{"slug": "yes-go-on-the-ai-conversations-leaking-into-your-search-console", "title": "\"Yes, Go On\": The AI Conversations Leaking Into Your Search Console", "summary": "Google's Search Console is recording AI Mode conversation fragments as queries, with SEO Anastasia Kourou first noticing odd terms like \"yes go on\" in August and Google's John Mueller confirming the mechanism. Developer Suganthan Mohanadasan analyzed 16 months of his own data, finding 1,127 such queries and 20,300 impressions, and built a free classifier into his Search Console MCP tool to sort them into seven types. The new Generative AI performance reports, live for everyone as of 11 August, show impressions but omit queries and clicks, making the leaked fragments the only way to see what AI conversations drive visibility.", "body_md": "# \"Yes, Go On\": The AI Conversations Leaking Into Your Search Console\n\nA classifier for the AI conversations leaking into Search Console. Seven kinds, from human replies and comparison questions to tracker bots and agent prompts, built from sixteen months of my own data and released as a free tool you run with one prompt.\n\nIn early August, an SEO called Anastasia Kourou noticed queries in her Search Console report that didn’t look like searches. “yes”. “yes go on”. “yes, pricing”.\n\nShe [posted a screenshot](https://www.linkedin.com/posts/anastasia-kourou-4b393034_hi-john-mueller-i-am-noticing-some-unusual-share-7489988919229353984-FJ0m/) and asked John Mueller whether Search Console was tracking what people say to AI.\n\n[His reply](https://www.linkedin.com/feed/update/urn:li:activity:7489988920529653761/?dashCommentUrn=urn%3Ali%3Afsd_comment%3A%287490337203936604160%2Curn%3Ali%3Aactivity%3A7489988920529653761%29) confirmed it.\n\nSearch Console includes AI Overviews and AI Mode data in the general performance report, and [Google’s documentation](https://support.google.com/webmasters/answer/7042828) explains the mechanism. A follow-up question inside AI Mode counts as a brand new query, and everything in the response gets attributed to it. When someone tells the AI “yes go on” and your page appears in what comes back, Search Console records an impression for your page against the query “yes go on”.\n\n[Search Engine Roundtable covered the thread](https://www.seroundtable.com/google-search-console-ai-mode-queries-41821.html) on 6 August, and Ross Tavendale asked the question everyone was circling. If Search Console is recording people’s responses to AI Mode, how do we reverse engineer it?\n\nNobody answered him.\n\nThis post is my answer.\n\nI pulled every one of these fragments out of sixteen months of my own Search Console data, worked out that they come in seven recognisable kinds, and built the classifier into my free [Search Console MCP](https://suganthan.com/blog/google-search-console-mcp-server/) so you can run it on your site with one prompt. Sorting them is what makes the leak usable.\n\nThe whole thing comes down to four ideas.\n\n- Google’s new AI report hides queries, but the queries leak into the report you already use.\n- The leaked fragments come in seven recognisable kinds, and a classifier can sort every one.\n- What you can see is a floor. Most of the conversation is in impressions Google anonymises.\n- Joined with the new Generative AI report’s export, the fragments turn into a page-level picture.\n\nI found 1,127 queries and 20,300 impressions across sixteen months on my site, against millions of ordinary impressions. Small, and every row of it is something a real session said or ran.\n\n## The report everyone asked for is missing its queries\n\nGoogle launched [Generative AI performance reports](https://developers.google.com/search/blog/2026/06/gen-ai-performance-reports) in June and, [as of 11 August](https://www.seroundtable.com/google-search-console-ai-report-live-41850.html), they’re live for everyone. The report shows how often your pages appear inside AI Overviews and AI Mode. It has [five data views](https://support.google.com/webmasters/answer/16984139). Impressions, pages, countries, devices, dates.\n\n**Queries and clicks are the two things it leaves out,** and there’s no API support either. I checked that last part on my own property just to be sure. The Search Analytics API’s [type parameter](https://developers.google.com/webmaster-tools/v1/searchanalytics/query) still ends at googleNews, the searchAppearance dimension returns nothing AI related, and the BigQuery bulk export schema doesn’t have an AI column. The export button in the UI is the only way this data leaves Google.\n\nSo the report tells you how much AI visibility you have and refuses to say for what. (I think we all know why lol) Meanwhile the ordinary performance report, the one you’ve been reading for years, has been picking up AI conversation fragments the whole time. Nobody filters them out because officially they’re just queries.