ChatGPT Desktop and Claude Cowork send no referral data. That was the headline Ross Hudgens of Siege Media posted on LinkedIn on 7 November 2026, from his conversation with SparkToro's Rand Fishkin, who argued that a Direct share climbing from 40% of a site's traffic towards 70% or 80% is the AI platforms at work, not chance.
The claim is right about the referrer and incomplete about the damage. We build LovedByAI, an AI search optimization platform for small businesses that logs AI referral visits on the sites it optimizes, so we can show which AI visits actually go dark, which ones stay labelled, and how a small business can find the dark ones in GA4 this week.
Why ChatGPT Desktop and Claude Cowork visits show up as Direct traffic
A browser fills in the referrer from the page you clicked on. A desktop app is not a page. When you click a link in the ChatGPT or Claude desktop app, the app hands the URL to your operating system, which opens it in your default browser the same way it would open a link from any other program. The browser starts that visit with no page behind it, so there is no referrer to send.
Google Analytics 4 then has two places left to look: a utm_source tag on the link, or nothing. Google defines its Direct channel as visits from a saved link or a typed URL, and in practice that is where every visit with neither a referrer nor a tag lands.
| How the visit arrives | Referrer | Tag | Where GA4 puts it |
|---|---|---|---|
Link with a utm_source tag, from any app or browser | Any | Yes | Labelled by the tag (for ChatGPT, the AI Assistant channel) |
| Link with no tag, clicked in a browser tab that shares its referrer | Yes | No | Referral or AI Assistant |
| Link with no tag, opened from a desktop or mobile app | No | No | Direct |
| Link with no tag, copied and pasted later | No | No | Direct |
So the useful question is not "does the app send a referrer", because app clicks rarely carry one. It is "does the link carry a tag".
Is ChatGPT app traffic really dark? The tag carries it, not the referrer
For ChatGPT, mostly not. OpenAI's publisher FAQ says ChatGPT adds utm_source=chatgpt.com to the links it shows. A tag is part of the URL itself, so it survives the trip out of an app where a referrer does not.
Our data shows how little the referrer matters already. From 15 to 21 October 2026, fewer than one in thirty tagged ChatGPT visits to the small business sites LovedByAI tracks also carried a chatgpt.com referrer. If GA4 relied on the referrer, ChatGPT would look almost absent. It doesn't, because of the tag.
The tag is also the weak point. When GPT-6 rolled out, many ChatGPT links lost it, and the typical site lost 60% of the visits it could identify as ChatGPT over 9 to 14 October 2026, compared with 1 to 8 October. Those visits were counted as Direct and stay Direct. We covered that week in where ChatGPT traffic goes in GA4.
Why Claude traffic shows as Direct in Google Analytics
Claude is the real dark spot. Anthropic does not document any tracking tag on the links Claude gives, and nothing in the visit names Claude when it comes from the desktop app or Cowork.
The web app does not help much either. When we requested claude.ai on 9 November 2026, its servers answered with a Referrer-Policy: same-origin header, which tells the browser to share a referrer only with claude.ai's own pages. Unless a link overrides it, another site gets nothing.
The same is true, to varying degrees, of every app that opens links in a browser: Slack, WhatsApp, Outlook and the LinkedIn and Instagram apps all produce Direct visits. That matters for the next step, because AI apps are not the only thing filling your Direct line.
How to find ChatGPT and Claude visits hiding in Direct traffic in GA4
You can't label an untagged visit after the fact. You can find where AI visitors go, then watch Direct on those same pages. Four steps, about twenty minutes in GA4.
1. List the pages AI already sends you to. In Reports → Engagement → Landing page, add a filter on Session source that matches the regex below, set the range to the last 90 days and export the list. These are the pages assistants recommend when they do leave a label.
chatgpt\.com|openai\.com|perplexity\.ai|claude\.ai|gemini\.google\.com|copilot\.microsoft\.com
2. Build a Direct exploration. In Explore → Free form, create a session segment where Session default channel group exactly matches Direct. Use Landing page + query string as rows and Sessions, Engaged sessions and Key events as values. Add a filter so Landing page + query string does not exactly match /.
3. Compare matched windows. Set the exploration to compare the last 28 days with the 28 days before. Split the rows into two groups: your AI pages from step 1, and every other deep page. If Direct on your AI pages is growing faster than Direct on the rest, and your tagged AI visits are growing too, the difference is your best estimate of untagged AI app traffic.
4. Check the fingerprints on the pages that stand out. AI app visits tend to share four traits:
- A deep URL nobody types. A long blog slug or a specific service page, not the homepage.
- New users. Add New / returning: a bookmark brings back a returning visitor; an AI answer brings someone new.
- Desktop operating systems. ChatGPT Desktop and Claude Cowork run on macOS and Windows, so add Device category and Operating system.
- Engaged sessions. People arriving from an answer came for that page, so check that the Direct rise is engaged visits, not one-second hits.
None of these proves a single visit came from AI. Together, on pages assistants already recommend, they turn "Direct went up" into "Direct went up on the pages ChatGPT sends people to".
Does AI traffic only land on the homepage? What small business data shows
Ross Hudgens tried a version of this method, built on long-tail landing pages, and gave it up: in his words, "almost all AI traffic goes to the homepage, besides data-driven content types." On the sites his team looked at, deep pages carried too little AI traffic to be a signal.
Small business sites look different. In LovedByAI's data on ChatGPT visits from 14 September to 13 October 2026, the homepage drew 42% of ChatGPT's visitors on the average service business site, and most ChatGPT visitors started somewhere other than the homepage on 109 of 175 service business sites and 129 of 148 online stores. Product pages, service pages and blog posts did the rest. The full breakdown is in where ChatGPT referral traffic lands.
That is good news for the recipe above: if your AI visitors mostly land on inner pages, the deep-page signal is there to find. If your own step 1 list is almost all homepage, accept that those visits are indistinguishable from typed ones and skip to the next section.
What you can't measure about AI app traffic, and what to stop worrying about
Some of this is permanently out of reach. Be honest about it in your reports:
- Homepage visits from AI apps. They look exactly like someone typing your address.
- Google's AI Overviews and AI Mode. Google counts those clicks as Organic Search, not Direct, so no Direct analysis will find them.
- Answers that mention you without a click. GA4 counts visits only. On the sites LovedByAI monitors, the median site saw 55 AI assistant page fetches for every AI referral visit over the 90 days to 23 October 2026.
- An exact AI share of Direct. Any tool claiming one is estimating, and should say how.
And three things to stop worrying about:
- A rising Direct line is not lost traffic. The visitors are real. GA4 just can't name the source.
- "How did you hear about us?" is not a fix on its own. Rand Fishkin told Siege Media he has soured on those surveys after an experiment in which customers credited an influencer who had never mentioned the brand. Use answers as a hint, never as a count.
- You don't need a perfect channel group. A custom AI channel catches the visits that carry a label. The dark ones are found by pages and trends, not by a better regex.
What you can do is make more of your visits carry a name. Tag every link you control that AI assistants tend to quote: your Google Business Profile, directory listings, partner pages and PDFs, as in our guide to UTM tags for AI traffic. Add a GA4 annotation for 8 to 15 October 2026 so nobody reads that week's Direct spike as a campaign win.
How LovedByAI counts AI visits that GA4 can't name
GA4 depends on what the visitor's browser tells it. LovedByAI also reads the server side: the AI assistants fetching your pages and the AI referral visits that follow, page by page. That doesn't light up the dark visits either, but it tells you which pages AI is reading and recommending, which is exactly the list step 1 of the recipe needs, and the fetch side keeps counting on the days a tag goes missing.

