Why Is 70% of Your AI Traffic Invisible to Analytics?

An analysis of 446,405 website visits that were independently verified as originating from AI assistants found that 70.6% of them arrived with no referrer header. GA4 dutifully filed those sessions under "Direct." A user read your article inside an AI assistant, clicked through to your site, and landed — and your analytics have absolutely no idea how they got there.

That 70.6% isn't a rounding error. It's the norm. And it matters quite a bit, because AI-referred visitors convert at roughly 4x the rate of Google organic traffic. So there's a decent chance the best-performing channel you have right now is one you can barely measure.

How Do We Know This?

The 70.6% figure comes from researchers at Seresa who cross-referenced 446,405 sessions that server-log analysis had independently confirmed as AI-sourced, then checked what GA4 reported for those same sessions. Nearly three quarters landed in Direct. A second analysis looked at what happened after Google launched its native AI Assistant channel in GA4 on May 13, 2026 — the first time Google officially recognised AI-referred traffic as a distinct channel — and found that even with the new channel active, between 35–70% of AI-referred sessions still don't get correctly attributed, depending on the platform and device type.

To understand why, you need to understand how the Referer header actually works.

Why Do Referrer Headers Disappear?

When you click a link in a desktop browser, the browser sends the URL of the page you were on as the HTTP Referer header. That's how GA4 knows you came from a Google results page or another website. So what breaks it?

AI assistants mostly aren't web browsers. ChatGPT on iOS is a native app. Perplexity on Android is a native app. When those apps open a link, the OS spawns a browser tab with no referrer — because there's no "previous webpage" in the conventional sense. The origin is gone before the request is even made. When an AI assistant fetches your page server-side to summarise it before answering a user query, no JavaScript runs at all — no GA4 tag fires, no session cookie is set, nothing. The visit happened; your reporting didn't register it.

This architectural mismatch is why the problem isn't going away. Privacy preferences and native app behaviour both push toward referrer stripping, and AI platforms aren't going to undo that to make your analytics easier.

What Traffic Is Actually Measurable?

There are two different things people tend to conflate when they talk about "AI traffic," and mixing them up leads to measurement strategies that solve neither problem.

Crawlers and indexers — GPTBot, PerplexityBot, ClaudeBot, AppleBot, and a growing list of others — visit your site on a schedule to ingest content for training or real-time index updates. They carry distinctive User-Agent strings. They show up in server logs. They generally respect robots.txt. This traffic is completely measurable if you look at access logs — it just doesn't show up in GA4 because crawlers don't trigger JavaScript events. Running a 30-day log analysis on a typical content site usually reveals AI crawlers accounting for 3–8% of total request volume. The question worth asking: which pages are they reading, how frequently, and which crawlers are most active on your domain?

End-user referral traffic is the separate issue — when an AI assistant cites your page and a user clicks through. That session is what should show up in GA4 as a referral, and mostly doesn't. ChatGPT's web interface has been appending utm_source=chatgpt.com to outbound citation links since June 2025, which is why chatgpt.com now appears as a referral source in many site analytics. Perplexity has a similar citation mechanism for its web UI. But web-interface traffic is a subset of total AI assistant usage; mobile app users generate no referrer data at all.

Of the fraction of AI-referred traffic that does carry a referrer header, the platform split in mid-2026 looks roughly like: ChatGPT leading at around 75% of measurable AI referrals, Google's AI search product at roughly 11–12%, Perplexity at about 7%, Copilot and others sharing the remainder. But remember — these figures represent only the attributable slice. The full volume, including dark traffic, is substantially larger and no one has a precise count.

Does GA4's New AI Channel Actually Fix This?

Google's May 2026 AI Assistant channel addition to the Default Channel Group was a genuinely useful step — the first time GA4 officially acknowledged AI-referred traffic as a category worth tracking separately. But it only captures sessions that carry a referrer header from a platform Google has added to its recognition list. Sessions arriving from a mobile app with no referrer (the majority) stay in Direct. Sessions from AI platforms Google hasn't whitelisted yet: same result. The 35–70% dark traffic estimate applies even after the native channel is active.

The channel is a partial fix to a structural problem. It helps with the easy, web-interface case. The hard case — mobile, non-major platforms, assistant API traffic — is still dark.

What Should Site Owners Actually Do?

Enable server-side logging and parse it for crawler traffic. Client-side analytics have a fundamental blind spot for anything that doesn't execute JavaScript. Your web server or CDN access logs don't have this problem. Setting up a filter for known AI crawler User-Agent strings gives you visibility into indexing frequency, popular paths, and which AI systems are most active on your site — information that's simply unavailable through GA4. If you discover that a particular AI crawler visits your /docs or /blog path every 48 hours, you're getting a directional signal about which content is being indexed for AI answers.

Configure a custom GA4 channel grouping for AI referrers. GA4's native AI Assistant channel is a start, but you'll want a custom grouping that catches referrer domains from platforms not yet in the native list. Perplexity, Copilot, and other AI search products have known referrer domains that can be matched with a regex rule. This doesn't recover dark traffic, but it correctly attributes what is already trackable and stops it from inflating your Direct bucket.

Treat recent Direct growth with scepticism. If your GA4 Direct traffic has been climbing over the past 12–18 months without a clear explanation from brand campaigns or offline media, a substantial fraction of that growth is almost certainly AI-referred visits. The misattribution is largest for mobile-heavy audiences and for sites that are frequently cited in AI assistant responses. If you know your content gets referenced in AI answers — and you can check this roughly by searching for your domain in the major AI chat products — assume your Direct growth is AI-driven until you can rule it out.

Don't confuse crawl activity with citations. Knowing that GPTBot hit your pricing page eleven times last month tells you your content is being indexed. It doesn't tell you whether AI assistants are recommending you to users. For that you need the attribution data from the end-user side — and with 70% of AI-driven clicks dark, that requires layering server-side triangulation on top of whatever UTM data you can capture. Match session fingerprints, timing correlations, and UTM parameters to reconstruct origin paths that the referrer header can't carry. It's messy, but it's the closest thing to a complete picture most site owners can get right now.