Search Console added Search Generative AI performance reports on June 3, 2026, giving site owners an official view of how often their pages show up in AI Overviews and AI Mode. It’s a genuine step forward, but it has a specific hole: the report shows impressions and pages, not the queries that triggered them. Neither the Search Analytics API nor the BigQuery bulk export exposes generative AI query data at that level either — that gap has been re-confirmed as recently as mid-August 2026.
Yet the queries are visible anyway, hiding in plain sight inside the standard performance report. Fragments of real AI conversations — replies like “yes go on,” follow-up questions, entire pasted prompts — show up as ordinary search queries. SEO Anastasia Kourou first flagged the odd strings on LinkedIn, and Google’s John Mueller confirmed the data had been sitting in the performance report all along, pointing back to Google’s own documentation.
Why These Queries Are There At All
When a chatbot runs a background search to ground an answer, that search often gets logged the same way a human-typed query would. The result is a performance report quietly seeded with machine-generated noise: prompt fragments, conversational replies, and follow-up pivots sitting alongside normal keyword traffic. Once SEOs realized what they were looking at, in mid-2025, the race was on to separate the signal from the noise.
Early Approaches: Word Count And Manual Regex
The first wave of detection, in mid-2025, was blunt: flag anything long — 32-plus words, for instance — as a likely conversational query. Barry Schwartz rounded up several of those early efforts, crediting contributions from Metehan Yeşilyurt, Vijay Chauhan, and David Konitzny. The approach worked as a rough filter but missed short replies and non-English strings entirely.
Four Ways SEOs Are Extracting AI Queries Today
A year on, the tooling has gotten considerably more specific. Four approaches are now in active use.
Full inventory via API and spreadsheet
The performance report UI caps any table at 1,000 rows, which hides most of the data on a large site before analysis even starts. One workaround, popularized by SEO consultant Glenn Gabe, pulls the complete query set directly through the Search Analytics API using an Excel add-in, then hands the resulting list to an AI assistant to flag likely conversational queries. It suits teams already living inside Excel who want the full dataset rather than a capped sample.
Custom regex inside the report filter
Rather than exporting anything, SEO Jean-Christophe Chouinard built a regex that flags conversational strings directly in the performance report’s query filter using the “Custom regex” option, publishing it in mid-August 2026. He built the pattern by comparing common LLM prompt phrasing against strings rarely seen in Search Console before AI Overviews existed — prompt verbs like “write,” “draft,” “summarize,” greetings, acknowledgements (“yes,” “sounds good,” “yes go on”), and follow-up phrasing (“more,” “continue,” “any other options”). It’s free, requires no export, and works inside the report SEOs already have open. Chouinard’s earlier case study on AI Mode clicks also found that the queries behind those clicks are almost always anonymized in Search Console — a limitation that applies to every method here, not just this one.
Browser-based visualization tools
Chrome extensions such as Amin Foroutan’s Advanced GSC Visualizer add charting, annotation, and an AI assistant on top of Search Console data with one-click API access. These tools are useful for exploring the data generally, though none currently ship a dedicated AI Mode filter.
MCP servers with a built-in conversation detector
Search Console and BigQuery MCP servers can run a purpose-built conversation-detection tool directly against a site’s own data — including the anonymized bulk-export pool, which is where most of these queries actually live. Because the analysis runs locally, data stays between the site owner and Google. The tradeoff is setup complexity: these tools require an MCP client and configuration, so they suit technical users more than a quick one-off check.
Where Pattern-Based Methods Break Down
Regex and word-count approaches share a common weak spot: they can’t reliably catch edge cases like rank-tracker probes, AI agent harness prompts, pasted strings, or location-probing queries phrased like “my location is.” They’re also built around English phrasing, so a conversational reply in Tamil, or a code-mixed Hinglish or Tanglish string, slips through unlabeled. Classifying every query correctly — across languages, including machine-generated strings that don’t look like natural language at all — is a genuinely hard problem for rule-based systems, which is why a trained classification model handles it more reliably than any single regex ever will.
One free tool built specifically for this problem uses a fine-tuned xlm-roberta-base model — pre-trained across 100 languages and further trained on labeled Search Console conversation data plus synthetic examples spanning eight languages, including code-mixed Tanglish and Hinglish. The classifier was validated against more than 120,000 scored queries and retrained twice to fix systematic errors. It runs on a lightweight, scale-to-zero container so it stays free to use with no signup required.
How A Query Classifier Actually Works
The workflow is straightforward: drop in a Search Console or BigQuery query export as CSV, and the tool sorts each row into one of seven buckets — full conversational queries, short replies (“yes go on”), follow-up pivots (“what about the pro plan”), rank-tracker probes, agent harness prompts, pasted strings, and ordinary searches. Deterministic rules catch the obvious cases directly in the browser, so that portion of the data never leaves the device; everything else is classified in memory and not stored. Each row comes back with a bucket label and a confidence score, filterable and exportable with no row cap. Pairing the classified queries with a Search Console generative AI report export also maps them against the specific “AI page” each query triggered.
In practice, the process is five steps: export queries from Search Console (or stack multiple filtered exports, pull the full set via the API, or dump the BigQuery bulk export for larger sites), upload the CSV, run the classifier, filter or search the labeled results, then export.
The Honest Limitation: An Undercount By Design
Every method for tracking AI Mode activity in Search Console — automated or manual — runs into the same ceiling: Google anonymizes rare queries, and conversational strings are disproportionately rare. A meaningful share of the underlying pool never appears in any export at all. One recent BigQuery analysis found 57.7% of impressions sat in the anonymized pool over a 59-day window measured in mid-August 2026. Any number pulled from these methods should be read as a floor, not a complete count.
This kind of measurement gap is exactly why it pays to track what Google’s generative AI Search Console data actually does and doesn’t show before building reporting around it, and to keep an eye on how Google’s dedicated AI search reports have evolved since they first rolled out in testing.
Frequently Asked Questions
Can I see the exact queries behind my AI Mode traffic in Search Console?
Not officially and not completely. Google’s generative AI performance report shows impressions and pages, not query-level detail, and the Search Analytics API and BigQuery export don’t expose that layer either. Conversational query fragments do leak into the standard performance report, but a meaningful share of them are anonymized and never appear anywhere.
What’s the fastest way to start spotting AI-driven queries?
A custom regex filter applied directly inside the performance report’s query filter is the lowest-effort starting point — no export, no tooling, results in minutes. It will miss edge cases and non-English queries, but it’s a reasonable first pass.
Why do pattern-based methods miss so many queries?
Regex and word-count rules are built around identifiable phrasing, mostly in English. Machine-generated strings, code-mixed languages, and short non-standard replies don’t fit any fixed pattern, which is exactly the kind of ambiguity a trained classification model handles better than a rule list.
Is the data I get from any of these tools a complete picture?
No. Google anonymizes rare queries, and AI-conversation queries are disproportionately rare, so a real share of them never shows up in any export, official or third-party. Treat every count as an undercount and track direction over time rather than an absolute number.