Similarweb Says Gemini Is Fading. What Is That Chart Counting?

Similarweb published its August generative AI traffic numbers on September 7, and the headline everyone picked up was that Gemini has slipped. It peaked at 27.8% in May and now sits at 25.6%.

Set the numbers next to the same chart from six months ago and a duller, more useful fact appears. ChatGPT was 56.7% then and is 55.5% now. Gemini was 25.4% then and is 25.6% now. Within a rounding error, the top of the consumer market hasn’t moved since March.

One line on the chart did move. Claude went from 6.0% to 9.3% over those six months, and from 1.9% twelve months ago.

Two charts, two different worlds

Now put that against how businesses spend.

Menlo Ventures surveys enterprise AI decision-makers annually, and its most recent read puts Anthropic at roughly 40% of enterprise spending on large language model interfaces, with OpenAI at about 27%, down from 50% in 2023. When Anthropic first passed OpenAI in Menlo’s mid-2025 update, partner Tim Tully described what the survey found:

“Some might be surprised to see Anthropic overtake OpenAI, given its first mover advantage. But our research puts real numbers behind what we’ve heard anecdotally from the market: Teams are prioritizing real performance in production.”

So the same company is roughly one in eleven of consumer web visits and something close to two in five dollars of enterprise model spend. Both figures are defensible. They measure different populations doing different things, and neither one is “AI market share” even though both get quoted that way.

The caveat that undoes most of the coverage

Similarweb measures website traffic, and nothing else.

The Decoder, which reported the August figures, names the gap plainly: Google pushes a large share of Gemini use through Android, and those interactions never touch a browser. OpenAI’s mobile app and its workplace product sit outside the count for the same reason.

Every assistant in the table is undercounted, and there’s no reason to assume they are undercounted evenly. A number that misses the phone in your buyer’s hand is a number about desktops.

Keep the data and drop the label. The two most confident claims in circulation this week, that Gemini is fading and that Claude is surging, both rest on the surface that captures the least of how people use these tools.

Why a small B2B team should care

The practical damage is in how answer-engine work gets scoped.

We’ve covered what happens when the tracker and the reality disagree before, in the half-life of an AI citation. This is the version that happens before anyone writes a line of copy.

An agency writing a retainer in September reaches for the public chart, sees ChatGPT at 55.5%, and weights the engagement accordingly: most of the effort at ChatGPT, a hedge at Gemini, Claude as an afterthought if it appears at all. That allocation is a reasonable read of a consumer traffic table.

It’s a poor read of a B2B buying committee. If the people who evaluate your product work at companies where Anthropic holds two in five model dollars, the assistant your buyer opens at work isn’t distributed like the assistant they open at home. We have written before about AI referrals converting three times better while being half a percent of traffic; this is the same failure in the other direction, a real signal weighted by the wrong denominator.

The honest position is that nobody has published the number a B2B marketer needs, which is share of buying-related queries by assistant, inside businesses, in your category. Until someone does, the public chart is a proxy and should be labelled as one.

The steelman

The strongest argument against all of this is that the consumer chart is the right chart, because consumer habits pull enterprise ones behind them. People bring the assistant they like into work, and Similarweb is measuring the funnel that eventually decides what a company standardises on.

That argument has real support. Anthropic’s enterprise position was built partly on developers choosing it personally first. The consumer surface is a leading indicator of something.

What it isn’t is a current picture of where your buyer asks about vendors, which is the question the AEO invoice claims to answer. A leading indicator and a scope document are different documents.

What to do with it

1. Stop quoting either chart as market share. Say “share of consumer web visits, Similarweb, August” or “share of enterprise model spend, Menlo survey.” A number without its population is how a scope goes wrong.
2. Run your own prompts across all three. Ten commercial queries in your category, on ChatGPT, Gemini and Claude, logged with which sources each one cites. That’s a two-hour job and it produces the only distribution number that describes your buyers rather than the internet’s.
3. Weight the retainer toward where you’re weakest, not where the crowd is largest. If you show up in ChatGPT and nowhere else, the marginal hour is worth more on the platform that quintupled its share while nobody was optimizing for it.

The chart everyone shared this week is real. It’s just a map of a different country from the one your buyers are standing in.