The figures
Similarweb published its August 2026 snapshot of generative AI website traffic share on 7 September. ChatGPT takes 55.5% of visits across the tracked assistants, Google’s Gemini 25.6% and Anthropic’s Claude 9.3%. DeepSeek is on 3.4%, Grok 2.4%, Microsoft’s Copilot 1.6% and Perplexity 0.9%, as reported by The Decoder from the same data.
Against three months earlier, ChatGPT is up from 52.7%. Against twelve months earlier, it is down from 73.3%. Both are true, and they are not the same story.
What actually moved this year
Similarweb’s own posting of the series puts the year-ago distribution at ChatGPT 73.3%, Gemini 12.9%, DeepSeek 4.0%, Grok 2.6%, Copilot 2.0%, Perplexity 2.0% and Claude 1.9%. Six months ago the same measure had ChatGPT at 56.7%, Gemini at 25.4% and Claude at 6.0%.
Read across those three points and the shape is clear. ChatGPT lost roughly eighteen points of share in a year, almost all of it in the first half of that period, and has been broadly flat since — recovering a couple of points over the summer rather than resuming a decline. Gemini roughly doubled its share year on year and has since stopped growing on this measure. Claude is the fastest riser in proportional terms, from 1.9% to 9.3%, off the smallest base of the three.
What this measure does not count
The caveat matters more than usual here. This is website traffic. It does not count mobile app usage, which is where a large share of consumer chatbot interaction happens, and it does not count assistant surfaces embedded in other products.
That cuts against Google most directly: Gemini reaches users through Android and through Google’s own products in ways a visit to a chatbot website never registers. It also understates OpenAI, whose desktop and work applications sit outside the browser. A share of web visits is a real measurement of a real thing, but it is not a measurement of how many people use these assistants.
Why the number is watched anyway
Web traffic share has become the default public proxy for consumer position in this market because the companies themselves disclose so little. OpenAI, Google and Anthropic do not publish comparable monthly active user figures on any regular schedule, so a third-party panel estimate fills the gap.
It fills it imperfectly. Panel-based estimates carry sampling error, they change when a provider changes how its site is structured, and they are revised. Treating a 2.8-point move over three months as a trend is reading more into the instrument than it can carry. Treating an eighteen-point move over twelve months as noise would be the opposite mistake.
What to watch
Whether Gemini’s flat summer holds through the autumn, and whether Claude’s rise continues from a base that is no longer small. The more interesting question is whether anyone starts publishing usage figures that do not have to be inferred from browser panels — which, for now, nobody has.
