On June 3, 2026, Google launched Search Generative AI efficiency reviews in Search Console, devoted views of impressions inside generative AI options on Search, similar to AI Overviews and AI Mode. Ultimately, there’s official AI Overviews reporting displaying how usually these options contain your pages. What it doesn’t embody is the queries behind the AI exercise, and the report is UI solely. I re-verified alone property on 11 August 2026 that neither the Search Analytics API nor the BigQuery bulk export exposes the generative AI knowledge, so there’s no official approach to observe AI Mode site visitors at question degree.
Besides the queries flip up anyway. Fragments of actual AI conversations, replies like “sure go on,” follow-ups, and entire pasted prompts, began showing as abnormal queries within the efficiency report. Anastasia Kourou posted in regards to the unusual strings on LinkedIn, and Google’s John Mueller replied, confirming this knowledge had all the time been out there within the efficiency report and pointing to Google’s documentation.
I’ve written the total account with 16 months of my very own knowledge individually and deployed my methodology of extracting these queries by way of my Google Search Console MCP, so what follows is simply the sensible methods to drag the AI queries in Search Console out into the open plus a free instrument powered by my very own custom-trained machine studying mannequin that does it for you.
The Scramble For The Queries
As soon as the leak was public, SEOs began constructing methods to drag these queries out.
The early wave, again in June 2025, was phrase rely regex, surfacing lengthy conversational queries by size alone, something of 32+ phrases, for instance. Barry Schwartz rounded up these early makes an attempt, crediting Metehan Yeşilyurt, Vijay Chauhan, and David Konitzny. The approaches have gotten much more particular since.
Technique 1: Glenn Gabe’s Full Stock In Excel
The efficiency report UI caps any desk at 1,000 rows, so on an enormous web site most of your question knowledge stays hidden earlier than you begin. Glenn Gabe’s reply is to skip the UI. He pulls the entire question set via the Search Analytics API utilizing Analytics Edge in Excel, then has Claude arrange the record and flag doubtless AI Mode queries. In case you already stay in Excel, this places the total stock contained in the workflow you already know.
Technique 2: Jean-Christophe Chouinard’s Customized Regex
Jean-Christophe Chouinard went after the identical downside from contained in the report. On August 14, 2026 he revealed a regex on LinkedIn that flags conversational strings immediately within the efficiency report’s Question filter, utilizing the Customized regex choice. He constructed it by evaluating prompts individuals generally sort into LLMs, from knowledge he has entry to, in opposition to strings hardly ever seen in GSC earlier than AI Overviews.
The result’s one lengthy alternation overlaying immediate verbs (write, draft, generate, summarize, clarify, act as), greetings, acknowledgements (sure, okay, sounds good, sure go on), refusals and observe ups (extra, proceed, present me extra, some other choices). Paste it into the filter and also you get an prompt look, free, contained in the report you have already got open.
Individually, his January 12, 2026 case examine confirmed AI Mode clicks are tracked in Search Console, however the queries behind them are nearly solely anonymized, a caveat that hangs over each methodology right here.
Technique 3: Amin Foroutan’s Superior GSC Visualizer
Amin Foroutan’s Superior GSC Visualizer is a free Chrome extension that bolts superior charting, annotations, and an AI assistant onto your Search Console knowledge, with one-click API entry. There’s no AI Mode-specific filter in it.
Technique 4: My MCP Servers
My Search Console MCP has the dialog detector inbuilt as a instrument known as genai_conversation_queries, and my BigQuery MCP runs the identical detector in opposition to the majority export, which is the place the anonymized pool of queries lives. Each run domestically, so your knowledge strikes solely between you and Google. What they do properly is automation and scale with query-level labels on each row. The restrict is setup slightly than detection, because you want an MCP consumer and a config, in order that they swimsuit technical customers.
The 4 Strategies Aspect By Aspect
| Technique | Robust at | Limits |
|---|---|---|
| Glenn Gabe, API into Excel with Claude | Full question stock in an Excel workflow | Per-run sorting, pattern-based |
| Jean-Christophe Chouinard, {custom} regex | Immediate free filter contained in the report | English sample record |
| Amin Foroutan, Superior GSC Visualizer | Charting and exploring with out exports | No AI Mode particular filter |
| My MCPs | Automation and scale, labels on each question | Wants an MCP consumer and config |
Look down the Limits column and the primary three share a spot. Sample lists and per-run sorting can’t reliably catch edge circumstances like rank tracker probes, agent harness prompts, pasted strings, or “my location is” probes. They’re constructed for English too, so a reply in Tamil or a code-mixed Hinglish string sails via unlabelled. Technique 4 will clear up this downside with a {custom} ML mannequin.
