Between November 2024 and May 2026, 166 websites logged 6.77 million visits that started inside an AI tool: someone asked ChatGPT, Claude, Gemini, or a competitor a question, got an answer with a link, and clicked through. Those sessions grew almost tenfold across that window, from roughly 65,000 a month to 644,000, according to a July 2026 analysis by the SEO firm Previsible. ChatGPT alone accounted for 92.4% of the trackable LLM referrals.
Here is the part that should stop anyone who earns money from a website. Most of that traffic never showed up labeled as AI traffic at all. It landed in your reports as “Direct,” sitting in the same bucket as people who typed your URL by hand. You have been looking at your best-converting channel this whole time and calling it something else.
The number the attribution studies keep landing on
A separate 2026 benchmark from the analytics company Loamly measured this directly. Across 446,405 visits in its detection database, 70.6% of AI-driven traffic arrived with no referrer header. Google Analytics 4, and almost every other tool built on the same signals, has nothing to attribute those visits to, so it files them under “Direct.” Only 29.4% carried a referrer that GA4 could recognize as ChatGPT, Claude, Perplexity, or Gemini.
The mechanics are boring and that is exactly why the gap persists. When a link opens from a native app, a server-side render, or certain in-chat browsers, the browser often sends no referrer string. No referrer, no attribution. The traffic is real, the buyers are real, the revenue is real, and the line in your dashboard says “Direct.”
If this were low-value traffic, the miscount would not matter much. It is the opposite. In Loamly’s cohort, AI-referred visitors converted at 10.21% against 2.46% for everyone else, roughly a 4x gap. Measured through Microsoft Clarity across about 1,200 publisher sites, AI-referred visitors signed up at 1.66% versus 0.15% for organic search, an 11x difference. The same sessions showed 27% lower bounce rates and 70% longer time on site. People who arrive after an AI has already vetted you show up warmer, further along, and closer to buying.
So the channel growing fastest is also converting best, and most of it is invisible in the reports you use to decide what to build next.
Why this is a measurement problem before it is a marketing problem
In 20-plus years running IT operations, and more recently through fractional COO engagements, the most expensive mistakes I watched businesses make were rarely bad decisions. They were confident decisions built on a number that was quietly wrong. A team optimizes hard against the metric it can see, and the metric it can see is missing a third of the story.
Play that out for a content site or an affiliate site. You review your analytics at the end of the quarter. One cluster of articles shows flat organic search and a big pile of “Direct” you assume is bookmarks and newsletter clicks. Another cluster shows steady search traffic. Guess which one you keep investing in. Guess which one you quietly stop updating. If that “Direct” pile was actually AI referrals converting at four to eleven times your search rate, you just defunded the pages doing your best work and poured effort into the ones being eroded by AI answers instead.
That erosion is not hypothetical. Content sites are watching AI tools answer informational queries in place of a click, a shift I wrote about in what AI Overviews are doing to publisher traffic. The AI-referral story is the flip side of the same coin. One motion takes clicks away from your top-of-funnel explainers. The other sends warmer, higher-intent visitors to whatever page the model decided to cite. If you cannot tell the two apart, you cannot make either decision well.
The landscape is moving fast, which makes stale reports worse
The platforms themselves are reshuffling underneath all of this. Similarweb’s data has ChatGPT’s share of worldwide generative-AI web traffic sliding from about 76% in mid-2025 to roughly 53% a year later, with Gemini climbing from under 9% to around 27% and Claude going from a rounding error near 2% to close to 9%. The category as a whole reached 9.5 billion monthly visits, up 70% year over year. Previsible’s session data tells a parallel story: Claude grew 64x and passed Perplexity in March 2026, while Microsoft’s Copilot referrals collapsed 96% from their 2025 peak.
Two practical consequences fall out of that churn. First, the referrers you need to watch are not the same ones you would have hard-coded a year ago. A tracking setup that only catches chatgpt.com misses a growing slice of Gemini and Claude. Second, the volume is not evenly spread. Adobe Analytics clocked year-over-year AI-referral growth of roughly 1,700% in travel, 1,200% in retail and e-commerce, and 1,200% in financial services, against about 50% in publishing and media. If you sell products, this channel is arriving at your store faster than almost anywhere else, and it is the channel most likely to be mislabeled.
There is also a crawl-to-click reality worth keeping in view. Digiday reported that Claude crawls roughly 500,000 pages for every referral it sends, ChatGPT about 3,700 to one, and Perplexity around 700 to one. The bots are reading everything and citing a sliver. Getting cited, and then measuring the visits that citation produces, is the whole game.
How to make AI referrals visible in GA4
None of this requires new software. It requires refusing to accept the default “Direct” bucket. Three moves, in order.
Build a custom channel group in GA4. Create a new channel defined by a regex that matches the source or referrer against the hosts that matter now: chatgpt.com, chat.openai.com, claude.ai, perplexity.ai, gemini.google.com, copilot.microsoft.com, and .oai variants. This pulls the 29.4% that does carry a referrer out of “Referral” and “Direct” and into one line you can actually read. It does not recover the referrer-less majority, but it stops you from missing the traffic that is already labeled.
Cross-check against server or CDN logs. Your web server and your CDN see requests that JavaScript analytics never fire on, including a chunk of the referrer-less sessions and the crawler activity behind them. Comparing GA4 sessions against raw log lines for the same pages is the closest thing to ground truth most small operators can get without paying for a dedicated attribution tool. Loamly, Seresa, and a handful of others now sell this as a product, but the log cross-check is free if you have access.
Report AI as its own line, never folded into “Direct.” Once the channel group exists, put it on the dashboard you actually use to decide what to write, refresh, or kill. The point is not a prettier chart. The point is that the next time you are deciding whether a page earns its keep, the AI-referral number is sitting in front of you instead of hiding inside a bucket you have trained yourself to ignore.
What I would do first
If you run a site that earns, spend an hour this week doing one thing: pull your last 90 days of “Direct” traffic and look at which landing pages it hits. Direct traffic, in theory, should skew toward your homepage and pages people bookmark. If instead it is landing deep on specific articles or product pages, especially ones that answer a clear question, you are almost certainly looking at AI referrals wearing a disguise.
That single view tends to reframe the whole conversation. The pages the models cite are the pages worth protecting and expanding, and they are frequently not the ones your search-only reports flagged as winners. It is the same principle behind treating owned assets as the real business rather than chasing platform reach, and it rhymes with the case for owning the last click and the customer as AI tools insert themselves between you and your buyers.
The uncomfortable truth in the 2026 data is not that AI is taking your traffic. It is that AI is sending you some of the best traffic you have ever had, and your analytics are handing you the wrong map to find it. You cannot optimize what you refuse to measure, and right now most sites are refusing to measure the channel that converts four to eleven times better than the one they obsess over. Fix the map first. The strategy gets a lot simpler once you can see where the money is actually coming from.
Sources: Previsible 2026 AI Traffic Report; Loamly State of AI Traffic 2026; Similarweb generative AI statistics; TapClicks on AI referral attribution; Seresa on the GA4 Direct problem.