To find AI Mode conversation turns in Google Search Console, open the main Performance report and filter the query table for short, conversational, dependent fragments — the follow-up replies that AI Mode counts as new queries. There's no AI Mode label, so you pattern-match by shape, then read the fragments as conversion-journey signals rather than exact metrics.
This is the practical companion to the finding that AI Mode conversation turns land in your main Search Console report. It's a distinct task from the broader regex method for mining Search Console for AI-search questions: that one treats your classic query data as a proxy for AI intent; this one isolates the AI Mode follow-up turns specifically.
Before you start: what this method can and can't do
State the scope up front so you don't over-read the output. As of mid-2026, Search Console has no filter, dimension, or label that isolates AI Mode queries. Every follow-up turn a user asks in AI Mode is logged as a separate query in the main report, but nothing marks it as AI Mode. So this method infers AI Mode turns from their conversational shape. It's directional evidence — good for spotting journeys and content gaps, not for reporting a clean "AI Mode traffic" number.
Step 1: Open the main Performance report, not the Generative AI report
The query text lives in the general report. Open Google Search Console → Performance → Search results. Do not use the Generative AI performance report for this — it isolates AI-feature impressions but carries no query data at all. If you want AI-feature impression totals, that report is the place; for the conversational question text, you need the main Performance report.
Step 2: Filter the query table for conversational fragments
Add a query filter set to Custom (regex) and apply the patterns below one at a time. Each isolates a different kind of follow-up turn. Treat them as starting points and adapt to your vocabulary:
| Turn type | What it surfaces | Regex (Query → Custom regex) |
|---|---|---|
| Confirmations | "yes", "yeah", "sure", "ok" as standalone replies | `^(yes |
| Continuations | "go on", "tell me more", "and then" | \b(go on|tell me more|and then|what about|explain more|keep going)\b |
| Short fragments | 1–3-word dependent replies | ^(\w+\s?){1,3}$ |
| Context-carrying refinements | replies starting with a pronoun/article | ^(the|that|this|it|those|these|they)\b |
The tell across all four is dependency: a query that is meaningless without a preceding turn is almost certainly a follow-up turn. Export each filter's matches so you can group them in the next step.
The signature of an AI Mode turn is dependency — a query that can't stand on its own, like "yes, pricing" or "the cheaper one." A person types a full question into a search box; they reply in fragments inside a conversation.
Step 3: Group the fragments by journey stage
Sort the exported fragments into conversation-journey stages, because the mix tells you where AI Mode conversations about your brand are heading:
- Confirm / continue ("yes", "go on") — early-to-mid funnel; the user is still exploring.
- Refine ("for a small team", "the cheaper one") — mid funnel; they're narrowing.
- Compare ("vs the other one", "any alternative") — mid-to-late; a decision is forming.
- Price / commit ("yes, pricing", "how much") — late funnel; buying intent.
A cluster of late-funnel fragments on a page means AI Mode conversations are reaching a decision while citing you — the pages worth protecting and refreshing first.
Step 4: Corroborate before you act
Because you can't cleanly segment AI Mode, treat the fragments as a lead, not a verdict. For each high-value pattern:
- Check the landing pages the fragments map to, and confirm each answers the predictable next turn (pricing, integrations, comparisons) as a self-contained, answer-first chunk.
- Pose the likely full questions in the engines — ask AI Mode, ChatGPT, and Perplexity the underlying question and its follow-ups, and note whether you're named and cited at each turn.
- Fold it into your baseline so a single noisy report never drives a decision, the way the five-metric reporting stack prescribes.
What are the limits of this method?
Read the output directionally, for three reasons. First, it's a keyhole view: only turns that produced an impression for your site appear, and Google samples and thresholds query data, so counts understate reality. Second, regex filters inflate metrics — Search Console sums matching rows across pages, so trust the patterns and proportions, not exact totals. Third, the patterns catch false positives: some genuinely short classic queries ("yes bank", "ok google") match too, so eyeball the list before trusting it. The honest frame: this surfaces candidate AI Mode turns to investigate, and the investigation happens in the engines, not in Search Console.
Taking these candidate turns, posing the real questions across AI Mode, ChatGPT, Perplexity, and Claude, and tracking whether your brand is named and your pages cited versus merely surfaced over time — turn by turn — is exactly what Buffy Intel is built to do. Questions: [email protected].