Six AI assistants logged about 76 billion sessions a month in September. AI referrals still account for less than 1% of traffic to the average website.
Both numbers come from the same report. The interesting question is how they can both be true.
The comparison that stopped working
Search Engine Journal covered the updated AI usage report from Graphite, an SEO agency, built on Similarweb data. Sessions across the six assistants grew 37% year over year. Usage had plateaued in late 2025, then rose 19% between January and September 2026.
The bigger news is what the report dropped. In March, the same agency estimated AI sessions at 56% of search worldwide. This edition says it can "no longer reliably compare" AI with search engines. The Google Android app alone accounts for an estimated 11.8 billion monthly sessions in the US, and nobody outside Google knows how many of those are searches versus translation, image lookups or Discover.
I like this correction. It is rare to see someone retract a headline number in public. It is also a warning about every AI versus search ratio currently sitting in a strategy deck.
Usage is huge, referrals are small
The referral data tells the other half. In the second quarter of 2026, organic search sent 49 times more US traffic than AI referrals. Paid search sent eight times more. AI referrals grew 92% year over year worldwide, but from a very small base.
The split between assistants is worth knowing. ChatGPT drove 60% of AI referral traffic worldwide in July, Gemini 14% and Claude 10%. Claude also posted the largest session growth of any assistant, up 1,901%, while Perplexity was the only one to decline, down 27%.
And 89% of sessions happened on mobile. People talk to these tools on their phones, often inside an app, which is exactly where clean referral tracking goes to die.
There is a second-order effect here that budget owners should notice. When a channel looks small in the dashboard, it gets a small budget. When it gets a small budget, nobody measures it properly, so it keeps looking small. AI is in that loop right now at most companies I talk to.
The loop breaks only when someone decides the dashboard is incomplete rather than accurate. That is a judgment call, not a data point, and it usually has to come from the top.
Where the influence actually lands
Here is the mechanism most referral dashboards miss. Someone asks ChatGPT for the best project management tool for a ten-person agency.
The answer names three brands. The person does not click a link. They open a new tab later, type one brand name into Google, and visit.
In your analytics, that visit is branded organic search or direct. The AI did the persuading and Google gets the credit.
The data supports this. Similarweb's June analysis found that among US desktop users who visited a brand's site after ChatGPT recommended it, 55.9% arrived through branded search. Graphite also cites anecdotal evidence of AI-driven visits being credited to direct traffic.
I explored the same pattern in AI mentions now move real traffic. The newer numbers make it sharper. The referral line in your analytics is the smallest visible part of the effect.
What I would measure instead
Start with branded search volume as a leading indicator. If AI assistants recommend you more often, branded queries should rise before referral traffic does. Track it monthly in Google Search Console and Google Trends, and annotate it against changes in your AI visibility.
Then watch direct traffic to deep pages. Direct visits to your homepage are normal. Direct visits to a specific product or pricing page, from people who never visited before, often mean someone pasted or typed a name they heard elsewhere.
Third, ask. Add a free-text "how did you hear about us" field to demo and contact forms. It is unglamorous and it works. When "ChatGPT told me" starts appearing in that field, you have evidence no referral report will give you.
Fourth, measure the answers themselves. Which prompts mention you, which competitors appear next to you, and which sources the assistant relied on.
That is the input layer. Referrals and branded search are the outputs. GEOflux.ai exists to measure that input layer, because the output metrics arrive too late to steer by.
Check how the ruler was built
The lesson from the Graphite correction is broader than AI. Before any ratio goes into a channel plan, check how the source counted. OpenAI's claim that ChatGPT handles about one fifth as many search-like queries as Google counts something different from Graphite's session data, which counts something different again from your own analytics.
None of these numbers are wrong exactly. They measure different things, and they get quoted as if they measured the same thing.
For a CMO, the practical takeaway is to stop asking "how much traffic does AI send us?" That question has a small, misleading answer. Ask instead how often AI recommends you, and whether branded demand moves when it does.
The assistant rarely sends the visitor. It sends the name. Your measurement has to follow the name.