The AI Conversations Leaking Into Your Search Console

Aug 18, 2026 02:00 AM - 1 hour ago 1

In early August, an SEO called Anastasia Kourou noticed queries successful her Search Console study that didn’t look for illustration searches: “Yes.” “Yes spell on.” “Yes, pricing.”

She posted a screenshot and asked John Mueller whether Search Console was search what group opportunity to AI.

His reply confirmed it.

Search Console includes AI Overviews and AI Mode information successful the wide capacity report, and Google’s documentation explains the mechanism. A follow-up mobility wrong AI Mode counts arsenic a marque caller query, and everything successful the consequence gets attributed to it. When personification tells the AI “yes spell on” and your page appears successful what comes back, Search Console records an belief for your page against the query “yes spell on.”

MCP covered the thread connected August 6, and Ross Tavendale asked the mobility everyone was circling. If Search Console is signaling people’s responses to AI Mode, really do we reverse technologist it?

Nobody answered him.

This station is my answer.

I pulled each 1 of these fragments retired of 16 months of my ain Search Console data, worked retired that they travel successful 7 recognizable kinds, and built the classifier into my free Search Console MCP truthful you tin tally it connected your tract pinch 1 prompt. Sorting them is what makes the leak usable.

The full point comes down to 4 ideas.

Google’s caller AI study hides queries, but the queries leak into the study you already use.
The leaked fragments travel successful 7 recognisable kinds, and a classifier tin benignant each one.
What you tin spot is simply a floor. Most of the speech is successful impressions Google anonymises.
Joined pinch the caller Generative AI report’s export, the fragments move into a page-level picture.

I recovered 1,127 queries and 20,300 impressions crossed 16 months connected my site, against millions of mean impressions. Small, and each statement of it is thing a existent convention said aliases ran.

The Report Everyone Asked For Is Missing Its Queries

Google launched Generative AI capacity reports successful June and, as of August 11, they’re unrecorded for everyone. The study shows really often your pages look wrong AI Overviews and AI Mode. It has five information views. Impressions, pages, countries, devices, dates.

Queries and clicks are the 2 things it leaves out, and there’s nary API support either. I checked that past portion connected my ain spot conscionable to beryllium sure. The Search Analytics API’s type parameter still ends astatine googleNews, the searchAppearance magnitude returns thing AI-related, and the BigQuery bulk export schema doesn’t person an AI column. The export fastener successful the UI is the only measurement this information leaves Google.

So, the study tells you really overmuch AI visibility you person and refuses to opportunity for what. (I deliberation we each cognize why lol) Meanwhile the mean capacity report, the 1 you’ve been reference for years, has been picking up AI speech fragments the full time. Nobody filters them retired because officially they’re conscionable queries.

Idea 1: Your Query Report Is Part Conversation Log

AI Mode looks for illustration a chatbot. Underneath, each connection is processed arsenic a Google search, including the follow-ups, and Google folds each of it into the web hunt type alongside the classical 10 bluish links. Your query study now holds 2 different things. Searches group typed, and fragments of conversations group had pinch a exemplary that happened to show your page.

The position information proves it’s the 2nd thing. My tract shows mean position 4.5 for the query “yes.” On the unfastened web, that ranking is impossible; “yes” belongs to songs and grammar sites. Inside an AI response, it makes sense, because Google’s documentation says links successful an AI Overview inherit the position of the full block, and AI Mode citations get counted nether the aforesaid rules erstwhile they scroll into view. Position 4.5 connected a reply connection intends my nexus sat wrong the reply block, not connected a results page.

That’s the leak. The mobility is whether the fragments tin beryllium told isolated from normal queries astatine scale.

Idea 2: The Fragments Come In 7 Kinds

I pulled 16 months of queries from my spot and classified everything that couldn’t beryllium a typed search. Seven kinds came out, and each 1 has a different origin.

Note: I haven’t included each of my queries for evident reasons, conscionable a sample. When you tally this connected your ain GSC account, you’ll get each of your data.

