ChatGPT's background searches shifted twice in four months, Yeşilyurt finds
A 169M-search sample shows query shapes moving with each model update, and the year itself now appears in 41% of searches.
Metehan Yeşilyurt tracked 169 million searches that ChatGPT ran in the background between July and November 2026. He posted the findings on LinkedIn on 10 November 2026.
After the GPT-5.6 update in August, searches using the site: operator jumped from 0.3% to 22%. Queries containing "best" and "vs" dropped over the same window. When GPT-6 reached all users in October, "best" and "vs" came back, and "review" doubled.
The year appears in 41% of searches in the sample. "2027" is already showing up, ahead of the calendar.
What the shifts suggest about model behaviour
The query shapes moved with each model release. The same assistant reached for different retrieval patterns depending on which version was answering. Writing to any one of these shapes would have been wrong twice in four months.
We think the steadier signal here is the year. If 41% of background searches name a year, content without a clear date is harder for the model to place. That holds across both model versions in the sample.
Checking this against your own pages
Before acting, look at your own logs. Pull the ChatGPT user-agent hits for the last 90 days. Count how many carry a year, a site: operator, or words like "review" and "vs".
If the pattern matches what Yeşilyurt saw, the work is ordinary. Show when a page was written and when it was last checked. Make comparisons honest, including cases where a competitor wins. Publish real reviews with named reviewers rather than summaries of other people's reviews.
None of that is written for the retriever. It is what a reader wants from a source worth citing, and the retrieval follows.