Sometime in 2026, SEOs started noticing queries in Search Console that no human would ever type into a search box: bare replies like "yes," "sure," and "really?", pivot follow-ups like "what about resend?" and "what about gemini," and even whole pasted prompts and error strings. They are not spam. They are fragments of real conversations happening inside Google's AI Mode, leaking into your performance report because of a specific mechanic: a follow-up message inside an AI Mode conversation is processed as a brand-new search query, and every source in the AI's response — including your page — is attributed to it. So when someone deep in an AI conversation types "yes, go on" and the answer that comes back cites you, Search Console dutifully records an impression for your page against the query "yes, go on." This matters because it is the closest thing you have to seeing the actual questions AI users ask about your topic after the first one — and because it is easy to misread. The data lives only in the regular performance report, not the dedicated generative-AI report (which shows impressions but withholds queries and clicks); it is UI-only, absent from the Search Analytics API and, until recently, the BigQuery export; and most of it never surfaces at all, because a large share of AI-driven impressions are anonymized with no query attached. This guide explains the two separate AI reports and which one leaks conversations, why a conversation becomes a query, the recognisable shapes these queries take, what you can and cannot conclude from them, how to actually surface and filter them, and the one category — the pivot follow-up — that is a content brief handed to you for free.
Sometime in 2026, people staring at their Search Console performance report started finding queries that no human would ever type into a search box. Bare replies: "yes," "sure," "really?" Half-finished follow-ups: "what about resend?", "what about gemini". Whole pasted prompts, error messages, even complete AI system instructions, all sitting in the queries table alongside normal searches, each with its own impressions. The first instinct is to call it spam or a bug. It is neither. These are fragments of real conversations happening inside Google's AI Mode, and they are leaking into your report because of one specific mechanic: a follow-up message inside an AI Mode conversation is processed as a brand-new search query, and every source in the AI's response — your page included — gets attributed to that fragment.
So when someone three turns deep in an AI conversation types "yes, go on" and the answer that comes back cites your article, Search Console records an impression for your page against the query "yes, go on." Read carelessly, that is noise to be filtered out. Read carefully, it is something valuable and genuinely new: a partial window into the actual questions AI users ask about your topic after the first one, in their own words. This guide is the read on that data — the two separate AI reports and which one actually leaks conversations, why a conversation turn becomes a query, the recognisable shapes these queries take, what you can and cannot conclude, how to surface and filter them, and the one category worth acting on immediately. It sits next to the measurement read on Search Console's AI impressions metric and the query-shape argument in AI search behavior is replacing keywords, but it is about a different thing: not the impression count, but the strange text in the query column.
The confusion starts because there are two different places AI shows up in Search Console, and they behave oppositely. The first is the dedicated Search generative-AI performance report, which Google launched on June 3, 2026 and rolled out widely over the following weeks. It reports how often your URLs appear inside AI Overviews and AI Mode as impressions — and it deliberately withholds queries and clicks. That is by design: it is a presence metric, covered in full in reading Search Console's AI impressions metric. The point for this guide is what it does not have: because it carries no query text, the conversation fragments never appear there. If you only look at the generative-AI report, you will never see "yes, go on."
The second place is the regular performance report — the same clicks, impressions, position, and queries view you have used for years. Google confirmed, through Search Advocate John Mueller, that information about AI Overviews and AI Mode is included in that general performance report. And because AI Mode's follow-up turns are logged as ordinary queries, they land here, in the query column, mixed in with genuine typed searches and with no label marking them as AI-driven. That is the counterintuitive part worth stating plainly: the queryless AI report hides the queries, and the ordinary report — which was never billed as an AI feature — is where the AI conversations actually show up. If you want to see the words, you look in the old report, not the new one.
The mechanic is simple once you see it. AI Mode is conversational: you ask something, read the answer, then ask a follow-up, and the model keeps the thread. But underneath, Google does not treat that thread as one long session — it treats each turn as its own retrieval. A follow-up question inside AI Mode counts as a brand-new query, the system runs a fresh search for it, and every source that feeds the resulting answer is attributed to that follow-up's text. The user experiences one continuous conversation; the retrieval layer experiences a series of separate queries, one per turn.
That is why the fragments look the way they do. A follow-up is rarely a clean, standalone keyword — it is whatever the person said next in context. Sometimes that is a bare continuation ("yes," "go on," "sure") that only makes sense as a reply. Sometimes it is a pivot ("what about Gemini?", "what about resend?") that assumes the prior turn's subject. Sometimes it is a pasted error message or a long instruction the user dropped in. In every case, the retrieval fires, an answer is built from sources, and if your page is one of them, you get an impression logged against a string that reads like half a sentence — because it is half a sentence. The visible query is a fragment of a conversation you cannot see the rest of.
Once you know to look, these queries fall into recognisable buckets, and telling them apart is what turns a pile of noise into a usable signal. A practitioner analysis that classified sixteen months of one site's Search Console data grouped them into roughly seven shapes; the useful distinction for strategy is between the ones that carry demand and the ones that are machine exhaust.