Reply Artifacts

Bare replies: “yes,” “sure,” “really?,” “show me.” A personification answered the AI mid-conversation; the reply was processed arsenic a search, and your page appeared successful the response. Positions present travel from the reply block, not a results page.

Pivot Follow-Ups

Mid-conversation comparisons: “what astir resend?,” “what astir gemini,” “how astir successful chinese?” The personification has an reply successful beforehand of them and asks the AI to trial an alternative, and the replacement they sanction is the 1 they attraction about.

Conversational Questions

Questions addressed to personification alternatively than typed astatine a hunt box: “can you jailbreak meta raybans,” “how do one waste it,” “is it free.” The giveaway is grammar that only useful pinch a listener, which gets its ain conception below.

Tracker Probes

Synthetic prompts from AI visibility tools, tally connected a schedule. Two signatures successful my data. Prompts ending “. my location is usa.” and prompts successful the shape “evaluate the [company] connected [facet].” Nobody searches the aforesaid condemnation each time for 2 months, truthful these are software.

Agent Harnesses

A machine’s complete instructions, logged whole: “search the web for… return the 3 astir applicable results you really recovered … do not invent results aliases urls.” Somewhere an technologist wrote a punctual template, and Google revenge it arsenic a query.

Pasted Strings

Error messages and spreadsheet headers searched as-is, by group aliases pipelines. My information includes a rank tracker’s afloat CSV file header arsenic a azygous query.

Long Uncategorized

Ten aliases much words pinch nary different marker. Some are quoted sentences, immoderate are agents, immoderate are people. The classifier files them for reappraisal alternatively of guessing, which is the correct magnitude of assurance for this bucket.

The classifier is simply a ladder of checks, applied successful order, and a query stops astatine the first 1 it matches.

Take “what astir resend?” It isn’t connected the reply list, truthful it passes the first rung. It matches the second, a pivot, “what about” followed by a short noun. Classification done, and the statement tells maine the rest.

One belief connected my newsletter costs post, position 10, 1 click. Somewhere, personification was asking an AI astir newsletter tools, asked “what astir resend?” sewage my self-hosted setup successful the response, and clicked through. That’s a personification mid-decision, and the query names the replacement they cared about.

There’s nary instrumentality learning successful immoderate of this. The patterns are a curated database anyone tin publication and disagree with, which is the point. Every classification is explainable.

How Is This Different From A Long-Tail Query?

The evident objection to each of this is that agelong queries existed earlier AI. [how to get not provided keywords successful google analytics] is 9 words, and cipher said it to a chatbot. The divided comes down to who the query is addressed to.

When I first showed this to my co-founder Andy, that was his question.

A long-tail query is simply a elaborate petition addressed to nobody. It has its ain taxable and names its ain tools.

A conversational query is addressed to someone, and 4 signals springiness it away.

SignalExample from my dataWhy it can’t beryllium agelong tail

Instructing an assistant	“give maine measurement by step”	Nobody instructs a hunt box
First personification context	“i americium utilizing lmstudio”	You don’t little Google astir your setup
Dangling pronouns	“is it free”, “does it work”	“It” has nary referent successful the query. The referent lives successful a conversation
Politeness	“please clarify”	Nobody says please to an input field

Any of those four, astatine immoderate length, and the query is conversational.

Length unsocial is the anemic signal, truthful it gets quarantined.

Ten aliases much words pinch mobility syntax gets classified arsenic conversational, because typed queries mean 2 to 4 words and almost cipher types 11.

Ten aliases much words pinch nary different marker goes to the reappraisal heap alternatively of into a claim. And thing astatine 9 words aliases less pinch nary of the signals is treated arsenic an mean search, which is why the not provided query supra ne'er enters the dataset.

One show validates the boundary, and I’m not utilizing it yet. Repetition. That not provided drawstring has 1,853 impressions connected my tract because thousands of group type the aforesaid long-tail query.

Conversational strings almost ne'er repeat; astir of excavation beryllium astatine 1 to 3 impressions, because nary 2 group building a follow-up identically. Impressions per drawstring arsenic a typed versus spoken awesome is the evident 2nd type of this classifier.

