Google now reports impressions for generative AI features such as AI Overviews and AI Mode, but the query-level signals that explain that activity remain partial and awkward to extract. That matters: without reliable query data you can’t measure which pages AI features surface, how users phrase follow-ups, or whether AI-driven discovery substitutes or supplements organic clicks.

What changed — and what still doesn’t

On June 3, 2026 Google added Search Generative AI performance reporting to Search Console, showing impressions tied to generative features. The new UI gives site owners visibility into how often pages appear inside AI Overviews and AI Mode, but Google did not initially expose the same generative-AI fields through the Search Analytics API or BigQuery bulk export. Reporting on this gap in August 2026 found that the API and BigQuery exports still did not include the generative-AI flags, so the UI remains the most direct place to see those impressions.

At the same time, practitioners noticed conversational fragments — pasted prompts, short replies like “yes go on,” and other AI-style strings — appearing among ordinary queries in the performance report. That leak of conversational text led SEOs and toolmakers to develop practical methods to surface and label AI-related queries from Search Console data.

Four practical ways to pull AI Mode queries

Four approaches have emerged as the most useful in production environments. Each balances coverage, setup effort and language support differently.

1) API export into Excel (full inventory)

The Search Console UI caps query tables at 1,000 rows, hiding long-tail queries on larger sites. Pulling the full query set with the Search Analytics API removes that cap and lets you work with the complete inventory in Excel or a similar workflow. This is simple to adopt if Excel is already central to your reporting; the typical limit is manual sorting and pattern-based flagging per run.

2) Regex filters inside the performance report

Custom regular expressions applied to the Query filter in the Search Console UI can surface conversational prompts and common AI-pattern strings instantly. These regexes are free and fast to use, but they rely on hand-crafted pattern lists that skew toward English and common prompt verbs, which makes them brittle for non-English or code-mixed queries.

3) Client-side explorer extensions

Browser extensions that connect to Search Console can add charting, annotations and quick API export. They improve exploration without full exports, but many extensions don’t include AI-specific filters out of the box — useful for visualization, less so for automated classification.

4) ML-powered classification at scale

Rule-based filters miss edge cases: pasted prompts, tracker probes, agent harness strings, and multilingual or code-mixed replies. The scalable answer is hybrid classification: deterministic rules for obvious categories and a machine-learning model to handle fuzzy boundaries and non-English inputs. That approach supports automated, query-level labels across large exports.

The machine-learning option in practice

Practical ML classifiers for this problem typically use a multilingual transformer as a base, fine-tuned on labeled Search Console queries plus synthetic examples in several languages. One reported implementation fine-tuned xlm-roberta-base across an annotated dataset and validated it on roughly 120,000 scored queries, then deployed a quantized model for low-latency runs. The classifier assigns queries to categories such as full conversational prompts, short replies (“yes go on”), pasted strings, rank-tracker probes and ordinary searches, and returns a confidence score for each label.

There is a material caveat: Google anonymizes rare queries in bulk exports. In the analysis referenced above, a significant share of impressions—measured at 57.7% over a 59-day window in that BigQuery export—fell into an anonymized pool that cannot be labeled. That creates an unavoidable undercount when estimating AI-driven activity from exported query tables.

What marketers and SEOs should do next

  • Export full query sets where possible (API or BigQuery) rather than relying on the 1,000-row UI cap.
  • Use regex filters in the UI to catch obvious conversational queries quickly; treat these as a fast, low-effort signal, not a complete solution.
  • Deploy ML classification if you need multilingual coverage or automated, repeatable labels for large datasets—hybrid rules + model works best.
  • Account for anonymization: expect a partial view and report ranges, not precise totals, for AI Mode traffic until Google exposes generative-AI fields in APIs or exports.

What to watch

The two changes that will materially improve measurement are (1) Google exposing generative-AI fields via the Search Analytics API and BigQuery exports, and (2) any adjustments to anonymization policies that reduce the share of queries placed in the anonymized pool. Until those happen, combine exports, pattern filters and automated classification to estimate AI-driven discovery and track trends over time.

Practically: if your search analytics depend on understanding how AI features affect content discovery, prioritize full exports, sample-labeling and automation now so you have a baseline when the platform-level signals evolve.