How much traffic do AI assistants actually send? Three dashboards, three answers

AI assistants send very little traffic, and the published numbers about where that traffic comes from disagree more than anyone admits. BrightEdge put ChatGPT at 95.1% of AI referral traffic on September 24, 2026. Trakkr's panel of 5,777 GA4 properties put it at 86.9% two days earlier. One site's own four-month audit put it at 42%. On mintec.co, GA4 records 388 sessions from AI assistants out of 89,299 over the last 180 days, which is 0.43% of everything, and ChatGPT took 265 of the 373 sessions that came from the four engines we track. Four datasets, four answers, none of them lying. The denominator, the panel and the window are doing all the work.

Four datasets, published within weeks of each other

Read that row by row and the spread is absurd: 95.1%, 86.9%, 71%, 42%. If you were choosing a GEO budget on the strength of one of these numbers, you would pick a different strategy depending on which tab you had open. We have argued before that visibility scores get sold as evidence when they are really samples. Referral shares are the same thing with a percentage sign on it.

What our own property says

We pulled the 180 days ending September 30, 2026 from GA4 and read the source field directly instead of trusting the channel grouping.

• 89,299 total sessions. 388 of them bucketed under the medium ai-assistant. That is 0.43%.
• Of those, 373 came from four engines: ChatGPT 265, Gemini 63, Perplexity 24, Claude 21.
• By month: 12 sessions in June, 56 in July, 95 in August, 224 in September through the 29th.
• Last 30 days by source: ChatGPT 176, Gemini 26, Perplexity 11, Claude 10. That is 223 sessions, about 0.9% of all sessions in the same window.
• ChatGPT's share of the sessions from those four engines: 71% over the full window, 79% over the last 30 days.

The growth is real. September is running at more than double August on a full-month basis, and ChatGPT alone went from 20 sessions in July to 61 in August to 170 in September. Both of the headline vendor numbers are directionally right about growth: Trakkr's index rose 46.1% against the prior 30 days, and BrightEdge reported ChatGPT referrals up 101% between January and August.

A growth percentage hides the arithmetic. Going from 95 sessions to 224 is "+136%", and it is 129 sessions. Our site served 43,116 sessions in August. This channel still reads as noise because the denominator is enormous and the numerator is a rounding error sitting on it. BrightEdge says as much in its own release: AI platforms collectively account for under 1% of total referral traffic in its measurements, with Google above 90% of search.

Where those sessions actually go

Volume is the wrong lens anyway. The page-level data is more interesting than the total.

ChatGPT's top destinations on our site over 180 days: the homepage (16 sessions), /es/contacto/ (13), a chatbot sales post (12), a Meta lifecycle strategy post (11), a Meta CPM post (10), /contact/ (9). Add the two contact pages and 22 sessions, about 8% of everything ChatGPT sent us, landed on a page whose only job is to start a conversation.

Engagement runs far above the site average. Average session duration is 161 seconds for ChatGPT, 228 for Gemini and 579 for Perplexity, against a site average of 22.7. Engagement rate: Perplexity 66.7%, Gemini 42.9%, ChatGPT 39.6%, site-wide 34.8%. Twenty-two sessions is not a channel. It is twenty-two people who arrived having already read something.

One honest gap before anyone asks: our property has zero key events configured, so we cannot claim a conversion rate for this traffic. We said AI traffic converts four times better based on published research in July. Our own numbers cannot confirm or deny that until we fix the event setup, and we would rather say so than publish a made-up multiplier.

Ratios drift the same way shares do. Le Monde's CEO has described ChatGPT's subscriber conversion advantage twice this year: 70 times Facebook and 173 times Google Discover in June, then 26 times and 260 times in a later interview covered by Search Engine Journal. No baseline rates, no volumes, no window behind either pair. Both times the direction was identical and the numbers were not, which is what happens when nobody publishes what the ratio was calculated against.

Why the four numbers disagree

Five things separate those rows, and none of them is a measurement error.

1. Denominator: share of AI referrals is not share of traffic. BrightEdge's 95.1% is a slice of a category that is itself under 1% of referrals. Our 0.43% is a slice of everything. Both are true and they answer different questions.
2. Panel: Trakkr aggregates 5,777 GA4 properties whose owners connected them voluntarily. BrightEdge reads its customer base. 1clickreport read itself. A panel of connected properties over-represents teams already obsessed with AI search.
3. Unit: Trakkr counts pageviews, BrightEdge counts referrals, we counted sessions. ChatGPT sent 176 sessions and 222 pageviews in our last 30 days, so the unit alone moves the number by a quarter.
4. Window: trailing 30 days against January-to-August against 180 days. This channel grew 136% month over month in our data, so windows are not comparable.
5. Grouping: on our property, AI sessions sit under the medium ai-assistant, separate from referral (420 sessions). Any tool that filters on the referral channel reads a different dataset from one that filters on source domain.

How to read any AI traffic dashboard

Before quoting a number from one, get answers to five questions. If a vendor cannot supply all five, the headline is marketing.

The last one is the trap that costs people a week. On our property the referral channel and the ai-assistant medium are different rows, and a report built on one will contradict a report built on the other while both are technically correct.

What we are doing about it

We report AI traffic as an absolute number next to the percentage, per page, and we read it against the pages that make money rather than the aggregate. The tooling side of this is covered in the stack we actually pay for, and the mechanics of what Search Console will and will not tell you are in our piece on AI Overview position counting. The conversion measurement gap is the next thing we fix: key events on the contact form, then a real read on whether AI traffic's reported conversion advantage shows up here too.

The number we watch

We watch neither the share nor the growth rate. We watch the twenty-two sessions ChatGPT sent to a contact page in 180 days. It is small, it is growing, and it is the only figure in this article attached to revenue. The rest is one dashboard disagreeing with another while neither one can see what happens on your contact form.