, browser-side matters more than usual. These are personalized, session-dependent pages — a cloud crawler fetching them from a datacenter isn't seeing what you see, which defeats the purpose. You want the answer that was rendered to a real session, which means the extraction has to happen where the session is.
Worth noting honestly: ScrapeMaster can't bypass logins, paywalls, or CAPTCHAs, doesn't rotate proxies or fingerprints, and heavy extraction on aggressive anti-bot sites can get you blocked. Use extraction delays. If a site blocks you, it blocked you.
Keep the receipts
One more habit worth building. Extracted rows capture the data; they don't capture what the page looked like. For anything you might need to defend later — a client dispute, a "we were cited and then we weren't" claim, a competitor's misleading citation — take a dated PDF snapshot alongside the extraction. Convert: Web to PDF does it locally, with selectable text and working links, no watermark. We've written specifically about saving SERP snapshots and saving AI Mode and Overviews answers as PDF.
Rows for analysis, PDFs for evidence. They're different jobs.
Frequently asked questions
Will Google add clicks to the AI performance reports later?
Unknown, and we won't speculate. What's true today: the reports expose impressions, pages, countries, devices, and dates. No clicks, no CTR, no queries. Plan around what exists.
Why doesn't Search Console show AI Overviews query data?
Google hasn't given a public reason we can cite, so we won't invent one. The observable fact is that the reports cover AI Overviews, AI Mode, and AI Overviews in Discover, and query-level data isn't among the dimensions.
Can I use ScrapeMaster to reconstruct my AI Overviews CTR?
No, and be suspicious of anyone claiming otherwise. CTR needs click data from Google's side. You can build a record of what the AI surfaces say and who they cite — a different, and in some ways more actionable, dataset.
Is scraping Google's SERPs allowed?
Google's terms of service address automated access, and you should read them and decide for yourself. ScrapeMaster reads pages you loaded in your own browser rather than crawling from a datacenter, but that's not a legal opinion and it's not permission. Google's anti-bot systems are also very good, so heavy extraction will get challenged. Use delays, keep volume low, and see our guide on whether web scraping is legal. Not legal advice.
How long until I have a useful time series?
Three or four samples before you see anything. Eight to twelve weeks before you can distinguish a trend from noise on a non-deterministic surface. Start now — the annoying thing about time series is you can't retroactively collect one.
Does auto-detect send my page content to your servers?
No. During auto-detect, the page's HTML structure — not its content — is sent to our analysis API so it can suggest columns. Extracted data stays in your browser's IndexedDB and is never uploaded.
Can it handle infinite-scroll SERPs and AI Mode?
Yes. Next-page buttons, load-more, numbered pagination, and infinite scroll are all supported, and it reads the post-render DOM so JS-heavy surfaces work. See handling pagination and infinite scroll.
What should I tell my client about sample quality?
The truth, in one line: a convenience sample of N queries, from one location, on a personalized non-deterministic surface, collected weekly. Then show the date column. Clients trust caveated data more than confident data, and they should.
Bottom line
The July 9, 2026 expansion put AI performance reports in front of more sites, and those reports say the same thing to all of them: here are your impressions, pages, countries, devices, and dates. Nothing about clicks. Nothing about queries. Nothing about what the AI said.
That gap is not going to be filled by waiting. It gets filled by picking a query set, sampling it on a schedule, keeping a date column, and building a record you own — while being straight about the fact that it's a small personalized sample and saying so out loud in every deck you put it in. That's still infinitely more than an impression count, and it's more honest than any CTR figure being sold to you right now.
ScrapeMaster is free, opens in a side panel, auto-detects structure in 2–4 seconds, follows detail links, handles pagination, exports to CSV/XLSX/JSON/clipboard, and keeps every extracted row on your machine. Start the time series this week. In October you'll be glad you did.