# AI Mode follow-ups count as new queries in Search Console

> Google counts an AI Mode follow-up as a new query. The rule is query refinement, documented years before AI Mode, whose worked example is clicking an image.

Canonical: https://brandonlazovic.dev/articles/gsc-query-report-conversation-turns/  
Author: Brandon Lazovic  
Published: 2026-09-03

## The short version

- Google's documentation confirms that an AI Mode follow-up counts as a new query, with that response's impressions, clicks and position attributed to it rather than to the search that started the conversation.
- The mechanism is query refinement, a rule Google documented long before AI Mode, whose worked examples are clicking a photo in a gallery and switching from web results to the video view.
- One conversation therefore produces as many rows as the reader asked questions, and those rows sit in the same list as typed searches with no field separating them.
- Google's own methodology section says its per-result-type list is not exhaustive, so the documented behaviors are a floor and not a complete account.

Google's documentation states that a follow-up question inside AI Mode counts as a brand-new query, with all of that response's impression, position and click data attributed to the new query. [1] The interesting part is why. Google classifies the follow-up as a **query refinement**, a rule it documented long before AI Mode existed, whose worked examples are clicking a photo in a gallery and switching from web results to the video view. [1] A conversation is being measured by a rule designed for a click. If you read your query list as a keyword report, you are now reading two different things stacked in one column.

> One AI Mode conversation produces as many query rows as the reader asked questions.

## What does Search Console do with an AI Mode follow-up?

In Search Console's accounting, an AI Mode follow-up is a separate search. Google's documentation is unambiguous: "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." [1]

Nothing about that is hidden or unannounced. It sits in the same help page that documents how clicks, impressions and position work for AI Overviews, Accelerated Mobile Pages (AMP), and every other result type. The behavior is public, consistent and easy to verify, which is worth stating plainly before any criticism of what it produces.

What follows from it is arithmetic. A reader who opens AI Mode, asks one question and then asks three follow-ups generates four query rows, and the original search receives credit for only the first response. The other three carry whatever the reader typed next.

## What is a query refinement, and why does it apply here?

A query refinement is Google's term for clicking a link inside Search results that runs a new query without leaving Google. The documentation's examples predate AI Mode by years: searching "cat breeds", then clicking one of the breed images in a gallery, or searching "fat cats" in web results and switching to the image, video or news view. [1] Each of those runs a new search inside the same session.

Google's rule for them reads: "If a user follows a query refinement link they are essentially performing the new query shown in the search terms box. All impression, position, and click data in the new result page are counted as coming from this new user query." [1] Put that beside the AI Mode passage and the two are the same sentence with one noun swapped.

Defining the pattern, **inherited classification** is what happens when a measurement rule written for one interaction absorbs a new one because the new one satisfies the old definition. The refinement rule made sense on its original terms. A refinement was a link, the resulting query appeared in the search box, and a site owner reading the report could see a real search string that a real person had effectively run.

An AI Mode follow-up satisfies the definition and matches none of the reasoning. There is no link, the reader typed a sentence in a conversation, and what lands in your report is that sentence stripped of the exchange it belonged to. My read: nobody decided to measure conversations this way, and the classification simply grew to cover them.

## What does a four-turn conversation look like in your report?

A four-turn conversation looks like four unrelated searches. Each turn becomes its own row with its own impressions and its own position, and the report offers no field indicating that they came from one continuous exchange or that any of them followed from the others. The conversational shape that made the fourth question meaningful is not carried into the data.

That produces two specific distortions. Query counts inflate, because one intent is now several rows. And average position becomes harder to reason about, since a follow-up's position reflects where you sat in a response the reader reached through three prior turns, not in a result set they searched for directly.

John Mueller pointed a confused site owner at this documentation after short conversational strings appeared in their Performance report. [2] Search Engine Journal wrote the exchange up on August 17, 2026. [3] That is one site owner's screenshot, and it illustrates the failure mode without measuring it. How often this happens, and to whom, has not been published by anyone.

## Can you separate AI Mode rows from typed searches?

Google's documentation describes no method for separating AI Mode rows from typed searches. The help page covers AI Mode's click, impression and position methodology in detail, and across that whole section it never names a filter, a dimension or a label that would isolate those rows from ordinary typed searches in the Performance report. [1]

Two honest limits on that statement. Documentation silence is not the same as absence, and I have not verified the current Performance report interface against a live property for this piece, so treat the absence of a filter as undocumented rather than as confirmed. The same help page also says its per-result-type methodology list "is not exhaustive," which is Google telling you the documented behaviors are a floor. [1]

Search Console also excludes data from experiments in Search Labs, which is a separate gap from mislabeling: some AI Mode activity is absent from the report entirely. [1]

## What breaks if you read the query list as demand?

Any analysis that treats a Search Console query row as one person's search decision breaks. Keyword research built on your own query export starts counting conversation turns as independent demand, and a rising query count can now reflect more talkative sessions instead of more interest. The direction of the error is upward, which is the direction least likely to prompt anyone to check.

This site argued in [your AI referral parameter is measuring a fraction of the effect](/articles/ai-referral-tag-undercount/) that AI-driven traffic is undercounted in the reporting a site owner can see. The same surface is now producing the opposite distortion, inflating the query count while suppressing the referral count, and neither error is labeled in the interface. Two errors running in opposite directions on one dashboard is a harder problem than either alone, because they partially mask each other in any aggregate.

## What should you do with query-level analysis now?

Move query-level analysis to the pages report, which the AI Mode follow-up distortion does not reach. The pages report attributes to URLs rather than to query strings, so a four-turn conversation that surfaced one of your pages contributes to that page's row without fragmenting into four. For questions about which content earns visibility, that is now the more trustworthy view.

If you want to see the effect in your own data, sort your query list by impressions ascending and read the tail. Conversational fragments and full sentences are the tell, since neither looks like something a person types into a search box. Finding none is a real answer about your own traffic mix, and not evidence the behavior is absent, and at low AI Mode volume you should expect to find nothing.

What does not work is trying to clean the list. There is no documented way to tell a follow-up from a typed search, so any filter you write is a guess about string shape, and a guess applied to a metric is worse than a known limitation you have written down.

## Sources

1. Google Search Console Help: Search result features and other elements in the Performance report — https://support.google.com/webmasters/answer/7042828?hl=en
2. John Mueller (Google Search Advocate) on LinkedIn — https://www.linkedin.com/feed/update/urn:li:activity:7489988920529653761/
3. Search Engine Journal: The AI Conversations Leaking Into Your Search Console — https://www.searchenginejournal.com/the-ai-conversations-leaking-into-your-search-console/585663/