How to measure AI search ROI when AI traffic hides in "Direct"

Every marketing team now gets the same question from the top: "Is AI search bringing us anything?" And most teams answer it from the wrong report.

Open GA4, filter the referral list for chatgpt.com or perplexity.ai, and you will usually find a small number. It looks like a rounding error, and it is easy to conclude the channel doesn't matter yet. The number is real. It is just incomplete — a meaningful part of the traffic AI sends you never arrives labelled as AI. One analysis estimates that 20–40% of AI-referred visits are recorded as "Direct" (Seresa). Treat that range as an estimate from one source, not a law. The direction, though, matches what anyone who has tried to track this has seen.

This article is a practical setup for measuring AI search honestly, in three layers: the clicks you can see, the clicks you can't see, and the influence that never produces a click at all.

Why the AI number in your analytics is too low

There are four common reasons an AI-driven visit loses its source on the way to your site:

• Apps and in-app browsers. When someone taps a link inside a mobile AI app, the page may open in an embedded browser that passes no referrer. GA4 files the session under Direct.
• Copy and paste. Users often copy a brand name or URL out of an answer and paste it into a new tab. That visit is Direct, or — if they search the name first — branded organic search.
• Privacy tools. Ad blockers, strict browser settings and some corporate networks strip referrer data.
• The "remember and come back" journey. A buyer hears your name in ChatGPT on Monday and visits on Thursday by typing it into Google. Analytics credits Google.

None of this means AI traffic is huge and hidden. For most sites, AI referrals are still a small slice of total traffic — and the mix is shifting fast: one report found ChatGPT's share of B2B AI referrals fell from 89% to 63% in eight months as Claude and Gemini grew (Goodie). It means the visible number understates the channel, and the understatement is uneven — it hits mobile and high-intent "remember the name" journeys hardest.

Layer 1: capture every AI click you can see

Most sites don't even get the visible part right, because AI sources are scattered across Referral and sometimes Organic. Fix that first.

Create an "AI assistants" channel in GA4

In GA4, go to Admin → Data display → Channel groups, copy the default group, and add a new channel called AI Assistants with the condition Session source matches regex:

chatgpt\.com|chat\.openai\.com|openai|perplexity|gemini\.google|bard\.google|copilot\.microsoft|copilot\.com|claude\.ai|deepseek|you\.com|meta\.ai|grok|kimi|doubao

Then drag the new channel above Referral in the list. Channel rules are evaluated in order, so if Referral comes first it will swallow these sessions.

Two details that matter:

• ChatGPT often appends utm_source=chatgpt.com to the links it cites. That is helpful, but it also means the same assistant can show up as both a referrer and a UTM source. The regex above catches both, because GA4 uses the UTM value as the session source when it is present.
• Keep the list maintained. New assistants appear and domains change; review it quarterly.

Compare it against the right baseline

Look at conversion rate, not just sessions. Published comparisons of AI versus organic conversion vary enormously — from a modest lift in large ecommerce datasets to claims of many times higher — depending on industry, sample and how "conversion" is defined (AirOps summarises several). Your own number, on your own funnel, is the only one worth reporting.

Layer 2: estimate the clicks you can't see

You can't recover a lost referrer, but you can measure the shadow it leaves. Three signals are cheap to set up.

1. Ask people

Add an "AI assistant (ChatGPT, Perplexity, etc.)" option to the "How did you hear about us?" field on your demo, contact or checkout form. Self-reported attribution is imperfect, but it is the only signal that catches the "heard it on Monday, visited on Thursday" journey. After a quarter, a line like "9 of 37 new customers this month said AI" is more persuasive to a CEO than any traffic chart.

2. Watch Direct traffic to deep pages

Nobody types a 60-character URL from memory. If Direct sessions landing on specific articles or product pages grow while your other channels are flat — especially pages you know AI cites — a share of that is likely AI traffic that lost its referrer. Track Direct entrances to non-home pages as a separate trend line.

3. Watch branded search

In Google Search Console and Bing Webmaster Tools, track impressions for queries containing your brand name. When AI starts recommending you, more people search for you by name. It's a lagging, noisy signal, but it is independent of referrer data.

Layer 3: measure influence that never produces a click

Here is the part traffic reports miss entirely. A buyer can read an AI answer that recommends you, compare you to two competitors, and decide — without clicking anything. That is still the channel working. So the core GEO metrics are about the answer itself, not the visit.

Visibility in the answers

Take the 20–100 questions your buyers actually ask ("best X for Y", "X vs Y", "is X worth it"), and run each one several times on each platform — answers vary from run to run, so one screenshot proves nothing. Then record:

• Recommendation rate — the share of answers that name you.
• Average position — where you appear when you're named.
• Share of voice — your mentions divided by all brand mentions in the set.
• Cited sources — which websites the AI used, so you know where to act.

Report these per platform. Averaging ChatGPT, Perplexity, Gemini and Grok into one "AI score" hides the fact that they read different sources and often disagree.

AI fetches in your server logs

Some AI crawlers visit in real time because a user just asked a question. OpenAI's ChatGPT-User, Perplexity's Perplexity-User and Anthropic's Claude-User are documented as user-triggered fetchers, distinct from training crawlers like GPTBot or ClaudeBot. A rising count of user-triggered fetches on a page is one of the closest things you can get to seeing "someone asked an AI a question and it read my page to answer".

Your analytics tool can't see these — they don't run JavaScript. You need server or edge logs grouped by user agent. On our own site we log every request at the edge and split crawlers into training, AI search indexing, user-triggered fetches and classic search engines; the user-triggered line is the one we watch. User agents can be spoofed, so if a number looks surprising, verify it against the IP ranges the AI companies publish.

Putting it together: an honest AI ROI report

A report that will survive questions from a CFO has four lines, reviewed quarterly rather than monthly (AI answers are noisy week to week):

Present the first line as the proven minimum and the second as the likely range. Use the last two to show direction. Don't multiply visible AI traffic by an assumed "hidden" factor and present it as fact — it's the fastest way to lose credibility on a channel that already attracts too much hype.

A 30-minute checklist

• Create the GA4 "AI Assistants" channel and move it above Referral.
• Add an AI option to "How did you hear about us?" on every lead and checkout form.
• Start a Direct-to-deep-pages trend line and a branded-search trend line.
• Pick 20 buyer questions and record a baseline on ChatGPT, Perplexity, Gemini, Grok and Google AI Overviews — several runs each.
• Make sure your logs can show AI user agents, and that robots.txt isn't blocking the ones you want (OAI-SearchBot and PerplexityBot govern whether you can appear in those products' search features at all).

The goal isn't a bigger number. It's a number you can defend — and a baseline, so that in three months you can show whether anything changed.