AI Mode queries are inflating your Search Console data

If you have opened Search Console in the last month and found impressions attached to the query yes, you did not misread the screen and your property is not broken. You are looking at somebody's second sentence in an AI Mode conversation, counted and filed as if it were a search.

The observation came from Anastasia Kourou, who posted screenshots of queries including yes, yes go on, and yes pricing appearing in her performance data and asked the obvious question. John Mueller answered it on August 6: Search Console includes information on AI Overviews and AI Mode in the general performance report, and a follow-up question inside AI Mode is essentially performing a new query, which is counted separately for impressions and clicks.

That is a bigger statement than it looks. It means the main report, the one every SEO dashboard, every client deck, and every year-over-year comparison is built on, now contains a class of query that nobody types into a search box, that carries no standalone intent, and that cannot be filtered out.

The query that started it

A conversational turn is not a search. When a user asks AI Mode about project management tools and then types yes, tell me about pricing, the second turn inherits all of its meaning from the first. Google resolves that reference internally, retrieves against the resolved intent, and may surface your page. The impression is real. The query string logged against it is not a thing a human would ever search for on its own.

The fourth row is the one that will burn somebody this quarter. Average position is a mean across impressions. Add a pile of impressions from conversational turns where few results compete, and the number improves without a single ranking changing. A team that reports average position monthly will show a win it cannot explain and cannot reproduce.

What Google actually confirmed

It is worth separating what has been stated from what is being inferred, because there is a lot of confident nonsense circulating on this. Two things are on the record from Google, one thing is conspicuously absent, and most of the advice being published this week quietly assumes the absent thing exists.

The practical consequence is that the contamination is unbounded from your side. You cannot compute what share of your impressions came from conversational turns, which means you cannot restate last year's numbers on a like-for-like basis. This is the same structural problem we described when AI search reached its not-provided moment: the data is technically present and analytically unusable.

Why the new AI report cannot fix this

Google spent August rolling out the generative AI performance report more broadly, and on the surface that looks like the answer. It is not. The report carries impressions, pages, countries, devices, and date granularity from hourly to monthly. It carries no queries and no clicks. AI Mode and AI Overviews are not broken out from each other. Mueller also confirmed it is still not on every domain, and like the Discover and News reports it only appears where there is sufficient data.

So one report has the queries but cannot tell you which are conversational, and the other knows they are AI but will not tell you what they were or whether anyone clicked. We went through what the AI report does and does not measure in detail when the generative AI impressions data first landed, and this is the follow-on consequence: the two reports do not reconcile, and they were never designed to.

“Two reports, neither complete, and the gap between them is exactly the number everyone wants. That is not an accident of engineering. It is a decision about what Google is willing to show.”

How to find conversational turns in your account

You cannot filter them, but you can estimate their footprint well enough to caveat a report, and an estimate with a stated method beats a shrug. The heuristic is that conversational turns are short, contain confirmation or continuation language, and carry impressions with almost no clicks attached. None of the three passes below is precise. Together they give you a defensible lower bound and a rough sense of whether this is a rounding error on your property or a real distortion.

Pass 1  Confirmation strings
        Regex filter, custom (regex), query matches:
        ^(yes|yeah|yep|no|ok|okay|sure|go on|continue|tell me more)\b
        Record impressions and clicks. Clicks should be near zero.
 
Pass 2  Continuation phrasing
        Query contains: 'yes ', 'and what about', 'what about the',
        'how about', 'compare those', 'the second one', 'that one'
 
Pass 3  The residual test
        Compare total impressions month over month against clicks.
        Impressions climbing while clicks are flat, with no new pages
        and no ranking movement, is the signature.
 
Report the pass 1 + pass 2 impression total as a FLOOR, never a total.

Treat the result as a floor and say so out loud. Most conversational turns will not match a pattern you wrote, because real follow-ups are phrased in ordinary language that is indistinguishable from a search. The measurable ones are the obvious ones, and the obvious ones are the minority. If your floor comes back at three percent of impressions, the honest statement is at least three percent, not about three percent.

One more diagnostic worth running before you blame AI Mode for anything. Pull page-level impressions for your top twenty pages and check whether the growth is concentrated or spread. Conversational turns spread thinly across many pages. A genuine ranking gain concentrates. If a technical SEO program is being judged on a number that moved for the wrong reason, that distinction is the whole argument.

The year-over-year problem nobody has priced in

Here is the part that will cause arguments in January. AI Mode rolled out gradually and its usage is still growing, which means the volume of conversational turns entering your performance report is growing too. Any year-over-year impression comparison you run from now on is comparing a period with few turns against a period with many. The baseline is moving underneath the metric, and it is moving upward.

The direction matters. Most teams have spent the last eighteen months explaining declining clicks to leadership. Impressions were the consolation metric, the number that let you say visibility is holding even as clicks fall. If impressions are now being inflated by a query class that never clicks, that consolation is worth less than it looks, and the CTR decline it produces will be read as a content quality problem rather than a counting change.

There is a second-order effect on anything that consumes Search Console data downstream. Automated reporting, anomaly alerting, forecasting models, and any bid or budget logic keyed to impression trends will all inherit the distortion without knowing it. If you have an alert that fires on a CTR drop, expect it to fire. If you have a model that predicts traffic from impressions, expect it to over-forecast. Neither system will tell you why. Anyone running a technical SEO and GEO audit this quarter should treat the reporting layer as in scope, not just the site.

Reporting rules while this is unlabeled

Until Google ships a filter, four rules keep a Search Console report defensible. None of them require new tooling.

The second rule deserves an argument, because plenty of teams have spent two years moving away from clicks toward impressions and visibility on the reasonable theory that AI answers reduce clicking. That theory is still right. It just cannot be measured with this metric in this report right now. Where visibility genuinely matters, measure it against the AI surfaces directly rather than borrowing a proxy from a report that is blending two populations. We laid out the shape of that reporting gap in the AI search visibility reporting problem, and nothing about this month's changes closes it.

Do this today: run the two regex passes, write the floor number down with the date, retire sitewide average position from your recurring report, and add the annotation. Fifteen minutes of work that stops a fabricated win from ending up in a board deck, and stops a real loss from hiding underneath one. Source: John Mueller via Search Engine Roundtable, August 6 and August 11, 2026 (coverage).

See where you are cited today

A free snapshot audit of your rankings and AI citations before we ever talk.