Model sensitivity: GPT-5.6 → GPT-6

We compared the six nights before the switch (8 to 13 October) with the first three after (15 to 17 October): around 5,800 ChatGPT answers, about 4,000 before and 1,800 after.

GPT-6 Luna gives shorter answers that name fewer businesses, so each answer has fewer slots, and it searches more to decide who fills them.

Fewer slots could have pushed many businesses out of the answers. On average they did not. Share of answers that name the business we track for each site:

Gemini and Copilot kept their models and drifted up a little over the same nights, so ChatGPT trails them slightly. Per site, the average move on ChatGPT was 4 points, about what Gemini and Copilot move with no change at all. Only one site moved 10 points or more. A hint to watch: four of the five sites ChatGPT named most often lost 5 to 7 points, which fits fewer slots per answer, but three nights is too early to call it. The flip side: a business that is still named now shares the answer with about three others instead of five, so each mention is worth more.

So a model change shows up as a step in the trend, not a verdict on your work. We mark it on the timeline, and we read our own changes against an engine whose model stayed the same.

Three nights so far, measured through OpenAI's API rather than the ChatGPT app. Claude changed model the same day, so it is left out here.