These are the bare continuations: "yes," "okay," "sure," "really?", "go on." They carry no topical information on their own — they are the glue of a conversation, not a question — so they are useless as keywords and should be excluded from any keyword analysis. Their value is diagnostic, not directional: seeing them at all confirms your pages are being pulled into multi-turn AI Mode conversations, and watching when they first appeared is a rough clock on when AI Mode started sending you impressions. In that same dataset, reply artefacts showed essentially zero impressions through most of 2025, a first flicker late in the year, then persistent monthly activity from around March 2026 — a timeline that tracks AI Mode's scale-up rather than anything you did.
This is the category that matters. A pivot follow-up is a comparison or extension question asked after the first answer: "what about Resend?" on a thread that started about a different email API, "what about the free plan," "how does it compare to X." Each one is a real, specific question a real person asked about your topic, with buying or comparison intent baked in — and it is a question your existing content evidently did not fully answer, or the conversation would not have continued. That makes each pivot follow-up a content gap with a live user behind it, which is exactly the kind of demand a volume tool can never show you, argued more broadly in AI search content opportunities and content gap analysis.
The rest is largely not human demand at all: synthetic tracker probes (structured prompts ending in things like ".my location is usa"), whole agent system prompts logged verbatim as queries, and pasted strings like error messages or spreadsheet headers searched as-is. These are AI systems and automated tools generating queries, not customers. They should be filtered out of any opportunity analysis, because leaving them in skews your keyword picture toward phantom demand — a query that only a bot ever "asked" is not a topic worth building around.
Hold this data to what it can actually support, because it invites over-reading. The honest ceiling comes from a single fact: most AI-driven impressions never show a query at all. In the sixteen-month analysis above, well over half of the site's impressions carried no query string — they were anonymized, with the visible fragments representing only a floor of the real AI activity. So the conversation queries you can see are a small, unrepresentative sample of a much larger stream you cannot. You can read them for texture and for specific opportunities; you cannot read them for volume, and you should never size a market from them.
Within that limit, two conclusions are safe and useful. First, a pivot follow-up is a genuine demand signal — a specific question, asked by a person, that you can answer better. Treat the good ones as briefs. Second, a fragment with high impressions but low or zero clicks is a strong hint that your content is being cited inside the AI answer rather than clicked through to — you are in the response, doing its work, without the visit. That reframes the goal for those queries from earning the click to being the passage the model extracts, the same answer-first, quotable-prose discipline covered in optimizing content for AI answers, not clicks and why specific, detailed content gets cited. What you cannot do is trust any single fragment's intent in isolation, or forget that machine exhaust is mixed in — the reading is a careful filter, not a raw export.
Finding these queries is a matter of knowing where they live and how to sieve them. Because they sit in the regular performance report, they are reachable three ways, each with a different ceiling. The Search Console UI is the quickest for a spot check: open the performance report's query table and filter. A simple regex to isolate the purest reply artefacts is a match on bare affirmatives — something like ^(yes|yeah|ok|okay|sure)[?!.,]*$ — which pulls the "yes," "okay," "sure" residue out in one pass so you can eyeball the rest. The UI's hard limit is that it exports only 1,000 rows per table, so on any site with real traffic you are seeing the top of the list, not the tail where most distinct fragments hide.
For anything beyond a spot check, go to the data exports — and here Google's own advice is explicit. John Mueller has recommended setting up the BigQuery bulk export precisely because it surfaces more of these queries than the interface: the bulk export has no row cap, where the UI stops at 1,000 rows per table and the Search Analytics API caps around 50,000 rows a day. On a large site, the long tail of conversation fragments only becomes visible in BigQuery. Once the data is out, the work is classification: separate the pivot follow-ups (keep — briefs) from reply artefacts (discard — diagnostic only) from machine exhaust (discard — phantom demand), then feed only the pivots into your content planning. Note the one thing you cannot do: pull the queryless generative-AI impression report through the API at all — it is UI-only, so the impression side and the query side of your AI picture come from two different places and cannot be joined in one query.
Strip everything else away and the actionable core of this is small and sharp: your own Search Console is now quietly listing the follow-up questions AI users ask about your topics, and most of your competitors are treating the whole lot as spam. A pivot follow-up like "what about Resend?" on your email-tooling post is not noise — it is the market telling you, in its own words, that the next thing people want after reading you is a comparison you have not written. Multiply that across every topic you cover and you have a demand feed no keyword tool can produce, because these questions were asked inside a conversation, not typed into a search box, and they carry the specificity and intent that conversational turns always do.
The catch is the shape of the opportunity. It does not arrive as a tidy list of ten head terms; it arrives as a long, scattered stream of tiny, specific questions, each worth a real answer and each wanting that answer in the format and on the surface where its asker will next look — a comparison section on the page, a short video, an image post that states the missing fact in extractable text. Answering one is trivial. Answering the stream, continuously, as new fragments appear each month, is a production problem — and it is the production problem, not the finding, that decides whether the signal ever turns into growth. That is the same gap between insight and capacity that runs through Search Console's social reporting and every other first-party-data guide: the report is easy, the response is the work.
