Four assistants out of twelve take more than 90% of the visits

Four assistants take 92.4% of the visits, more than nine in ten. The other three the survey measures take 6.7%, about one in fifteen; the 0.9% missing from a hundred Similarweb does not assign, and five assistants this survey does not measure at all.

The source is Similarweb, April 2026 (the figures from its post of 12 October 2026), which counts visits to the assistants' own sites. That is where we take the order we line them up in, and it is the only source we use for ordering.

• ChatGPT 53.7%
• Gemini 26.7%
• Claude 8.0%
• DeepSeek 4.0%
• Grok 3.2%
• Copilot 2.0%
• Perplexity 1.5%

What to look at is the drop after the first two rows, because from the third down nobody reaches one visit in ten.

Mistral, Meta AI, Qwen, GLM and Kimi have no share, and not because it is low: there isn't one. Similarweb measures 7 of the 12, and those five stay out of the count.

What it means for a local business

If you only come out on Claude you are on 8.0%, fewer than one visit in ten, and the rest of those visits never see you.

If you are missing on ChatGPT you are missing on more than half. On its own it is worth 53.7%, more than one visit in two, more than the other eleven put together.

We put the cut at four ourselves

The first three make 88.4%, just under nine visits in ten: only the fourth takes it past 90%. A rule like "the shortest group above 90%" would give four today as well, but the first three gaining one point six would be enough to bring it down to three on its own. We put the line where we wanted it, and it moves by hand.

Three sources, three numbers

On ChatGPT the three sources say 53.7%, 76.9% and 46.4%: depending on who counts, more than three in four or less than one in two. Thirty points between the lowest and the highest are not noise, because the first counts visits to the sites, the second the clicks those sites send out and the third people across apps and web. They are three different things.

We do not average them. An average would be a fourth number nobody measured, and it would look like the truest one precisely because it would be the only one of ours.

How many brands they look at changes too: Statcounter measures 6, Sensor Tower 3.

What this does not say

These are visits to websites, not people looking for a plumber. None of the three says how many times an assistant was opened to find somewhere to eat tonight: nobody measures that, and the only way we know of getting close is to put the questions to the assistants and read what they answer.

The percentages move every month and these are April 2026 ones, so in six months the row may look different.

The three sources do not agree and we have not made them agree: if you find them averaged into a single number somewhere, that number was measured by nobody. And the five without a share are not worth zero. They are worth a part this survey did not look at.