ChatGPT named the brand on 12.4% of buyer questions before GPT-6 and 13.0% after
OpenAI introduced GPT-6 in ChatGPT on October 7, 2026, starting with paid plans, and said Free and Go users would get it from the next day. We read ChatGPT's answers from its logged-out page as a US visitor, and this analysis uses only those, so we were on the free side of that rollout. The model name the page reported in our answers switched from gpt-5-6 to gpt-6 late on October 8, UTC, and stayed there.
That gave us a before and after on the same questions. We took every buyer question (one that doesn't name the brand) that ChatGPT answered under both labels between October 4 and 18: 188 questions for 16 brands, 362 answers on GPT-5.6 and 349 on GPT-6.
Averaged per question, ChatGPT named the brand 12.4% of the time before the switch and 13.0% after. The difference is 0.6 points, with a 95% interval of minus 1.8 to plus 2.9. Of the 30 questions where the brand showed up at least once, 4 went from usually named to usually not, and 3 went the other way.
The answers got 40% longer and cited 64% more sources
The same 188 questions got very different answers in every other respect.
- Length: 529 words per answer on average before, 738 after. 167 of the 188 questions got a longer answer.
- Structure: headings per answer roughly doubled, from about 4 to 8. Bullet points and tables stayed about the same, so the extra length went into more sections.
- Citations: 5.1 cited links per answer before, 8.4 after, from 4.0 distinct domains to 5.9.
- Searching: ChatGPT ran fewer searches to get there, 3.0 per answer before and 2.0 after.
Claude, Perplexity and AI Overviews barely changed on the same questions
A calendar can fool you here. Brands edit pages, roundups get updated, and our own pipeline changes. So we split the same buyer questions at the same date on the three engines that didn't change model that week.
Claude's answers went from 256 words to 255, Perplexity's from 233 to 243, and Google AI Overviews' from 293 to 334. Citations per answer moved by less than one on all three. The brand rate on each moved by under 2.5 points, about as much as ChatGPT's. So the length and sourcing change belongs to ChatGPT's model, and the steady brand rate isn't something the other engines lacked.
Your own site might be cited a little more, and we can't tell yet
The brand's own site was cited in 7.6% of ChatGPT's answers to these buyer questions before the switch and 10.8% after. That's the direction you'd hope for when an answer carries more sources. But the interval on the change runs from minus 0.1 to plus 6.8 points, and it rests on 9 questions that started citing the brand's site against 5 that stopped. Too few to call.
What to check when an engine ships a new model
The common reflex is to treat a model release as a reset and start rewriting pages. On this evidence, wait a week and check instead.
- Find out which model produced each answer. A trend line that mixes two models without saying so will hide the switch or invent one.
- Compare the same questions before and after. A different question set on each side tells you nothing about the model.
- Compare against an engine that didn't change in the same week. If both moved, the model wasn't the cause.
- Read the citations, not only the mention rate. In our data the model changed what ChatGPT cited much more than whether it named the brand.
How Proofsource handles it
Every answer Proofsource stores keeps the model name the engine reported, next to the answer text, the sources it cited and the searches it ran. The answer view shows that model, and the method line under the Visibility and Prompts charts lists both versions when an engine changed model inside the date range, so a before and after like this one is visible instead of blended.
Where this could be wrong
This is logged-out ChatGPT only. OpenAI's release notes say Free and Go use GPT-6 Luna and paid plans GPT-6 Sol, and we didn't measure Sol. OpenAI's model help page still described GPT-5.6 Luna as the Free default when we read it, so the gpt-6 label is what the page reported to our data vendor and we haven't confirmed which variant it was.
GPT-6 can answer with visuals and interactive elements. We capture the answer as text, so anything that only appeared in those wouldn't be counted.
It's a small sample: 188 questions for 16 brands, mostly small companies with Proofsource workspaces, and about half the questions had a single answer on each side, one day apart. An effect of two or three points in either direction is still possible. "Named" means our extraction found the brand in the answer, and a miss counts the same on both sides.
Ten days after a release is early. OpenAI can change the model behind a label without announcing it, and the next update may behave differently from this one.
