On June 3, 2026, Google introduced Search Generative AI capacity reports successful Search Console, dedicated views of impressions wrong generative AI features connected Search, specified arsenic AI Overviews and AI Mode. At last, there’s charismatic AI Overviews reporting showing really often those features impact your pages. What it doesn’t see is the queries down the AI activity, and the study is UI only. I re-verified connected my ain spot connected 11 August 2026 that neither the Search Analytics API nor the BigQuery bulk export exposes the generative AI data, truthful there’s nary charismatic measurement to way AI Mode postulation astatine query level.
Except the queries move up anyway. Fragments of existent AI conversations, replies for illustration “yes spell on,” follow-ups, and full pasted prompts, started appearing arsenic mean queries successful the capacity report. Anastasia Kourou posted astir the unusual strings connected LinkedIn, and Google’s John Mueller replied, confirming this information had ever been disposable successful the capacity study and pointing to Google’s documentation.
I’ve written the afloat relationship pinch 16 months of my ain data separately and deployed my method of extracting these queries via my Google Search Console MCP, truthful what follows is conscionable the applicable ways to propulsion the AI queries successful Search Console retired into the unfastened positive a free instrumentality powered by my ain custom-trained instrumentality learning exemplary that does it for you.
The Scramble For The Queries
Once the leak was public, SEOs started building ways to propulsion those queries out.
The early wave, backmost successful June 2025, was connection count regex, surfacing agelong conversational queries by magnitude alone, thing of 32+ words, for example. Barry Schwartz rounded up those early attempts, crediting Metehan Yeşilyurt, Vijay Chauhan, and David Konitzny. The approaches person sewage a batch much circumstantial since.
Method 1: Glenn Gabe’s Full Inventory In Excel
The capacity study UI caps immoderate array astatine 1,000 rows, truthful connected a large tract astir of your query information stays hidden earlier you start. Glenn Gabe’s answer is to skip the UI. He pulls the complete query group done the Search Analytics API utilizing Analytics Edge successful Excel, past has Claude shape the database and emblem apt AI Mode queries. If you already unrecorded successful Excel, this puts the afloat inventory wrong the workflow you know.
Image Credit: Suganthan MohanadasanMethod 2: Jean-Christophe Chouinard’s Custom Regex
Jean-Christophe Chouinard went aft the aforesaid problem from wrong the report. On August 14, 2026 he published a regex connected LinkedIn that flags conversational strings straight successful the capacity report’s Query filter, utilizing the Custom regex option. He built it by comparing prompts group commonly type into LLMs, from information he has entree to, against strings seldom seen successful GSC earlier AI Overviews.
The consequence is 1 agelong alternation covering punctual verbs (write, draft, generate, summarize, explain, enactment as), greetings, acknowledgements (yes, ok, sounds good, yes spell on), refusals and travel ups (more, continue, show maine more, immoderate different options). Paste it into the select and you get an instant look, free, wrong the study you already person open.
Image Credit: Suganthan MohanadasanSeparately, his January 12, 2026 lawsuit study showed AI Mode clicks are tracked successful Search Console, but the queries down them are almost exclusively anonymized, a caveat that hangs complete each method here.
Method 3: Amin Foroutan’s Advanced GSC Visualizer
Amin Foroutan’s Advanced GSC Visualizer is simply a free Chrome hold that bolts precocious charting, annotations, and an AI adjunct onto your Search Console data, pinch one-click API access. There’s nary AI Mode-specific select successful it.
Image Credit: Suganthan MohanadasanMethod 4: My MCP Servers
My Search Console MCP has the speech detector built successful arsenic a instrumentality called genai_conversation_queries, and my BigQuery MCP runs the aforesaid detector against the bulk export, which is wherever the anonymized excavation of queries lives. Both tally locally, truthful your information moves only betwixt you and Google. What they do good is automation and standard pinch query-level labels connected each row. The limit is setup alternatively than detection, since you request an MCP customer and a config, truthful they suit method users.
Image Credit: Suganthan MohanadasanThe 4 Methods Side By Side
| Glenn Gabe, API into Excel pinch Claude | Full query inventory successful an Excel workflow | Per-run sorting, pattern-based |
| Jean-Christophe Chouinard, civilization regex | Instant free select wrong the report | English shape list |
| Amin Foroutan, Advanced GSC Visualizer | Charting and exploring without exports | No AI Mode circumstantial filter |
| My MCPs | Automation and scale, labels connected each query | Needs an MCP customer and config |
Look down the Limits file and the first 3 stock a gap. Pattern lists and per-run sorting can’t reliably drawback separator cases for illustration rank locator probes, supplier harness prompts, pasted strings, aliases “my location is” probes. They’re built for English too, truthful a reply successful Tamil aliases a code-mixed Hinglish drawstring sails done unlabelled. Method 4 will lick this problem pinch a civilization ML model.