\n\n## Idea 1. Your query report is part conversation log\n\nAI Mode looks like a chatbot. Underneath, every message is processed as a Google search, including the follow-ups, and [Google folds all of it into the web search type](https://developers.google.com/search/docs/appearance/ai-features) alongside the classic ten blue links. Your query report now holds two different things. Searches people typed, and fragments of conversations people had with a model that happened to show your page.\n\nThe position data proves it’s the second thing. My site shows average position 4.5 for the query “yes”. On the open web that ranking is impossible, “yes” belongs to songs and grammar sites. Inside an AI response it makes sense, because [Google’s documentation](https://support.google.com/webmasters/answer/7042828) says links in an AI Overview inherit the position of the whole block, and AI Mode citations get counted under the same rules once they scroll into view. Position 4.5 on a reply word means my link sat inside the answer block, not on a results page.\n\nThat’s the leak. The question is whether the fragments can be told apart from normal queries at scale.\n\n## Idea 2. The fragments come in seven kinds\n\nI pulled sixteen months of queries from my property and classified everything that couldn’t be a typed search. Seven kinds came out, and each one has a different origin.\n\n**Note:** I haven’t included all of my queries for obvious reasons, just a sample. When you run this on your own GSC account, you’ll get all of your data.\n\n### Reply artefacts\n\nBare replies. “yes”, “sure”, “really?”, “show me”. A person answered the AI mid-conversation, the reply was processed as a search, and your page appeared in the response. Positions here come from the answer block, not a results page.\n\n### Pivot follow-ups\n\nMid-conversation comparisons. “what about resend?”, “what about gemini”, “how about in chinese?”. The person has an answer in front of them and asks the AI to test an alternative, and the alternative they name is the one they care about.\n\n### Conversational questions\n\nQuestions addressed to someone rather than typed at a search box. “can you jailbreak meta raybans”, “how do i sell it”, “is it free”. The giveaway is grammar that only works with a listener, which gets its own section below.\n\n### Tracker probes\n\nSynthetic prompts from AI visibility tools, run on a schedule. Two signatures in my data. Prompts ending ”. my location is usa.” and prompts in the form “evaluate the [company] on [facet]”. Nobody searches the same sentence every day for two months, so these are software.\n\n### Agent harnesses\n\nA machine’s complete instructions, logged whole. “search the web for… return the 3 most relevant results you actually found… do not invent results or urls”. Somewhere an engineer wrote a prompt template, and Google filed it as a query.\n\n### Pasted strings\n\nError messages and spreadsheet headers searched as-is, by people or pipelines. My data includes a rank tracker’s full CSV column header as a single query.\n\n### Long uncategorised\n\nTen or more words with no other marker. Some are quoted sentences, some are agents, some are people. The classifier files them for review instead of guessing, which is the right amount of confidence for this bucket.\n\nThe classifier is a ladder of checks, applied in order, and a query stops at the first one it matches.\n\nTake “what about resend?”. It isn’t on the reply list, so it passes the first rung. It matches the second, a pivot, “what about” followed by a short noun. Classification done, and the row tells me the rest.\n\nOne impression on my [newsletter cost post](https://suganthan.com/blog/newsletter-cost-at-scale/), position 10, one click. Somewhere, someone was asking an AI about newsletter tools, asked “what about resend?”, got my self-hosted setup in the response, and clicked through. That’s a person mid-decision, and the query names the alternative they cared about.\n\nThere’s no machine learning in any of this. The patterns are a curated list anyone can read and disagree with, which is the point. Every classification is explainable.\n\n### How is this different from a long tail query?\n\nThe obvious objection to all of this is that long queries existed before AI. “how to get not provided keywords in google analytics” is nine words and nobody said it to a chatbot. The split comes down to who the query is addressed to.\n\nWhen i first showed this to my co-founder Andy that was his question.\n\nA long tail query is a detailed request addressed to nobody. It has its own subject and names its own tools.\n\nA conversational query is addressed to someone, and four signals give it away.\n\nSignal | Example from my data | Why it can’t be long tail |\n|---|---|---|\n| Instructing an assistant | ”give me step by step” | Nobody instructs a search box |\n| First person context | ”i am using lmstudio” | You don’t brief Google about your setup |\n| Dangling pronouns | ”is it free”, “does it work\" | \"It” has no referent in the query. The referent lives in a conversation |\n| Politeness | ”please clarify” | Nobody says please to an input field |\n\nAny of those four, at any length, and the query is conversational.