Hi everyone, this is Neo.

Quick question: have you ever spotted queries in your Google Search Console that don’t look like searches at all?

Things like “yes”, “yes go on”, “yes, pricing”, “what about resend?”…

Don’t laugh — this might be one of the most valuable data leaks you’ll ever pick up in 2026.

In early August, an SEO named Anastasia Kourou posted a screenshot on LinkedIn showing exactly these kinds of words sitting in her Search Console query report. She tagged John Mueller with a pointed question — are you recording what people say to AI?

Mueller’s answer set the whole SEO industry on fire: yes.

And then came the plot twist: Google confirmed the mechanism, but nobody knew how to exploit it. Until mid-August, when an SEO named Suganthan Mohanadasan dug through 16 months of his own Search Console data, sorted these “AI conversation queries” into 7 distinct types, and published a complete identification and exploitation playbook.

In this post, I’ll walk you through the whole story, the 7 query types, and how to put this data to work for your independent site.

The backstory: why is there a “yes” in my Search Console?

Let’s get the timeline straight first.

Early August: Anastasia Kourou spots weird entries in her Search Console query report — one-word replies and sentence fragments like “yes”, “yes go on”, “yes, pricing”. These look nothing like traditional searches — who types just “yes” into a search box?

She posts a screenshot on LinkedIn and asks John Mueller whether Search Console is tracking what people say to AI.

Mueller’s reply quotes an official help center document:

“Search Console includes information on AI Overviews and AI Mode in the general performance report… If a user asks a follow-up question within AI Mode, they are essentially performing a new query. All impression, position, and click data in the new response are counted as coming from this new user query.”

In plain English: Search Console folds AI Overviews and AI Mode data into the main performance report; when a user asks a follow-up question inside AI Mode, that’s essentially a new query, and every impression, position, and click in the new response gets attributed to it.

August 6: Search Engine Roundtable covers the thread. In the comments, Ross Tavendale asks the question everyone was circling: if Search Console is recording people’s responses to AI Mode, how do we reverse engineer it?

Nobody answered.

August 13: Suganthan publishes the answer on his blog — he classified every one of these fragments across 16 months of his own data into 7 recognizable types, and built the logic into a free, open-source classifier so anyone can run it on their own site.

Why this is a “leak”: the official report withholds query data

To understand why this data is valuable, you first need to see what Google’s official AI visibility reporting gives you — and what it hides.

On June 3, 2026, Google launched the Search Generative AI performance report, dedicated to your site’s visibility inside AI Overviews and AI Mode. By August 11, it had rolled out widely — most site owners can see it now.

The report has exactly 5 data views:

• Impressions
• Pages
• Countries
• Devices
• Dates

No queries. No clicks.

There’s no API support either — the Search Analytics API’s type parameter still ends at googleNews, the searchAppearance dimension returns nothing AI-related, and the BigQuery bulk export schema has no AI column. The export button in the UI is the only way the data leaves Google.

So Google tells you how much AI visibility you have, and refuses to tell you what you got it for.

(As Suganthan put it: we all know why, lol.)

Meanwhile, the ordinary performance report you’ve been reading for years has been quietly collecting AI conversation fragments the whole time — because officially, they’re just queries.

How do you prove those “yes” rows really come from AI conversations and not real people? Suganthan did it with position data, and it’s a beautiful argument:

His site shows an average position of 4.5 for the query “yes”. On the open web, that ranking is impossible — “yes” belongs to songs and grammar sites. The only explanation: his links sat inside an AI answer block, because Google’s documentation says links in an AI Overview inherit the position of the whole block, and AI Mode citations are counted the same way once they scroll into view.

A position-4.5 “yes” means your link is inside an AI’s answer, not on a results page.

That’s the leak. The only question left: how do you separate these fragments from normal queries at scale?

The 7 types of AI conversation queries

Suganthan pulled 16 months of queries from his own property and classified everything that couldn’t be a typed search. He found 1,127 queries and 20,300 impressions — tiny against millions of ordinary impressions, but every single row is a trace of a real session.

1. Bare replies

Signature: 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.

2. Mid-conversation comparisons (pivots)

Signature: what about resend?, what about gemini, how about in chinese?

This is the most valuable bucket. The person already has an answer in front of them and is asking the AI to test an alternative — the alternative they name is the one they actually care about.

Suganthan’s pivots line up with his posts one for one: “what about claude” hit his WebMCP guide, “what about xcode?” hit his Xcode post, “what about wayback machine” hit his Wayback tutorial. Every one is a reader asking the AI to compare something against what his page covers.

3. Questions addressed to someone

Signature: can you jailbreak meta raybans, how do i sell it, is it free

The grammar only works with a listener — this is speech directed at a person-like entity, not text typed into a search box.

4. Synthetic prompts from AI visibility tools

Signature: prompts ending “. my location is usa.”; prompts in the form “evaluate the [company] on [facet]”.

