Your AI Traffic Numbers Are Wrong: The GA4 Measurement Gap

When Google added a native AI Assistant channel to GA4 in May 2026, it looked like the measurement problem for AI traffic had finally been solved. After eighteen months of stitching together custom regex rules and squinting at referral reports, the industry got what it asked for: a default channel that tags sessions from ChatGPT, Gemini, and other AI assistants automatically.

The reality is messier. The AI Assistant channel undercounts AI traffic for almost every property it touches, and the gap isn't subtle. A single source like chatgpt.com can land in three different channels simultaneously — AI Assistant, Referral, and Unassigned — depending on factors most analysts never see. If you're reporting AI traffic numbers to a client or a board, those numbers are almost certainly low.

What GA4 Actually Did

On May 13, 2026, Google added the AI Assistant channel to GA4's Default Channel Group. The mechanics are straightforward in principle. When GA4 identifies a referrer as an AI assistant — ChatGPT, Gemini, Copilot, DeepSeek, Grok — it tags the session with the medium `ai-assistant`, assigns it to the AI Assistant channel, and stamps the campaign as `(ai-assistant)`. No configuration required.

The rollout was gradual. Most properties didn't see the channel until early June 2026. Anything before that — all those ChatGPT referrals from the first five months of the year — sits in Referral or Direct, untagged and effectively invisible in historical comparisons.

Google's list of recognized AI platforms has also shifted since launch. The initial set named ChatGPT, Gemini, and Claude. By June, Claude had been quietly dropped and replaced with DeepSeek, Copilot, and Grok. Perplexity — one of the most significant AI referral sources — still isn't on the list. It lands in Referral by default.

The Fragmentation Problem

Here's where the measurement breaks. GA4 determines the channel using both source and medium, not source alone. When you break chatgpt.com down by session source/medium, it splits into three distinct rows:

chatgpt.com / ai-assistant lands in the AI Assistant channel. This is the slice GA4 recognized and tagged correctly. It's the number most people read when they check their AI traffic.

chatgpt.com / referral lands in Referral. These are sessions that arrived before the channel rollout reached your property, plus sessions GA4 failed to tag for other reasons. If your property switched on in late May or early June, an entire quarter of ChatGPT traffic is sitting here.

chatgpt.com / (not set) lands in Unassigned — the channel almost nobody opens. The medium is missing entirely, which means no channel rule can catch it. The most common cause is the ChatGPT mobile app and its embedded browser, which strip referrer headers while preserving the source. Links opened inside the app's in-app browser arrive with a source but no medium, and GA4's response is to shrug.

Three channels. One source. And the number you're reporting — the AI Assistant channel read on its own — captures roughly one-third to one-half of your actual AI referral traffic.

Why Standard Fixes Fall Short

The instinctive responses to this problem each have a flaw.

Reading the AI Assistant channel alone misses the Referral and Unassigned fragments. You'll undercount every time, and the gap is larger for properties that had significant AI traffic before June.

Comparing month-over-month breaks because the native channel only counts forward from its rollout date. Any range that crosses the rollout boundary is comparing tagged traffic against untagged traffic. The trend line you see isn't a trend — it's a measurement artifact.

Checking rank tracking answers a different question. Rankings tell you whether an AI assistant mentions your brand or links to your domain. They don't tell you whether anyone actually clicked through.

Building a Custom Channel That Works

The fix is a custom channel group that matches on source while ignoring medium entirely. This collapses the `ai-assistant`, `referral`, and `(not set)` versions of each AI platform into a single line.

In GA4, navigate to Admin → Data Streams → Configure traffic → Channel groups, and create a new custom channel group. Add rules that match the source against known AI platforms:

• `chatgpt.com`
• `gemini.google.com`
• `perplexity.ai`
• `claude.ai`
• `copilot.microsoft.com`
• `grok.com`
• `deepseek.com`

Set the medium to match anything — or simply leave it unspecified. The rule will capture every session from that source regardless of how GA4 tagged it.

Two things happen when you do this. First, the fragmented sessions collapse into a single channel, giving you an accurate AI traffic number for the first time. Second — and this matters more than people realize — the custom channel applies its rules retroactively across your entire date range. All those old ChatGPT sessions sitting in Referral from January through May get reclassified. You get clean historical data going back as far as your property has been collecting.

Include Perplexity in your custom rules even though Google doesn't. It's one of the highest-intent AI referral sources, and it converts well above its weight. Leaving it in Referral means undercounting by a meaningful margin.

What the Numbers Actually Look Like

Similarweb's clickstream data has ChatGPT referrals converting at roughly 7%, ahead of organic search and not far behind paid search. Invoca's 2026 Lead Conversion Benchmark Report found that calls referred by ChatGPT have a 49% lead qualification rate — about 10 percentage points higher than the average across all tracked channels.

That's the context for why measurement matters. AI traffic is still a small slice of overall sessions for most sites. But it punches well above its weight in conversion quality. A channel with that profile deserves to be measured accurately, not eyeballed through a broken default channel.

The typical pattern we've seen after implementing the custom channel fix: real AI referral traffic is 2 to 3 times higher than what the AI Assistant channel reports on its own. The gap is largest for properties with significant mobile app traffic, where the `(not set)` medium problem is most pronounced.

Beyond GA4: The Attribution Layer GA4 Can't See

Even a perfect custom channel doesn't capture everything. When someone asks ChatGPT for a restaurant recommendation and then searches for the restaurant name on Google two hours later, that's an AI-driven visit that shows up as branded organic search. When someone researches a product in Perplexity and types the URL directly into their browser, that's an AI-driven visit that shows up as Direct.

Similarweb documented this pattern in June 2026: brands recommended by ChatGPT saw 2.5x higher visit probability within seven days, but most of that traffic appeared as branded search rather than direct referral. The digital trail from AI research to website visit is getting longer and more convoluted, and standard analytics tools are built for a world where the last click tells the whole story.

Server-side tracking, first-party data collection, and brand-lift studies all help close this gap. But the first step — the one most organizations haven't taken yet — is simply fixing the channel group so the traffic you can see is counted correctly.

A Practical Audit

If you're not sure whether this affects you, run this check in your GA4 property today.

Add Session source/medium as a secondary dimension to your Acquisition report. Filter for chatgpt.com. If you see more than one row — if chatgpt.com appears with different mediums — you're affected. Count the sessions in each row and compare the total to what the AI Assistant channel reports. The difference is your measurement gap.

Then build the custom channel group, wait a few minutes for it to process, and check again. The number will change. That's your real AI traffic.

The gap between those two numbers — the default channel and the custom channel — is the gap between what you think your AI search visibility is worth and what it's actually worth. For most organizations, that gap is significant enough to change strategy. It's hard to justify investment in GEO and AI visibility optimization when your analytics tell you the channel is smaller than it actually is.

Measurement isn't a footnote to strategy. It's the foundation. And right now, for AI traffic, that foundation has a crack in it that's easy to fix and expensive to ignore.