What 16 Months Of My Data Shows

Start pinch the timeline. Reply artifacts connected my site, by month. Zero impressions from April to November 2025. A first flicker successful December. Then March 2026 switches the people on, and it’s tally astatine 20 to 30 impressions a period since.

ProTip: Go to Search console → Search results → Add select → Query → Select custom.

Add this regex → ^(yes|yeah|ok|okay|sure)[?!.,]*$

You tin spot each of the artifacts.

A query people that didn’t beryllium for 8 months and past becomes persistent is simply a behaviour alteration pinch a day connected it. “Yes” unsocial has 110 impressions, six clicks, and an mean position 4.5 crossed the window. Six group replied to Google’s AI and past clicked done to my tract disconnected the backmost of their ain “yes.”

The pivots statement up pinch my posts 1 for one. [what astir claude] hit my WebMCP guide.

[what astir xcode?] hit the Xcode post. [what astir wayback machine] hit the Wayback guide.

Each 1 is simply a scholar asking the AI to comparison thing against the point my page covers.

The conversational bucket is the biggest, and it has a signature. 559 queries, 8,834 impressions, 13 clicks.

That ratio is what being publication wrong answers looks like.

My biggest azygous statement is [do ai crawlers for illustration gptbot support contented speech for markdown] astatine 2,998 impressions, and my favourite is [what ai hunt monitoring devices fto maine harvester ga4 convention data, gsc click data, and ai citation rates successful a azygous study truthful one tin understand the afloat hunt picture?].

Sixty-five impressions of personification asking an AI for the merchandise class I was building while they asked.

Then location are the machines.

124 of my queries are locator probes pinch 2,902 impressions, and the facet matrix aimed astatine my Roam reappraisal ran regular for 59 consecutive days.

The supplier harnesses full 2,181 impressions, and 1 of them ran astir 2,160 times crossed 5 days successful July, the Dubai upwind punctual from the database above.

Every tally surfaced my ChatGPT teardown successful its results, presumably because the station is astir searching the web for sources. I still don’t cognize whose mirage checker that was. If it was yours, the upwind successful Dubai is hot. (41 degrees today, ugh)

Three rows for the road. Someone pasted a afloat rank locator CSV header, 17 columns of it, and it earned 146 impressions arsenic a query.

Someone other pasted X Support’s full rejection email, the 1 that says the username is available, and Google revenge it against my grip station astatine position 1.3. And a azygous belief exists for [you didnt springiness maine the link], a personification complaining astatine the AI, recorded by Google, revenge against my Wayback post.

Google recorded a personification losing an statement pinch a robot haha. Classic!

Idea 3: What You Can See Is The Tip

Before you tally this connected your site, the main limitation. Google anonymizes uncommon queries, and conversations are almost by meaning uncommon strings. Nobody phrases a follow-up the measurement you do. So the fragments that past into the study are the repeated ones, and the bulk of the speech excavation is hidden.

My BigQuery export shows that complete the past 59 days, 57.7% of my web impressions transportation nary query drawstring astatine all. 454,720 anonymized against 333,651 visible. The 1,127 classified queries are the visible extremity of that pool. Read each number successful this station arsenic a floor.

3 Quick Caveats Before You Run This

The classifier judges really a query looks, not what the personification meant. [is it supplier ready] looks for illustration a dangling pronoun aimed astatine an assistant, and it astir apt is, but personification whitethorn person typed it astatine my agent readiness guide deliberately. Single rows are evidence, not statistics.

AI Overviews and AI Mode can’t beryllium separated astatine the query level. Google reports some nether web search, truthful a part tells you a speech happened, not which aboveground hosted it.

And the shape room is English. A conversational query successful Tamil aliases German only gets caught if it trips the magnitude rung aliases a instrumentality signature. The 1 Spanish probe successful my information was caught by its “. my location is spain.” suffix, not its words. The multilingual exemplary astatine the extremity of this station is built to adjacent that gap. It isn’t retired yet.