This is a demand-mining problem attached to a throughput problem, and the second half is exactly what Kompozy exists to solve. The loop is concrete: pull the pivot follow-ups from your Search Console — the "what about X" and "how does it compare" fragments that AI conversations left behind — and treat each as a brief. Kompozy is a full AI content generation and multi-platform publishing engine, not a single-format tool, so one such brief becomes a spread of answers at once: the blog section or article that answers the comparison in crawlable, citable text; the text and image posts that state the missing fact in real, extractable form; a carousel rendered brand-exact through HyperFrames for the step-by-step; and a face-locked short video for the same answer where AI search increasingly retrieves from video and social.
Two properties of the engine map directly onto what this data rewards. Because every output descends from one Persona Brief, the answers come out in your specific voice with your facts stated consistently — which is what a model needs to corroborate before it cites you, and what keeps you from flooding the tail with the generic filler answer engines ignore. And because autopilot schedules that spread across the supported surfaces — eight social platforms plus blog and email — behind a per-post review gate, you can keep answering new fragments as they appear each month instead of stalling after the first batch, with a human still signing off on every fact. The honest scope: Kompozy cannot see your Search Console for you, cannot decide which follow-ups are worth answering, and cannot make Google cite the result — that judgment and that outcome stay yours and the model's. What it removes is the throughput ceiling that makes acting on a continuous stream of specific, conversational demand impossible by hand, which is the difference between noticing "what about Resend?" in a report and being the page that answers it before anyone else does. The broader discipline this sits inside is generative engine optimization, and the channel-level version is AI search visibility.
AI conversations in Google Search Console are not a glitch and not spam — they are the multi-turn reality of AI Mode surfacing in your data, one follow-up at a time, because Google logs each conversational turn as its own query and attributes every cited source to it. They live in the regular performance report, not the queryless generative-AI report; they are only partly visible, since most AI impressions are anonymized; and they range from useless reply residue to machine exhaust to the one thing that matters, the pivot follow-up that is a content gap with a live user behind it. Read them with a filter, not a magnifying glass: discard the noise, keep the comparison and extension questions, and treat those as the briefs they are. The teams that win here are not the ones who spot the strange queries — everyone will eventually — but the ones who can answer them, across formats and surfaces, faster than the demand moves on.
They are fragments of real conversations inside Google's AI Mode. A follow-up message in an AI Mode conversation is processed as a new search query, and every source in the AI's response gets attributed to it. So when a user deep in a conversation types "yes, go on" and the answer cites your page, Search Console records an impression for your page against the query "yes, go on." Other examples include bare replies like "yes" or "sure," pivot follow-ups like "what about gemini," and whole pasted prompts. They are not spam — they are conversational AI turns logged as queries.
The regular performance report. Google confirmed, through John Mueller, that AI Overviews and AI Mode information is included in the general Search performance report. The dedicated Search generative-AI report launched June 3, 2026 shows impressions and pages but deliberately withholds queries and clicks, so the conversation fragments never appear there — the only place you see them as text is the standard performance report you have used for years, mixed in with ordinary searches.
Not cleanly. The generative-AI impression data is UI-only — the Search Analytics API does not expose an AI type, and the report itself is not in the interface's usual export paths. The conversation fragments do land in the regular performance data, so they reach the Search Analytics API and the BigQuery bulk export as ordinary queries. John Mueller has recommended the BigQuery bulk export specifically, because the UI export caps at 1,000 rows per table while the export has no cap — larger sites surface far more of these queries there.
Mostly demand you could not see before. The most useful category is the pivot follow-up — questions like "what about Resend?" on a post about a competitor — because it is a real comparison someone asked after reading about your topic, i.e. a content gap with a live user behind it. High impressions with low clicks on these fragments also signal that your content is being cited inside AI answers rather than clicked. What you cannot conclude is volume or intent from any single fragment: the data is partial, most AI impressions are anonymized, and machine-generated probes are mixed in.
The pivot follow-ups leaking into your Search Console are literal, first-party briefs — the exact next questions AI users ask about your topic — but they arrive as a scattered stream of tiny, specific queries, each wanting its own answer in the format the asker expected. Kompozy is an AI content generation and multi-platform publishing engine: point it at a mined follow-up and it produces the comparison, the blog section, the image post with the facts in real text, and the short video that answer it, all governed by one Persona Brief so they read as your expertise, then publishes across eight social platforms plus blog and email on autopilot. It turns a stream of conversational demand into answered content at a pace a manual team cannot match.
AI conversations in Google Search Console are fragments of real Google AI Mode conversations — replies like "yes go on," follow-ups like "what about gemini," and pasted prompts — that appear as ordinary queries in your performance report. They leak in because each follow-up inside an AI Mode conversation is logged as a new query and every cited source, including your page, is attributed to it. They live in the regular report, not the queryless generative-AI report, and reveal the questions AI users ask about your topic next.
Get started → · ← All guides · Compare Kompozy vs other tools