Three New AI Models in Three Weeks: Whose Jump Is It?

On 28 September 2026 Anthropic released Claude Sonnet 5.5. On 7 October OpenAI began moving the chat in ChatGPT to GPT-6, Sol for paying plans and, from the following day, Luna for Free and Go. On 13 October Google made Gemini 4 Argon available in AI Mode in Search for its Pro and Ultra subscribers. Three AI model updates from three providers in a little over a fortnight. If the visibility figures for your brand moved in October, some of that movement may have nothing to do with you.

That is the whole point of this note. A step change in measured AI visibility has two possible authors: you, or the provider. Before anyone takes credit for a rise or goes looking for a culprit behind a fall, it is worth establishing which.

Why a model change moves the numbers is not mysterious. An assistant's answer about suppliers in your category is produced by a particular model, trained on a particular body of text, with its own habits about when to search the web and which sources to trust. Replace the model and you replace all of that at once. The question your prompt asks is the same. The thing answering it is not. Some brands will be named more often afterwards and some less, and the reshuffle will have been decided in a lab, not in your marketing plan.

There is a complication worth stating plainly. A switch in the consumer product is not automatically a switch in what gets measured. A tool that calls a named model through an API keeps getting that model until it is retired or the tool moves to another one. So the GPT-6 rollout in ChatGPT may not move a measured series at all. What it can do is widen the distance between what is measured and what your buyer sees on the screen, which we have written about in the model your buyer uses is not the one we measure. The same applies to Gemini 4 in AI Mode: by Google's own announcement it went to paying subscribers, so most people using AI Mode were not yet getting it.

Four checks before crediting an AI model update, or yourself

None of these is conclusive on its own. Together they usually settle the matter.

• The date. Line the jump up against the release dates. With two scans a month, a model change shows up on the first scan after it, not on the day itself. A jump that predates every release cannot be blamed on any of them.
• One assistant or several. A model change touches one assistant. If your figure rose in Claude and stayed flat in ChatGPT, Gemini and Perplexity, look at Claude's side first. Something you did, such as a new comparison page or a mention in a trade publication, is more likely to show up in more than one place, though not necessarily in all four.
• The competitors. A new model reshuffles the whole category, not only you. If the competitors measured alongside you also moved on the same assistant and the same scan, that is the signature of the model. If you moved and they did not, the explanation is more likely to be specific to you.
• The sources. When an assistant cites its sources, read them. If the new answers point to a page you published last month, that is real evidence in your favour. If they cite the same pages as before and simply rank the names differently, the change is in how the model reads, not in what there is to read.

One check overrides the others: wait for the next scan. Answers vary from one run to the next even when nothing changes, which is why a single check proves nothing. A model change produces a new level that holds. Noise produces a bump that goes back. Two scans a month means a fortnight of patience, which is less expensive than rewriting a content plan around a reading that turns out to be a blip.

For a British exporter measuring in English, French and German, there is one more useful habit: keep a short log of your own actions with dates. Pages published, press coverage, a distributor that updated its catalogue. When a jump arrives, the log tells you whether you had anything to do with it. Without one, everyone remembers having done something that week.

PSentry measures four assistants through their official APIs, twice a month, with a score per language and market and the competitors measured under the same conditions. That makes the four checks above possible. It does not make them automatic. When a provider retires a model and the measurement moves to a newer one, that date belongs next to the series, because a step on one assistant on that date is more likely to be the model than your market.

October 2026 will not be the last month like this. Providers publish models on their own timetables and are under no obligation to warn anyone whose figures depend on them. A rise that coincides with a release is not a result. Neither is a fall a failure. Both are a reason to check the four things above before saying anything at the next meeting.