AI Traffic ROI Has a Denominator Problem. The Referrer Strip Is the Other Half.

Duane Forrester argued this week that AI search ROI looks bad because measurement compares a stable AI share against a shrinking organic base. He's right, but only half right. The bigger issue: 60 to 70 percent of ChatGPT-referred traffic lands in GA4's Direct bucket because the referrer header is stripped before it reaches your property. AI traffic isn't 1.08 percent of the web. Half of it is just labeled wrong.

The denominator argument, briefly

Forrester's Search Engine Journal piece frames the issue as a math problem most teams skip. AI's percentage of total traffic looks flat. But the total has been falling fast. Citing Similarweb, he points out organic traffic to news publishers dropped from about 2.3 billion monthly visits in mid-2024 to under 1.7 billion by May 2025. That's a 26 percent slide in roughly a year. Zero-click searches climbed from 56 percent to 69 percent over the same stretch. A Reuters Institute survey of media leaders puts the expected three-year decline at another 43 percent on average.

So when teams say "AI is 1 percent of traffic, why bother," they're measuring against a number that's already shrinking. Same share, smaller pie, and the pie keeps getting smaller. Forrester's framing of citations as "grounding artifacts" rather than traffic mechanisms is a useful reframe. LLMs answer the question. The citation is a courtesy, not a clickable funnel.

I think this is correct, but it's the easy half of the argument. The hard half is that even the 1 percent number is wrong, because most AI traffic doesn't get attributed at all.

Where the AI traffic actually goes

The Conductor 2026 AEO/GEO Benchmarks study analyzed 13,770 domains across 10 industries between May and September 2025. The headline number: AI referral traffic averages 1.08 percent of all website traffic. ChatGPT alone accounts for 87.4 percent of that. The number is real, but the methodology relies on traceable referrals, and for AI platforms the referrer header is mostly stripped.

The mechanics are simple, and most analytics teams haven't audited them:

• Free ChatGPT strips referrer headers on inline links in chat responses. Paid Plus and Team subscribers keep the referrer on the same kinds of links. So you get attribution from your premium audience and nothing from the much larger free user base.
• Mobile AI apps strip referrers by default. The session opens in an in-app browser or hands off to Safari or Chrome with no referrer set.
• HTTPS redirect chains lose referrers in transit, especially through link shorteners and tracking middlemen.
• Privacy-first browsers (Brave, Firefox with strict tracking protection) strip referrers as policy.
• Strict referrer policies in IT-managed environments affect another 5 to 10 percent of B2B traffic.

The estimates vary, but the consensus across the Clickport GA4 audit, MarTech's analysis of Comet and Atlas referrals, and Traffic Masters' breakdown lands in the same range. Roughly 60 to 70 percent of ChatGPT-referred traffic shows up as "Direct" in GA4. Roughly 70 percent of all AI referrals arrive with no referrer header, full stop.

When ChatGPT referrals appear to fall in your reports, the spike in Direct traffic is usually the same traffic with the source label scrubbed off.

What the math looks like once you fix it

If Conductor's 1.08 percent represents only the roughly 30 percent of AI traffic that carries a referrer, the real share is closer to 3 to 3.5 percent of total traffic. Still small. But on a shrinking organic base where AI Overviews already skip the browsy half of the SERP, it compounds quickly.

This is where Forrester's denominator argument bites harder than he framed it. AI is taking a bigger share of a smaller pie, and most teams aren't capturing that share correctly. The line between "AI doesn't drive traffic" and "AI is the new search front door" is hidden in the gap between your referrer report and your Direct bucket.

The pattern shows up everywhere once you know to look for it. Direct traffic growing year-over-year while organic search drops by double digits. No UTM strategy on AI-cited pages. No custom channel group set up in GA4. The Direct line is doing more work than most teams realize, and the infrastructure to actually attribute AI traffic was never built in the first place.

The 30-minute audit

This is the part worth doing this week, and it's not complicated. Three steps.

1. Build a custom channel group for AI in GA4. Add the known AI source domains as a custom group: chatgpt.com, chat.openai.com, openai.com, claude.ai, perplexity.ai, gemini.google.com, copilot.microsoft.com, you.com, kagi.com. This catches the roughly 30 percent of traffic that does carry a referrer and gives you a clean trend line. Most teams skip this because it's not in the default GA4 setup, which is exactly why it's a quick win.

2. UTM-tag your AI-targeted content. Anything you've optimized for AI Overview citations or ChatGPT browse, things like pillar guides, comparison pages, and glossary entries, should have a canonical version with a UTM parameter set as the recommended share URL in your meta tags. When AI tools surface the URL, the UTM survives even when the referrer doesn't. From what I've seen, this is the only attribution method that actually works for free ChatGPT users.

3. Watch brand search lift as the real proxy. Forrester is right that citations are grounding artifacts, not traffic mechanisms. The actual signal is people seeing your name in an AI answer and searching for it later. Brand search volume already beat backlinks as the top AI citation predictor. Watch the 28-day brand search trend in Search Console against your AI citation count from a tool like Profound or Otterly. That's the closest thing to a real ROI line you'll get on AI traffic until referrer policies change.

For benchmarks: a healthy account should see brand search volume growing 5 to 15 percent month-over-month if AI citations are working, even when the AI referral line in GA4 looks flat. The two should diverge in the same direction. If they don't, the AI optimization isn't working, or the brand isn't getting cited yet.

What I'd stop measuring entirely

Last-click attribution on AI traffic is fiction. Same with "AI as a percentage of total traffic" benchmarks pulled from public reports. The methodologies almost all rely on referrer headers, so they're systematically undercounting by the same 60 to 70 percent factor. When Search Engine Land covers Conductor's 1 percent number, it's a referrer-visible floor, not a real share.

The thing I'd actually track, beyond brand search and the custom channel group, is the ratio of Direct traffic to organic search over time. If Direct is growing faster than organic is shrinking, you're probably watching AI referrals get filed as Direct. It's not perfect, but it's the cheapest sniff test that runs in 10 minutes.

Anyway, Forrester's denominator argument is the headline. The referrer strip is the part nobody's writing about, and it changes the size of the conversation by roughly three times. AI is probably bigger than the public numbers suggest. Most of the gap is a custom channel group and an afternoon of cleanup.

Notice Me Senpai Editorial