\n\nLength alone is the weak signal, so it gets quarantined.\n\nTen or more words with question syntax gets classified conversational, because typed queries average two to four words and almost nobody types eleven.\n\nTen or more words with no other marker goes to the review pile instead of into a claim. And anything at nine words or fewer with none of the signals is treated as an ordinary search, which is why the not provided query above never enters the dataset.\n\nOne tell validates the boundary, and I’m not using it yet. Repetition. That not provided string has 1,853 impressions on my site because thousands of people type the same long tail query.\n\nConversational strings almost never repeat, most of mine sit at one to three impressions, because no two people phrase a follow-up identically. Impressions per string as a typed versus spoken signal is the obvious second version of this classifier.\n\n## What sixteen months of my data shows\n\nStart with the timeline. Reply artefacts on my site, by month. Zero impressions from April to November 2025. A first flicker in December. Then March 2026 switches the class on, and it’s run at 20 to 30 impressions a month since.\n\n**ProTip:** Go to Search console → Search results → Add filter → Query → Select custom\n\nAdd this regex → ^(yes|yeah|ok|okay|sure)[?!.,]*$\n\nYou can see all of the artefacts.\n\nA query class that didn’t exist for eight months and then becomes persistent is a behaviour change with a date on it. “*yes*” alone has 110 impressions, 6 clicks and average position 4.5 across the window. Six people replied to Google’s AI and then clicked through to my site off the back of their own “*yes*”.\n\nThe pivots line up with my posts one for one. “*what about claude*” hit my [WebMCP guide](https://suganthan.com/blog/webmcp-implementation-guide/).\n\n“*what about xcode?*” hit the [Xcode post](https://suganthan.com/blog/build-macos-ios-apps-with-ai-xcode/). “*what about wayback machine*” hit the [Wayback guide](https://suganthan.com/blog/how-to-use-wayback-machine/).\n\nEach one is a reader asking the AI to compare something against the thing my page covers.\n\nThe conversational bucket is the biggest and it has a signature. 559 queries, 8,834 impressions, 13 clicks.\n\nThat ratio is what being read inside answers looks like.\n\nMy biggest single row is “d*o ai crawlers like gptbot support content negotiation for markdown*” at 2,998 impressions, and my favourite is “*what ai search monitoring tools let me combine ga4 session data, gsc click data, and ai citation rates in a single analysis so i can understand the full search picture?*”.\n\n65 impressions of someone asking an AI for the product category I was building while they asked.\n\nThen there are the machines.\n\n124 of my queries are tracker probes with 2,902 impressions, and the facet matrix aimed at my Roam review ran daily for 59 consecutive days.\n\nThe agent harnesses total 2,181 impressions, and one of them ran roughly 2,160 times across five days in July,\n\nthe Dubai weather prompt from the list above. Every run surfaced my [ChatGPT teardown](https://suganthan.com/blog/how-chatgpt-picks-sources/) in its results, presumably because the post is about searching the web for sources. I still don’t know whose hallucination checker that was. If it was yours, the weather in Dubai is hot. (41 degrees today ugh)\n\nThree rows for the road. Someone pasted a full rank tracker CSV header, seventeen columns of it, and it earned 146 impressions as a query.\n\nSomeone else pasted X Support’s entire rejection email, the one that says the username is available, and Google filed it against my handle post at position 1.3. And a single impression exists for “*you didnt give me the link*”, a user complaining at the AI, recorded by Google, filed against my Wayback post.\n\nGoogle recorded a person losing an argument with a robot haha. Classic!\n\n## Idea 3. What you can see is the tip\n\nBefore you run this on your site, the main limitation. Google anonymises rare queries, and conversations are almost by definition rare strings. Nobody phrases a follow-up the way you do. So the fragments that survive into the report are the repeated ones, and the bulk of the conversation pool is hidden.\n\nMy BigQuery export shows that over the last 59 days, 57.7% of my web impressions carry no query string at all. 454,720 anonymised against 333,651 visible. The 1,127 classified queries are the visible tip of that pool. Read every number in this post as a floor.\n\n**Three quick caveats before you run this.**\n\nThe classifier judges how a query looks, not what the person meant. “is it agent ready” looks like a dangling pronoun aimed at an assistant, and it probably is, but somebody may have typed it at my [agent readiness guide](https://suganthan.com/blog/how-to-make-website-agent-ready/) deliberately. Single rows are evidence, not statistics.