Open your GA4 dashboard right now and look at Direct traffic for the last 90 days. If it’s up double digits year-over-year and you can’t point to a reason, you’re not imagining things. You’re also probably not looking at what you think you’re looking at.
“Direct” is supposed to mean someone typed your URL from memory or clicked a bookmark. Increasingly, it means someone asked ChatGPT a question, got your site recommended, and clicked through. GA4 has no way to tell the difference. It just dumps both into the same bucket and calls it a day.
We went looking for the best data on how big this problem actually is. Not vendor blog posts asserting a scary number to sell a tool. Real studies, with real methodology behind them. Here’s what they found, and more importantly, what to do about the fact that they don’t agree with each other.
The mechanism nobody disputes
Every credible source traces this back to the same three causes.
Mobile apps are the biggest one. When ChatGPT, Claude, or Perplexity surface a link, and someone taps it on their phone, that click often opens in an in-app browser that never passes a referrer header to your site. Native apps typically strip the referrer before the request leaves the device, a pattern that’s become more pronounced as mobile platforms tighten cross-app tracking.
Copy-paste behavior makes it worse. A lot of people don’t click links inside AI chat responses at all. They copy the URL and paste it into a new tab, which strips referrer data entirely and leaves that visit indistinguishable from someone who typed your URL from memory.
And privacy-first browsers pile on top of that. Safari and Brave strip cross-origin referrer data by default, so a Safari user clicking a Perplexity citation still shows up as Direct traffic even though the click came straight from an AI answer.
None of this is controversial. Where the studies start disagreeing is how much traffic this actually adds up to.
What the numbers actually say
Here’s the spread, and it’s a big one.
Loamly analyzed 446,405 real visits and found 70.6% of AI traffic lands as Direct. Dark AI traffic converted at a 10.21% transactional rate compared to 2.46% for non-AI traffic, a 4.1x gap.
Averi.ai puts the number much lower. 30-50% of AI-driven pipeline hides in Direct traffic at typical B2B SaaS companies, citing Forrester’s estimate that AI-generated traffic overall makes up 2-6% of total organic traffic and is growing more than 40% a month.
Wheelhouse DMG didn’t estimate anything. They measured it directly. Their server logs captured 56 visits from Gemini on iOS in a given window. GA4 recorded only 5 of those as referrals, catching 9% of the actual traffic. That’s not a rounding error. That’s a 91% blind spot on the one AI platform they could verify against ground truth.
Three studies, three different numbers, and none of them are wrong exactly. They’re measuring different things.
Why the disagreement isn’t a red flag
Different sample sizes. Different industries (general web traffic vs. B2B SaaS specifically). Different methods.
This last part is the one worth sitting with. Loamly and Averi are both estimating AI traffic from patterns inside GA4 itself, which means they’re using the flawed tool to measure the flaw in the tool. Wheelhouse compared GA4 against actual server logs, an independent source of truth. That makes their number the most trustworthy of the three, even though it’s also the narrowest, since they could only verify it for Gemini iOS specifically.
So which number is right? Wrong question. The real finding across all three studies isn’t a specific percentage. It’s that every single one of them found the true number is meaningfully bigger than what your dashboard shows. Whether that’s 30% or 70% of your AI traffic hiding in Direct, the dashboard is lying to you either way.
The part that’s actually getting better
This isn’t a permanent, unfixable gap. Some of it is shrinking on its own.
Wheelhouse’s research tracked something interesting: as of April 1, 2026, ChatGPT’s iOS, Android, and macOS apps, along with Perplexity’s iOS app, started passing referrer data correctly where they hadn’t before. The capability to preserve a referrer has existed the whole time. Most platforms just weren’t using it, and some are starting to now.
Even Google’s own systems have tripped over this. In May 2025, Google briefly mislabeled AI Mode search traffic as Direct instead of organic, because AI Mode links were carrying a noreferrer tag that blocked attribution. Google’s John Mueller confirmed it was a bug on their end and it was fixed within days.
Two things follow from that. First, this problem is dynamic, not fixed, and part of it is closing on its own as platforms clean up their referrer behavior. Second, the part that isn’t closing is the hardest part: an AI Overview click and a normal blue-link click both pass google.com as the referrer, with no technical way to tell them apart. That piece stays stuck no matter what anyone does on the GA4 side.
Every study agreed on this number:
Forget the disagreement on volume for a second. Every study that measured conversion rate found the same direction, even if the exact multiplier moved around.
Loamly found dark AI traffic converting at 4.1x the rate of non-AI traffic. Averi found AI-referred visitors converting at 14.2% versus 2.8% for standard Google organic, roughly 5x. The specific multiplier isn’t stable across studies. The fact that it’s always well above 1x is.
That consistency is worth more than any single headline percentage. When three independent studies with three different methodologies all land on “this traffic converts significantly better,” that’s a finding you can act on even without knowing the exact volume.
What to actually do about it
Don’t chase an exact percentage. Nobody has one, and chasing it is a distraction from the two things that actually move the needle.
Set up a GA4 custom channel group with a regex match on known AI referrer domains: chatgpt.com, claude.ai, perplexity.ai, gemini.google.com, and the rest. This takes about 30 minutes and it’s a permanent improvement. It won’t catch everything, since it only recovers traffic that still carries a referrer header. But it’s the floor, not the ceiling.
Then go looking for the traffic that’s still hiding. Pull the pages you know are getting cited in AI responses, cross-reference them against Google Search Console, and flag the ones ranking below position 10 organically. If those pages are getting meaningful Direct traffic anyway, that traffic isn’t coming from people bookmarking a page they can’t find in search. It’s coming from AI.
What this doesn’t solve: Google AI Overview clicks. That piece stays invisible no matter how good your setup is, and any tool that claims otherwise is selling you something.
Here’s the actual point. Your Direct traffic number was never really a measure of brand loyalty, not anymore. Some meaningful chunk of it is AI doing your top-of-funnel work for free, and your dashboard is taking credit for it under the wrong name. Stop reporting it like it’s still 2019.