ChatGPT referrals look like direct traffic: what we tell clients

A client writes on a Monday: "Direct traffic jumped and nobody changed anything. Did ChatGPT stop sending us visitors, or did something break?" That message has been landing in agency inboxes since early October 2026, and the honest answer takes a little care. Both readings can be wrong at the same time.

Here is what we know from a dated source. According to the Promptwatch changelog, ChatGPT stopped sending a referrer on October 8, so visits coming from ChatGPT started showing up as direct traffic. The same entry says ChatGPT acknowledged the bug and that it still adds utm_source=chatgpt.com to its links. We have not seen anything that says the bug is permanently gone, so we treat the situation as live until a client's own data proves otherwise.

Why a referrer loss looks like a traffic drop

Analytics tools decide where a visit came from by reading the referrer the browser passes along. When the referrer is missing and there are no campaign parameters, the visit lands in the direct bucket. Nothing about the visitor changed. The label did.

For a client, that produces two confusing lines on the same chart. The AI referral line sags, the direct line rises, and total sessions look flat. Anyone reading only the AI line will conclude that visibility fell. Anyone reading only the total will conclude that nothing happened. Neither is quite right.

This is also a reminder of something we say in every reporting kickoff: referrer-attributed AI visits are a measured floor, not the full count. People see a recommendation in a chat, close the tab, and type the brand into a browser an hour later. That was always true. The October bug just made the floor lower for one engine, for a period.

The check you can run without any new tool

Before we say anything to a client, we run a manual check in their analytics. It takes about ten minutes and it works in GA4 or any tool that records UTM parameters.

1. Pick a date range that straddles October 8, with the same number of days on each side.
2. Filter sessions by source or by the utm_source value chatgpt.com. If your tool records the full landing page URL, you can also search the landing page query string for utm_source=chatgpt.com.
3. Compare those sessions before and after the date. If the UTM-tagged sessions held steady while the referrer-based ChatGPT line fell, the engine is still sending people. They are just arriving without a referrer.
4. Look at the direct channel by landing page. A rise on deep, specific pages that nobody types by hand (a pricing comparison, a long guide) points to hidden referrals. A rise on the homepage points to something else, such as a campaign or an email.

That last step matters more than it sounds. Direct traffic to the homepage is usually people who know the brand. Direct traffic to a long URL is nearly always someone who clicked a link. If the jump sits on pages we know ChatGPT cites, we are much more comfortable with the explanation.

What we do not do

We do not reassign the direct channel in a client's analytics after the fact. Rewriting history makes later audits painful. We also do not publish an estimate of "hidden ChatGPT traffic" as if it were a measurement. A reasoned range, labelled as an estimate with the method written down, is acceptable. A single confident number is not.

We also do not assume the UTM parameter is a complete fix. It only covers links that carried it, and we cannot promise it appears on every ChatGPT surface. It is a good signal that closes much of the gap, and we say that and no more.

What we tell the client

The message we send is short and has three parts. First, what happened: a dated referrer problem on ChatGPT's side, acknowledged by ChatGPT, that moved some visits into direct. Second, what we checked: UTM-tagged sessions and direct traffic by landing page, with the numbers attached. Third, what changes in the report: ChatGPT visits will be shown with the UTM-attributed sessions included, and the report will say so in the scope block until the referrer behaviour is stable.

We keep the tone flat. A client who has just seen a scary chart does not need reassurance language. They need to see that someone checked the data and that the reporting method is written down.

We also remind them that conversions and revenue live in their own analytics and CRM. If the direct channel gains conversions on the same landing pages that gained sessions, that supports the story. If it does not, we keep investigating.

Where Promptwatch fits

Running that manual check every month for every client is tedious, and it is easy to get inconsistent. Promptwatch is the platform we run client programs on, and its Visitor Analytics has a toggle called "Include ChatGPT UTM referrers", added on October 16 per the same changelog. When it is on, visits that arrive without a referrer are attributed to their UTM source and added to the ChatGPT card and trend line. A real referrer always wins, so a visit is never counted twice.

The practical effect is that the ChatGPT line in the client report stops depending on which day someone remembered to run the GA4 filter. We turn the toggle on for clients whose ChatGPT line dropped after October 8, note it in the scope block, and keep the manual GA4 check as an occasional cross-check against the platform number. If the two disagree by a large margin, that is a reason to look at the tracking script placement, not to pick the number we like.

One boundary worth stating plainly: the toggle fixes attribution inside Promptwatch. It does not change what the client's GA4 or CRM shows. If the client's board reports come from GA4, we still send the explanation and the filter steps so their internal numbers get the same treatment.

A short checklist for your own clients

• Compare UTM-tagged ChatGPT sessions across October 8, not just the referrer line.
• Read the direct channel by landing page before drawing conclusions.
• Write the method into the report scope, with the date of the source.
• Turn on UTM attribution in your visitor analytics if it offers it, and leave the original data untouched.
• Revisit the setup when ChatGPT's referrer behaviour changes again, and expect that it might.

The goal is not to make the ChatGPT line look better. It is to make sure the client's decision about content and budget rests on a chart that still means what it says.