Why This Wants An ML Mannequin
Classifying each question into named buckets, throughout languages, with the bizarre machine-generated strings included could be very exhausting to do precisely with standard strategies. As a result of there are such a lot of edge circumstances, and so forth. So the one dependable approach to get this working is to coach a machine studying mannequin.
So I made a decision to construct one.
How I Constructed The Detector
The detector has two elements. Deterministic guidelines personal the precise lessons (reply artifacts, tracker probes, agent harness prompts), and a skilled mannequin owns the fuzzy boundary between conversational, lengthy tail, and abnormal.
The mannequin is FacebookAI’s xlm-roberta-base, pre-trained on 100 languages and fine-tuned alone labelled Search Console dialog exhaust plus artificial rows throughout eight languages, together with code-mixed Tanglish and Hinglish. I validated it throughout simply over 120,000 queries scored, and retrained it twice on no matter broke. It runs quantized on a scale-to-zero Cloudflare container, and the identical mannequin now powers a free instrument anybody can use with none setup.
The Device
That instrument is the AI Mode and AI Overview question classifier, free on this web site, and also you don’t have to enroll or give me your e-mail.
If the GSC MCP route gave the impression of effort, this is identical detector with that effort eliminated.
How It Works
Drop in Search Console or BigQuery question exports as CSV, and a number of other information stack into one run.
Step 1: The mannequin will classify your queries into considered one of seven buckets.
Step 2: In case you add a Search Console generative AI report, the instrument will extract web page knowledge and map this in opposition to your queries. (In your report, you’ll see an “AI web page” column. You’ll see the total URL while you export the CSV.)
The deterministic guidelines label the plain buckets proper in your browser, and people queries by no means go away your machine. I do that as a result of I wish to decrease the quantity of information I course of and to assist pace issues up.
The whole lot else goes to the mannequin, will get labeled in reminiscence, and none of that is saved. Every question comes again with a bucket and a confidence rating. In plain phrases, the buckets are full conversational queries, brief replies like “sure go on,” follow-up pivots like “what in regards to the professional plan,” rank tracker probes, agent harness prompts, pasted strings, and abnormal searches. You possibly can filter by bucket, classifier and confidence, then export each labelled row to CSV with no row cap.
How To Use It
5 steps, begin to end.
- Get your queries out. Below about 1,000 queries, the efficiency report’s regular export is already full. Greater websites can stack a number of filtered exports, pull the total record via the API with Search Analytics for Sheets or Looker Studio, or dump the BigQuery bulk export desk to CSV. In case you’re an search engine optimization, you already know the way to get greater than 1000 queries off GSC.
- Drop the CSV or CSVs onto the web page.
- Press Classify.
- Filter or search the labelled desk.
- Export the labelled CSV.
Limitations
The instrument comfortably handles as much as 100,000 distinctive queries per run within the browser, and an enormous run wants the tab stored open, which I settle for is a really 2010 approach to compute. The second restrict applies to each manner of monitoring AI Mode in Google Search Console, mine included. Google anonymizes uncommon queries; dialog strings are uncommon, and a part of the pool by no means exhibits, so deal with the output outcomes as an undercount. My BigQuery export exhibits 57.7% of my impressions sat within the anonymized pool over the past 59 days, measured on August 11, 2026.
The mannequin prices me actual cash to run, and as you already know me, I don’t promote something, and none of my instruments require any type of sign-ups or paywalls. So I’ve to place some limitations and caps so everybody can use this instrument.
So, please play good. You don’t wish to be the one that ruins issues.
The instrument is in beta, and if a great deal of individuals begin utilizing it, it’ll decelerate, and processing might take longer. In that case, wait and check out once more.
Additionally, kindly observe I supply my time and experience freed from cost, so don’t demand buyer assist, and so forth. Sure, I’ve to write down these things as a result of some individuals don’t act like affordable adults.
Over 100,000 Queries
Above 100,000 distinctive queries, we have to do a batch run as a substitute. Ship yours in, and my company runs it as a correct batch job, then emails you the entire outcomes. We will additionally flip the findings right into a full, actionable report, or assist together with your wider AI search engine optimization work.
Extra Assets:
This put up was initially revealed on Suganthan.
Featured Picture: Roman Samborskyi/Shutterstock