\n\nAI Overviews and AI Mode can’t be separated at query level. Google reports both under web search, so a fragment tells you a conversation happened, not which surface hosted it.\n\nAnd the pattern library is English. A conversational query in Tamil or German only gets caught if it trips the length rung or a tool signature. The one Spanish probe in my data was caught by its ”. my location is spain.” suffix, not its words. The multilingual model at the end of this post is built to close that gap. It isn’t out yet.\n\n## Run it on your site\n\nThe classifier is `genai_conversation_queries`\n\n, a free tool inside my [Search Console MCP](https://github.com/Suganthan-Mohanadasan/Suganthans-GSC-MCP), released today as v2.4.0.\n\nIf you don’t have the MCP yet, one command sets it up.\n\n```\nnpx -y suganthan-gsc-mcp setup\n```\n\nThe wizard signs you into Google, verifies the connection with a live call, lets you pick your property from a list, and writes the Claude config for you. The [full setup guide](https://suganthan.com/blog/google-search-console-mcp-server/) covers Claude Desktop, Claude Code and the manual routes with screenshots. It’s free and open source, and the server runs on your machine, so your data moves between you and Google and passes through nobody else’s servers.\n\nIf you already use it, the default config starts the server through npx, which picks up the latest published version.\n\nRestart Claude and you’re on v2.4.0. If you installed the one click desktop bundle instead, download the new one from [the releases page](https://github.com/Suganthan-Mohanadasan/Suganthans-GSC-MCP/releases).\n\nEither way, the next step is one sentence.\n\n```\nRun genai_conversation_queries on my site\n```\n\nOne call pulls sixteen months of your queries through Google’s regex filter, classifies every match down the ladder, attaches the landing pages, and returns the seven buckets with the monthly artefact timeline. BigQuery isn’t required, there are no new permissions beyond the Search Console access you already granted, and your data goes to Google and back like every other tool in the server.\n\nIf you run the [BigQuery version of the MCP](https://suganthan.com/blog/bigquery-mcp-server/), v4.1.0 adds `gsc_genai_conversation_queries`\n\n, the same classifier over your bulk export. That twin has no API row limits, which matters on large sites, and it reports the anonymised split, so you get your own iceberg percentage next to your fragments.\n\n## Idea 4. Add the Generative AI report export\n\nThe tool you just ran shows what people ask. The Generative AI report shows how much of your visibility sits inside AI answers, page by page. Joining them gives you both numbers for every page.\n\nThe join needs one manual step from you, and here’s why. Google still doesn’t provide this data via the API, which I verified earlier in this post, so the tool has no way to fetch it. The export button in the UI is the only way.\n\nFive steps, about two minutes.\n\n- In Search Console, open Performance, then Generative AI.\n- Set the date range to the full sixteen months.\n- Press Export, top right, and choose Download CSV. It downloads as a zip.\n- Unzip it. The file the join needs is Pages.csv, your AI impressions per page.\n- Attach Pages.csv to the Claude chat where you ran the tool, and ask.\n\n```\nMap this export against my genai_conversation_queries results. For each\npage, show its AI impressions next to the fragments Google recorded on it.\n```\n\nNow every page has two numbers. How much of its visibility is inside AI features, and which actual conversation fragments Google recorded against it.\n\nThe overlap gives you a directional hint I haven’t seen anywhere else. (We all know all the SEO tools are going to magically come up with this idea soon LOL)\n\nReply artefacts and pivots come from AI Mode specifically, because that’s where follow-up conversations happen, while the export counts AI Overviews and AI Mode together. So a page with high AI visibility and rich fragments skews towards AI Mode conversations, and a page with high AI visibility and no fragments skews towards one-shot AI Overview citations. It’s an inference, not a measurement, but it’s the only wedge anyone has into a split Google won’t provide.\n\nWhen Google adds an API for this report, I will update the MCP to fully automate this process.\n\n## What to do with what you find\n\nEach bucket wants a different response, and two of them want you to do nothing at all.\n\nPivot follow-ups are content instructions. “what about resend?” on my newsletter post means the comparison readers want isn’t in the post. Writing the Resend section answers a question people are already asking an AI about my page.