Does ChatGPT really send most AI referral traffic?

ChatGPT looks dominant in most analytics reports because it is the easiest AI source to identify. It stamps many of its outbound links with a UTM parameter, utm_source=chatgpt.com. Any analytics tool can read that tag and file the visit under ChatGPT.

Other AI surfaces behave differently:

• Google AI Overviews and AI Mode create no referral event of their own. A click from an AI Overview arrives looking like ordinary Google traffic.
• Most other assistants do not add a tag to outbound links. Their clicks land as a plain referral from their domain, or as direct traffic when no referrer is passed.
• ChatGPT itself is not fully tagged. Clicks from the free tier and the mobile app frequently arrive without the tag, which is why some of it shows up as direct traffic.

So a share-of-AI-referrals number measures which platforms label their traffic. It does not measure which platforms people use most. The next move is to stop reading that share as market data and start reading it as a visibility map of your own instrumentation.

The claim that "ChatGPT is 92% of AI traffic" traces to a study reporting 92.4% of trackable referral sessions across a self-selected panel of 166 properties run by one agency. Three limits sit inside that sentence.

1. Trackable only. The denominator is sessions the panel could attribute to an AI source. Untagged clicks fell outside it.
2. Google AI Overviews excluded. They were left out entirely because they produce no referral event. That removes one of the most visible AI surfaces from the comparison.
3. Self-selected panel. 166 sites chosen by one agency are not a sample of the web.

The result is a measure of instrumentation asymmetry as much as usage. It tells you ChatGPT is the most countable AI source on those sites. It cannot tell you ChatGPT sends most of the clicks AI search produces. Anyone repeating the number as a usage share is overreading it.

Every engine leaves a different trail, and some leave almost none. As of October 2026, this is what you can see.

Google states that AI Overviews and AI Mode "may use different models and techniques" and can return different links, so even Google's two surfaces are not one channel. Search Console's generative AI report completed its global rollout around 11 August 2026, but it reports impressions only. It cannot show whether an AI Overview appearance produced a visit. Our guide to AI Overviews and your traffic covers how to estimate that gap.

Your own data is more useful than any panel study, as long as you read it honestly. Work through it in this order.

1. Isolate tagged ChatGPT sessions. Filter on utm_source=chatgpt.com. Treat the count as a floor for ChatGPT traffic, since untagged clicks are missing.
2. Group referrals by AI domain. Build a segment for each assistant's referrer domain. GA4 added a native AI Assistant channel in May 2026, which helps, but check which sources it includes before trusting it.
3. Watch direct traffic to deep pages. A rise in direct visits to pages nobody types by hand, such as a pricing comparison, often hides untagged AI clicks.
4. Read Google separately. Put Search Console's generative AI impressions next to organic clicks. You will not get an AI click count, so do not invent one.
5. Report shares with their caveat. If ChatGPT is 80% of your identifiable AI referrals, say "of identifiable referrals". It is a ceiling on ChatGPT's real share.

Our step-by-step guide to measuring AI referral traffic shows the segment setup in detail. If you need revenue figures, attributing leads to AI search covers what is and is not possible.

Referral data misses more than untagged clicks. Three gaps matter.

Citations without clicks

An engine can name or cite you without sending anyone. A cited source is not always a clicked one, and a retrieved source is not always cited. If you see mentions but no visits, read why AI cites a site but sends little traffic.

Answers without search

ChatGPT search runs at the model's discretion, and OpenAI says it typically rewrites a question into one or more targeted queries. Many answers use no live retrieval at all. Those answers can mention your brand and never produce a link.

Samples versus engine counts

In OtterlyAI's 2,000 AI blogs experiment, Microsoft's AI Performance report credited one site with 42,817 citations, against 3,934 logged by Otterly's prompt-based tracker. The time windows may differ, so it is not an exact ratio. It shows that any prompt sample sees less than the engine's own count.

Stop using ChatGPT's referral share as proof of where your audience is. Use it as one input among several.

• Keep a tagged-ChatGPT segment, an AI-domain segment and a watch on direct traffic to deep pages.
• Check Search Console and Bing Webmaster Tools monthly, each on its own terms.
• Track whether each engine names you for real buyer questions, platform by platform, rather than blending them into one number. Our playbook on measuring AI share of voice walks through it.

Traffic is the end of the chain. Being named in the answer comes first, and the free checker on this site shows whether AI engines name you for up to three buyer questions.

Why does ChatGPT traffic sometimes show up as direct? Not every ChatGPT click carries the utm_source=chatgpt.com tag. Clicks from the free tier and the mobile app frequently arrive untagged. When an app also passes no referrer, analytics has nothing to attribute the visit to and files it as direct. That means tagged ChatGPT sessions understate ChatGPT's real traffic, while direct traffic quietly absorbs some of it.

Can I see clicks from Google AI Overviews in Google Analytics? Not as a separate source. AI Overviews produce no referral event of their own, so clicks blend into Google organic traffic. Search Console's generative AI report shows impressions for AI Overviews and AI Mode, with no clicks, click-through rate or position. You can compare those impressions against organic clicks over time, but you cannot isolate an AI Overview click count.

Is the 92% ChatGPT figure wrong? The number may be accurate for what it measured: 92.4% of trackable referral sessions on a self-selected panel of 166 properties. The error is in how it gets repeated. It excluded Google AI Overviews entirely and counted only traffic that could be identified. It describes which AI source is easiest to track, not which one people use or which sends the most clicks.

Should I add my own UTM tags to links for AI traffic? You cannot tag links an AI engine generates, because the engine writes the URL. You can tag links you control, such as links in your own listings or campaigns, which keeps your data clean. For AI engines themselves, rely on their own tagging, referrer domains and first-party reports, and accept that some AI clicks will stay unattributed.