Somebody typed the word “yes” into Google, and it showed up in your Search Console.
Not a keyword. Not a question. Just “yes”, sitting in the performance report next to your real search terms, carrying impressions against one of your pages.
That is what a keyword report looks like when the thing on the other side stops being a search box and starts being a conversation.
The report is recording half a dialogue
A recent analysis on Search Engine Journal by Suganthan Mohanadasan documents this properly. Google Search Console has been logging AI Mode follow-up turns as queries since March, and they read exactly like what they are: fragments.
He sorts them into seven types. Reply artifacts like “yes” and “ok”. Pivot follow-ups like “what about resend?”. Full conversational questions, tracker probes, automated agent traffic, and pasted strings that were never typed into a search box by a human at all.
Across sixteen months of data he found 1,127 visible queries carrying 20,300 impressions. That part is measurable.
The part that matters more is the rest of it. 57.7 percent of those impressions arrived with no query string attached. Anonymized, aggregated, gone.
So the fragments you can see are the surviving tip. The majority of the conversation happened, moved impressions against your pages, and left no record of what was actually asked.
Position 4.5 and almost nobody clicked
The second finding in that data is the one I would put in front of a CMO. The AI-cited pages held an average position around 4.5 and generated very few clicks.
Good position. Real impressions. No traffic.
Every instinct built over fifteen years of search says that combination is a bug. It is not. It is the correct outcome of a system where the answer is assembled before the user ever decides whether a link is worth opening.
Which means the number you have used as a proxy for commercial visibility, sessions from organic search, is now measuring a smaller and smaller share of the thing you actually care about.
I have watched marketing directors spend an entire quarterly review explaining a traffic decline that was not a performance problem. The content did not get worse. The counting method stopped matching the behaviour, which is the same diagnosis I made when your metrics broke, not your content.
Read the fragments as intent, not as keywords
Here is the practical part, because the temptation is to file this under “interesting” and move on.
Those fragments are not noise. A pivot follow-up like “what about resend?” tells you that a competing option entered the conversation at the exact moment your page was on screen. That is a competitive signal you would previously have paid a research agency for.
Conversational questions tell you the shape of the second question, the one that comes after the buyer already understood the basics. Most content strategies are still built entirely around the first question.
Pasted strings tell you something different again. They usually mean a machine, not a person, and they are worth separating out before they distort your averages.
So the first job is segmentation. Pull your query report, isolate anything under three words that reads like a reply rather than a search, and stop letting it sit in the same average as your commercial terms.
The second job is accepting what you cannot see. With more than half the impressions anonymized, Search Console cannot be the system of record for AI-driven demand. It was never designed to be.
This is the gap we built GEOflux.ai to close. Search Console tells you that something happened near your page. It does not tell you whether a model recommended you, which sources it used to decide, or what it said about you when your page was not cited at all. That distinction matters more now that AI mentions already move real traffic.
The measurement layer moved before the budget did
The uncomfortable operational reality is that most reporting stacks are now instrumented for a behaviour that is shrinking.
Sessions, bounce rate, pages per visit. All of it assumes the user arrives. A growing share of your commercial influence now happens in an answer the user reads without arriving anywhere.
That does not make the old metrics useless. It makes them incomplete in a way that gets worse every quarter, which is a different and more dangerous problem, because incomplete numbers still look healthy on a dashboard.
The fix is not complicated. Run two reports. One for the visits you still get, one for the answers you appear inside. Treat them as separate systems, because they are.
And when a fragment like “yes” shows up against a product page, read it for what it is. Somebody was in a conversation about you, and it went well enough to keep going.
You just were not invited.