Run It On Your Site

The classifier is genai_conversation_queries, a free instrumentality wrong my Search Console MCP, released coming arsenic v2.4.0.

If you don’t person the MCP yet, 1 bid sets it up.

npx -y suganthan-gsc-mcp setup

The wizard signs you into Google, verifies the relationship pinch a unrecorded call, lets you prime your spot from a list, and writes the Claude config for you. The full setup guide covers Claude Desktop, Claude Code, and the manual routes pinch screenshots. It’s free and unfastened source, and the server runs connected your machine, truthful your information moves betwixt you and Google and passes done cipher else’s servers.

If you already usage it, the default config starts the server done npx, which picks up the latest published version.

Restart Claude, and you’re connected v2.4.0. If you installed the one-click desktop bundle instead, download the caller 1 from the releases page.

Either way, the adjacent measurement is 1 sentence.

Run genai_conversation_queries connected my site

One telephone pulls 16 months of your queries done Google’s regex filter, classifies each lucifer down the ladder, attaches the landing pages, and returns the 7 buckets pinch the monthly artifact timeline. BigQuery isn’t required; location are nary caller permissions beyond the Search Console entree you already granted, and your information goes to Google and backmost for illustration each different instrumentality successful the server.

If you tally the BigQuery type of the MCP, v4.1.0 adds gsc_genai_conversation_queries, the aforesaid classifier complete your bulk export. That copy has nary API statement limits, which matters connected ample sites, and it reports the anonymised split, truthful you get your ain iceberg percent adjacent to your fragments.

https://cdn.searchenginejournal.com/wp-content/uploads/2026/08/screendrop-2026-08-12-19-53-25-3dfcfc-1-a7kt5.mp4

John Mueller’s Tip (New)

I woke up to this bully remark from John Mueller connected my LinkedIn station astir it. He suggested mounting up the BigQuery information export, because it mightiness aboveground much of these queries.

He’s correct for ample sites, and the logic is really overmuch information each way retired of Search Console hands over. The UI export stops astatine 1,000 rows per table. The API goes acold deeper but tops retired astir 50,000 rows a time per hunt type. The bulk export has nary statement headdress astatine all. On a large property, the tail the API drops is wherever these fragments live, since conversational strings are uncommon and almost ne'er repeat.

On a tract my size, the routes agree. I checked connected August 14. My busiest time successful the past 4 weeks held 988 chopped visible queries, astir 2% of the regular API ceiling. The classifier recovered 625 speech queries via the API versus 621 done the export, pinch the export lagging only because its information stops 2 days earlier. Small properties get identical answers from either route. If yours is big, tally the BigQuery twin, which is the constituent John was making.

The excavation from Idea 3 stays retired of scope either way. Anonymized impressions beryllium successful the export arsenic rows pinch nary query string, truthful BigQuery tells you really large the hidden excavation is and keeps the strings to itself.

If you person already connected your Search Console spot pinch BigQuery, past you tin usage my MCP to propulsion each of the queries pinch the aforesaid prompt.

You tin entree my BigQuery MCP here.

Idea 4: Add The Generative AI Report Export

The instrumentality you conscionable ran shows what group ask. The Generative AI study shows really overmuch of your visibility sits wrong AI answers, page by page. Joining them gives you some numbers for each page.

The subordinate needs 1 manual measurement from you, and here’s why. Google still doesn’t supply this information via the API, which I verified earlier successful this post, truthful the instrumentality has nary measurement to fetch it. The export fastener successful the UI is the only way.

Five steps, astir 2 minutes.

In Search Console, unfastened Performance, past Generative AI.
Set the day scope to the afloat 16 months.
Press Export, apical right, and take Download CSV. It downloads arsenic a zip.
Unzip it. The record the subordinate needs is Pages.csv, your AI impressions per page.
Attach Pages.csv to the Claude chat wherever you ran the tool, and ask.

Map this export against my genai_conversation_queries results. For each page, show its AI impressions adjacent to the fragments Google recorded connected it.