Nobody searches the same sentence every day for two months — that’s software. AI visibility trackers run scheduled prompts, and Google files them as queries.

5. A machine’s complete instructions

Signature: “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 logged the whole thing as one query.

6. Error messages and spreadsheet headers

Signature: his data includes a rank tracker’s full CSV column header as a single query (146 impressions), and someone pasted X Support’s entire rejection email.

7. Long queries (10+ words, quarantined)

Signature: ten or more words with no other marker — quoted sentences, agents, or people. These go to a review pile instead of a claim.

The iceberg effect: 1,127 is just what’s above the waterline

Suganthan cross-checked his findings against a BigQuery bulk export, and the reality check is sobering:

Over the last 59 days, 57.7% of his web impressions carried no query string at all — 454,720 anonymized 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.

The timeline tells its own story: zero reply-artifact impressions from April to November 2025. A first flicker in December. Then March 2026 flips the switch on, and it’s run at 20-30 impressions a month ever since.

A query class that didn’t exist for eight months and then became persistent is a behavior change with a date on it — that’s when AI Mode usage took off.

A few headline numbers:

• “yes” alone: 110 impressions, 6 clicks, average position 4.5. Six people replied to Google’s AI and then clicked through to his site off the back of their own “yes”.
• The conversational bucket: 559 queries, 8,834 impressions, 13 clicks.
• Tracker probes: 124 queries, 2,902 impressions; one “facet matrix” ran against his article daily for 59 straight days.
• Agent harnesses: 2,181 impressions; one ran roughly 2,160 times in five days.

And the best one: a single impression for [you didnt give me the link] — a user complaining at the AI, recorded by Google, filed against his Wayback post.

Google recorded a person losing an argument with a robot. Classic.

How to run this on your own site

Suganthan packaged the whole classifier into a free, open-source Search Console MCP. But even without his tool, there are things you can do right now:

1. Quick regex filter

Go to Search Console → Performance → Add filter → Query → Custom, and paste:

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

You’ll instantly see every “bare reply” AI conversation query on your site.

2. Treat pivots as your content roadmap

These queries tell you exactly what people are comparing against your content. A “what about X” query hitting your page means your page is missing the X comparison — that’s a free demand list. Write the X section and you’ll catch questions people are already asking AI about your content.

3. Impressions without clicks: fix the body, not the title

This one matters. In conversational queries, the user never sees your title — they see a generated answer. What travels into an AI response is the passage that answers the question: the first paragraph under a heading, the table, the named fact.

So title tweaks do almost nothing here. Optimize the quoted passage, not the clickable headline — the same conclusion I keep reaching from the ChatGPT side of this research.

4. Exclude the machine buckets from opportunity analysis

Tracker probes and agent harnesses aren’t real demand. Leave them out of any keyword opportunity work, or some tool will eventually recommend you optimize for the query “yes”.

5. Know your data egress limits

• UI export: 1,000 rows per table
• API: ~50,000 rows per day per search type
• BigQuery bulk export: no row cap

Conversational strings almost never repeat, so most of them live in the tail the API drops. On a large property, go straight to the BigQuery export.

Neo’s take

1. AI conversations have infiltrated the most basic data pipeline

The wildest part: no new feature, no announcement — AI conversation data just showed up inside the report you’ve been reading for a decade. Google’s design decision to process every AI Mode message as a search made everyone’s data “dirty” overnight — but it also handed every site owner a free stream of AI behavior data.

2. This is Google’s gift-wrapped “AI behavior audit”

Look at buckets 4 and 5: if you run an AI visibility tracker, the prompts your tool runs are being independently logged by Google. Your Search Console just became a free audit of what your tracker actually does. And if you don’t run one, you’re watching third parties sweep your ranking keywords every day. Either way, it’s intelligence.

3. “Keyword” is being redefined

Traditional keyword tools process typed searches. Your query report now mixes “what people said to AI” with “what machines said to AI”. Without classification, you’re making keyword decisions on dirty data. I believe every serious keyword tool will need a conversation-query classifier built in — whoever ships it first wins.

4. Direct advice for independent site owners

• Spend two minutes a week running the regex above
• Turn pivot queries into a content backlog — they’re demand with names attached
• Optimize the first paragraphs, headings’ lead-ins, tables, and named facts — those are what get cited into AI answers
• Large sites: use the BigQuery export, don’t rely on the UI or API

Wrapping up

From a LinkedIn screenshot on August 6 to a 16-month deep-dive on August 13, and SEJ pushing it to the world on August 17 — the speed of this story’s spread says everything about how hungry the SEO world is for AI data you can actually act on.

The conversations are already happening, and Google has been taking notes. Your move: read your query report, find the searches that don’t look like searches, and look at what your site says in the passages the AI keeps choosing.

That’s where your next traffic opportunity is hiding.

I’m Neo, see you in the next one.