AI Conversation Fragments Are Leaking Into Search Console

TL;DR: Google’s general Search Console performance report is recording fragments of AI Mode conversations — “yes,” “what about gemini,” “is it free” — as ordinary queries attached to your pages. The new Generative AI report shows AI impressions but strips out queries and clicks entirely. The leaked fragments in your standard report are the only query-level window into what real AI sessions looked like when your page appeared. Operators who classify and use them move faster than those who don’t.

The Report Google Built Has a Deliberate Hole in It

Google rolled out Generative AI performance reports in June 2026 and made them available to all properties by August 11. The report shows impressions, pages, countries, devices, and dates for AI Overviews and AI Mode. It does not show queries. It does not show clicks. There is no API access — the Search Analytics API’s type parameter still stops at googleNews, the searchAppearance dimension returns nothing AI-related, and the BigQuery export schema has no AI column. The export button in the UI is the only way the data leaves Google at all.

So Google tells you how much AI visibility your pages are getting and refuses to tell you for what. Meanwhile, the ordinary performance report — the one operators have been reading for years — has been logging AI conversation fragments the entire time. Nobody filtered them because officially they are just queries. Running a full marketing audit against this data is how you find out what your site is actually being cited for inside AI conversations.

Why Position 4.5 on the Query “Yes” Is Not an Anomaly

When SEO practitioner Anastasia Kourou noticed queries like “Yes,” “Yes go on,” and “Yes, pricing” appearing in her Search Console data in early August, she posted a screenshot and asked Google’s John Mueller directly. Mueller confirmed it: AI Mode processes every follow-up message as a new Google search, and the result gets folded into the standard web search type report alongside classic results. If your page appears in the AI response to that follow-up, Search Console records an impression for your page against that fragment as the query.

Position data exposes the mechanism. A site logging average position 4.5 for the query “yes” is not ranking on the open web for that word — songs and grammar references own that. Inside an AI response block, it makes complete sense. Google’s documentation states that links inside an AI Overview inherit the position of the block, and AI Mode citations follow the same rules once they scroll into view. Position 4.5 on a reply word means the page sat inside an answer block, not on a results page. That is the leak in concrete terms.

Across 16 months of one practitioner’s own Search Console data, this analysis surfaced 1,127 qualifying queries and 20,300 impressions. Small against millions of ordinary impressions, but every row represents a real session where a real person (or an agent) interacted with AI and saw that site’s content.

Seven Fragment Types, One Classifier

The fragments are not random noise. They fall into seven recognizable categories, each with a distinct origin and a different set of signals operators can act on.

Reply artifacts — bare replies like “yes,” “sure,” “really?,” “show me.” A person answered the AI mid-conversation; the reply was processed as a search and the page appeared in the response. Pivot follow-ups — mid-conversation comparisons like “what about resend?” or “what about gemini.” The user already has an answer and is testing an alternative; the alternative they name is the one they actually care about. Conversational questions — grammar that only works with a listener: “can you jailbreak meta raybans,” “how do i sell it,” “is it free.” These use first-person framing, dangling pronouns, or politeness markers that nobody types into a search box. Tracker probes — synthetic prompts from AI visibility monitoring tools, identifiable by repetition patterns and structured suffixes like “. my location is usa.” Agent harnesses — complete machine instructions logged whole, such as “search the web for… return the 3 most relevant results you actually found… do not invent results or urls.” Pasted strings — error messages or spreadsheet headers searched verbatim. Long uncategorized — ten or more words with no other classifier marker; these go to a review pile rather than a forced classification.

The classifier is a ladder of checks applied in order. A query stops at the first rung it matches. Everything is rule-based and explainable — no machine learning, no black box. Every classification can be audited and contested, which matters when you are using the output to change spend or content decisions. This is the kind of signal that sharpens precision targeting decisions at the page and topic level.

Conversational Queries vs. Long-Tail Searches: The Actual Distinction

The obvious objection is that long queries existed before AI. A nine-word search like “how to get not provided keywords in google analytics” predates AI Mode by years. The split comes down to who the query is addressed to.

A long-tail query is a detailed request addressed to nobody. It names its own tools, has its own subject, and is self-contained. A conversational query is addressed to someone, and four signals give it away. First: instructions directed at an assistant (“give me step by step”) — nobody instructs a search box. Second: first-person context (“i am using lmstudio”) — you do not brief Google about your current setup. Third: dangling pronouns (“is it free,” “does it work”) — “it” has no referent in the query; the referent lives in the prior conversation turn. Fourth: politeness (“please clarify”) — nobody says please to an input field.

Length alone is a weak signal and gets quarantined. Ten or more words with question syntax lands in the conversational bucket because typed queries average two to four words. Ten or more words with none of the four signals go to the review pile. Nine words or fewer with no signals are treated as ordinary searches. One validation signal the classifier does not yet use: repetition. A genuine long-tail query accumulates thousands of impressions because many people type the same thing. Conversational fragments almost never repeat — most cluster at one to three impressions — because no two conversations follow the same path.

What This Means for Performance Marketing Operators

For operators running paid and organic programs in high-CAC verticals, this data changes the content brief. Every pivot follow-up in your Search Console is a named competitor or alternative that a real person tested mid-session while your page was already in front of them. That is not keyword research — it is live decision-stage intelligence.

Consider what this looks like in practice for iGaming acquisition: a pivot query like “what about betway” attached to an impression on your sportsbook comparison page tells you exactly which brand a user was weighing when they saw your content. For Forex acquisition, agent harness fragments showing up against your broker review pages confirm that automated research tools are indexing your content as a source — a signal about technical authority, not just traffic. For law firm marketing, conversational questions like “is it free” or “how do I file” attached to practice area pages reveal the objections real users voiced to the AI before they reached your site.

The mechanics apply equally to crypto operator acquisition and CDL recruitment campaigns where intent signals are often thin and expensive to acquire through traditional means. A fragment like “what about owner operators” showing up against a fleet recruitment page is a job-type objection raised in a real AI session — more specific than any keyword planner output.

The operational move is straightforward. Export your Search Console performance data. Run a classifier — rule-based, no ML required — against the query column. Tag every row by fragment type. Join the output against the new Generative AI report export at the page level. The result is a page-level map of what AI sessions looked like when your content appeared, what alternatives users were weighing, and which objections came up mid-conversation. Managed performance programs that feed this signal back into creative briefs and landing page tests will convert AI-driven sessions at higher rates than programs running on search console data alone. The data is already in your account. Most operators are not reading it yet.

Teams uncertain where to start should treat this as a structured discovery exercise — the same mindset that drives a lead qualification audit when you suspect your funnel is leaking at the intent layer. Pull the fragments, classify them, and map them to pages before optimizing anything else.

Originally reported by Search Engine Journal, August 2026.