Why This Needs An ML Model
Classifying each query into named buckets, crossed languages, pinch the weird machine-generated strings included is very difficult to do accurately pinch accepted methods. Because location are truthful galore separator cases, etc. So the only reliable measurement to get this moving is to train a instrumentality learning model.
So I decided to build one.
Image Credit: Suganthan MohanadasanHow I Built The Detector
The detector has 2 parts. Deterministic rules ain the nonstop classes (reply artifacts, locator probes, supplier harness prompts), and a trained exemplary owns the fuzzy bound betwixt conversational, agelong tail, and ordinary.
The exemplary is FacebookAI’s xlm-roberta-base, pre-trained connected 100 languages and fine-tuned connected my ain labelled Search Console speech exhaust positive synthetic rows crossed 8 languages, including code-mixed Tanglish and Hinglish. I validated it crossed conscionable complete 120,000 queries scored, and retrained it doubly connected immoderate broke. It runs quantized connected a scale-to-zero Cloudflare container, and the aforesaid exemplary now powers a free instrumentality anyone tin usage without immoderate setup.
The Tool
That instrumentality is the AI Mode and AI Overview query classifier, free connected this site, and you don’t person to motion up aliases springiness maine your email.
If the GSC MCP way sounded for illustration effort, this is the aforesaid detector pinch that effort removed.
Image Credit: Suganthan MohanadasanHow It Works
Drop successful Search Console aliases BigQuery query exports arsenic CSV, and respective files stack into 1 run.
Image Credit: Suganthan MohanadasanStep 1: The exemplary will categorize your queries into 1 of 7 buckets.
Step 2: If you adhd a Search Console generative AI report, the instrumentality will extract page information and representation this against your queries. (In your report, you’ll spot an “AI page” column. You’ll spot the afloat URL erstwhile you export the CSV.)
The deterministic rules explanation the evident buckets correct successful your browser, and those queries ne'er time off your device. I do this because I want to minimize the magnitude of information I process and to thief velocity things up.
Everything other goes to the model, gets classified successful memory, and nary of this is stored. Each query comes backmost pinch a bucket and a assurance score. In plain words, the buckets are afloat conversational queries, short replies for illustration “yes spell on,” follow-up pivots for illustration “what astir the pro plan,” rank locator probes, supplier harness prompts, pasted strings, and mean searches. You tin select by bucket, classifier and confidence, past export each labelled statement to CSV pinch nary statement cap.
How To Use It
Five steps, commencement to finish.
- Get your queries out. Under astir 1,000 queries, the capacity report’s normal export is already complete. Bigger sites tin stack respective filtered exports, propulsion the afloat database done the API pinch Search Analytics for Sheets aliases Looker Studio, aliases dump the BigQuery bulk export array to CSV. If you’re an SEO, you already cognize really to get much than 1000 queries disconnected GSC.
- Drop the CSV aliases CSVs onto the page.
- Press Classify.
- Filter aliases hunt the labelled table.
- Export the labelled CSV.
Limitations
The instrumentality comfortably handles up to 100,000 unsocial queries per tally successful the browser, and a large tally needs the tab kept open, which I judge is simply a very 2010 measurement to compute. The 2nd limit applies to each measurement of search AI Mode successful Google Search Console, excavation included. Google anonymizes uncommon queries; speech strings are rare, and portion of the excavation ne'er shows, truthful dainty the output results arsenic an undercount. My BigQuery export shows 57.7% of my impressions sat successful the anonymized excavation complete the past 59 days, measured connected August 11, 2026.
The exemplary costs maine existent money to run, and arsenic you already cognize me, I do not waste anything, and nary of my devices require immoderate shape of sign-ups aliases paywalls. So I person to put immoderate limitations and caps truthful everyone tin usage this tool.
So, please play nice. You don’t want to beryllium the personification who ruins things.
The instrumentality is successful beta, and if loads of group commencement utilizing it, it will slow down, and processing whitethorn return longer. In that case, hold and effort again.
Also, kindly statement I connection my clip and expertise free of charge, truthful do not request customer support, etc. Yes, I person to constitute this worldly because immoderate group don’t enactment for illustration reasonable adults.
Over 100,000 Queries
Above 100,000 unsocial queries, we request to do a batch tally instead. Send yours in, and my agency runs it arsenic a due batch job, past emails you the complete results. We tin besides move the findings into a full, actionable report, aliases thief pinch your wider AI SEO work.
More Resources:
- Google Now Reports AI Search Impressions. Here’s How To Read Them
- How To Track AI Traffic In GA4 Without Undercounting It
- AI Visibility Measurement: What To Track & What To Ignore
This station was primitively published connected Suganthan.
Featured Image: Roman Samborskyi/Shutterstock
English (US) · 
Indonesian (ID) · 