\n\nConversational questions with impressions and no clicks show which pages get read inside answers. For those, title tweaks do very little, because the person never sees a title. What travels into an AI response is the passage that answers the question, the first paragraph under a heading, the table, the named fact. That’s the part to work on, and it’s the same conclusion I keep reaching from [the ChatGPT side](https://suganthan.com/blog/how-chatgpt-picks-sources/) of this research.\n\nReply artefacts are corroboration, and a warning. They prove pages live inside multi-turn conversations, and they’ll wreck a naive analysis if left in. A keyword tool that doesn’t know about this class will eventually recommend optimising for the query “yes”. Exclude the machine buckets, the probes and harnesses, from any opportunity analysis, and treat artefacts as a signal rather than demand.\n\nAnd the probes deserve a minute of your attention on their own. If you run an AI visibility tracker, some of those daily prompts are yours, independently logged by Google, which makes your Search Console a free audit of what your tracker actually runs. If you don’t run one, third parties are sweeping topics you rank for, every day, and you can watch them do it.\n\nWhich is the general point about all seven buckets. These are prompts from real sessions where Google’s AI reached for your pages. Every query fan-out tool, mine included, generates synthetic questions and hopes they resemble reality. This is reality, small sample, floors and all. Use it as a guide to how people phrase things at your content, not as a volume dataset.\n\n## The answer to the question\n\nRoss asked how we reverse engineer Search Console recording people’s responses to AI Mode.\n\nThe answer turned out to be a pattern library, a strict order of checks, and the discovery that the recording includes more than people.\n\nIt includes the AI industry’s own machinery too, tracker probes and agent harnesses showing up in everyone’s query reports.\n\nRun the tool, read your fragments, and go look at what your site says in the passages the AI keeps choosing. The conversations are already happening. Google’s been taking minutes.\n\n## Why Google isn’t showing query or click data\n\nI know this is a small dataset with leaked fragments. Even so, **there are barely any clicks**, despite having decent impression and very good average positions.\n\nWhich answers why Google isn’t including this in Generative AI reports.\n\nI doubt they’ll ever add it, but only time will tell.\n\n## The multilingual model\n\nThe pattern list has edges. It only works in English, and on the border between a conversational query and an ordinary long one it makes rule-based guesses. Both problems want **machine learning** rather than more regex, so I’ve already trained a model.\n\nIt understands languages the patterns can’t, including “kannst du mir das bitte zeigen”, and it handles the border cases the rules guess at. (Don’t ask what it is because i picture a german politely asking me if i want some coffee)\n\nI’m not releasing it yet. It’s currently trained on some real and synthetic data.\n\nFor example on a recipe site “can you freeze cooked chicken” is a regular search that looks like a conversation. So I’m testing it against other niches and making it better first.\n\nThe plan is a free tool where you drop your export in and get your buckets back, with the classification running in your browser so your queries never leave your machine. (As usual respecting your privacy)\n\nIt’s coming when I’m happy with the output.\n\nWant this kind of AI search forensics run on your site?\n\nIf you want help applying this on your own site, my agency Snippet Digital takes on this kind of work. Send an enquiry and I will be in touch.\n\n[Work with me](/work-with-me/)\n\nEnjoying this? Get the next one in your inbox.\n\nI email when I publish something new. Usually 2 to 4 emails a month. Unsubscribe anytime.\n\nJoin 1,230+ readers. Unsubscribe anytime.\n\n[Suganthan Mohanadasan](/about/)\n\nNorwegian entrepreneur with 20+ years in SEO. Co-founder of Keyword Insights and Snippet Digital. Based in Dubai.", "url": "https://wpnews.pro/news/yes-go-on-the-ai-conversations-leaking-into-your-search-console", "canonical_source": "https://suganthan.com/blog/ai-mode-queries-search-console/", "published_at": "2026-08-13 00:00:00+00:00", "updated_at": "2026-08-13 08:45:50.214795+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-tools", "ai-policy"], "entities": ["Google", "Search Console", "Anastasia Kourou", "John Mueller", "Suganthan Mohanadasan", "Search Console MCP", "Search Engine Roundtable", "Ross Tavendale"], "alternates": {"html": "https://wpnews.pro/news/yes-go-on-the-ai-conversations-leaking-into-your-search-console", "markdown": "https://wpnews.pro/news/yes-go-on-the-ai-conversations-leaking-into-your-search-console.md", "text": "https://wpnews.pro/news/yes-go-on-the-ai-conversations-leaking-into-your-search-console.txt", "jsonld": "https://wpnews.pro/news/yes-go-on-the-ai-conversations-leaking-into-your-search-console.jsonld"}}