GPT-6 reverses GPT-5.6 search patterns, Konitzny finds in 2.2M responses
Comparison queries, listicles and editorial domains climb again after the default model switch.
David Konitzny compared 5.7 million fan-out queries and 2.2 million ChatGPT responses from before and after OpenAI switched the default model to GPT-6, posting the findings on LinkedIn on 9 October 2026. Many of the search behaviours introduced by GPT-5.6 have moved in the opposite direction.
How ChatGPT now searches
Konitzny reports fan-outs per chat dropped from 3 to 2, a 40% fall. Half of all fan-outs now contain 11 or more words, up from 27%. site: queries fell from 22.6% to 8.3%, and quoted-phrase queries from 9.9% to 2.4%. Fewer searches per chat, each one longer and less narrowly targeted.
What it searches for
Comparison and question phrasing returned. In Konitzny's dataset, "official" doubled from 12.3% to 24.6% of queries. "Compare" and "comparison" tripled from 1.9% to 5.7%. "vs" rose from 1.7% to 3.7%. Question phrasing climbed from 4.0% to 9.8%. GPT-5.6 had moved away from these patterns.
What it cites
Konitzny's citation data shows listicles rising from 6.45% to 9.70% of cited URLs, a 50% increase after declining under the previous model. Editorial domains grew from 7.54% to 10.94%. Product pages slipped from 20.41% to 18.10%, and corporate domains from 58.11% to 53.65%.
How it answers
Answers ending with a follow-up question rose from roughly 0.1% to 22.1% in the same comparison, per Konitzny. ChatGPT now asks users for more detail before continuing, where GPT-5.6 would answer and stop.
Reading the shift
The sample is large and the direction is consistent across four separate measures. We think the sensible first step is to pull your own citation and query logs for the weeks either side of the model switch and check whether the pattern holds for your category before changing anything on the page. Model defaults change, and visibility built around one default can erode quietly when the next one lands. Comparison coverage and editorial placements that were discounted under GPT-5.6 may be worth revisiting on their own merits, not as a tactic aimed at the retriever.