Now each page has 2 numbers. How overmuch of its visibility is wrong AI features, and which existent speech fragments Google recorded against it.

The overlap gives you a directional hint I haven’t seen anyplace else. (We each cognize each the SEO devices are going to magically travel up pinch this thought soon LOL.)

Reply artifacts and pivots travel from AI Mode specifically, because that’s wherever follow-up conversations happen, while the export counts AI Overviews and AI Mode together. So a page pinch precocious AI visibility and rich | fragments skews towards AI Mode conversations, and a page pinch precocious AI visibility and nary 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 divided Google won’t provide.

When Google adds an API for this report, I will update the MCP to afloat automate this process.

What To Do With What You Find

Each bucket wants a different response, and 2 of them want you to do thing astatine all.

Pivot follow-ups are contented instructions. [what astir resend?] connected my newsletter station intends the comparison readers want isn’t successful the post. Writing the Resend conception answers a mobility group are already asking an AI astir my page.

Conversational questions pinch impressions and nary clicks show which pages get publication wrong answers. For those, title tweaks do very little, because the personification ne'er sees a title. What travels into an AI consequence is the transition that answers the question, the first paragraph nether a heading, the table, the named fact. That’s the portion to activity on, and it’s the aforesaid conclusion I support reaching from the ChatGPT side of this research.

Reply artifacts are corroboration, and a warning. They beryllium pages unrecorded wrong multi-turn conversations, and they’ll wreck a naive study if near in. A keyword instrumentality that doesn’t cognize astir this people will yet urge optimizing for the query “yes.” Exclude the instrumentality buckets, the probes and harnesses, from immoderate opportunity analysis, and dainty artifacts arsenic a awesome alternatively than demand.

And the probes merit a infinitesimal of your attraction connected their own. If you tally an AI visibility tracker, immoderate of those regular prompts are yours, independently logged by Google, which makes your Search Console a free audit of what your locator really runs. If you don’t tally one, 3rd parties are sweeping topics you rank for, each day, and you tin watch them do it.

Which is the wide constituent astir each 7 buckets. These are prompts from existent sessions wherever Google’s AI reached for your pages. Every query fan-out tool, excavation included, generates synthetic questions and hopes they lucifer reality. This is reality, mini sample, floors and all. Use it arsenic a guideline to really group building things astatine your content, not arsenic a measurement dataset.

The Answer To The Question

Ross asked really we reverse-engineer Search Console signaling people’s responses to AI Mode.

The reply turned retired to beryllium a shape library, a strict bid of checks, and the find that the signaling includes much than people.

It includes the AI industry’s ain machinery too, locator probes and supplier harnesses showing up successful everyone’s query reports.

Run the tool, publication your fragments, and spell look astatine what your tract says successful the passages the AI keeps choosing. The conversations are already happening. Google’s been taking minutes.

Why Google Isn’t Showing Query Or Click Data

I cognize this is simply a mini dataset pinch leaked fragments. Even so, there are hardly immoderate clicks, contempt having decent impressions and very bully mean positions.

Which answers why Google isn’t including this successful Generative AI reports.

I uncertainty they’ll ever adhd it, but only clip will tell.

The Multilingual Model

The shape database has edges. It only useful successful English, and connected the separator betwixt a conversational query and an mean agelong one, it makes rule-based guesses. Both problems want machine learning alternatively than much regex, truthful I’ve already trained a model.

It understands languages the patterns can’t, including [kannst du mir das bitte zeigen], and it handles the separator cases the rules conjecture at. (Don’t inquire what it is because I image a German politely asking maine if I want immoderate coffee.)

I’m not releasing it yet. It’s presently trained connected immoderate existent and synthetic data.

For example, connected a look site, [can you frost cooked chicken] is simply a regular hunt that looks for illustration a conversation. So I’m testing it against different niches and making it amended first.

The scheme is simply a free instrumentality wherever you driblet your export successful and get your buckets back, pinch the classification moving successful your browser truthful your queries ne'er time off your machine. (As usual, respecting your privacy.)

It’s coming erstwhile I’m happy pinch the output.

More Resources: