Why clicks from AI assistants end up as Direct
A click from an assistant becomes Direct when the browser passes no referrer and the link carries no utm tag. That happens on some surfaces by design (links opened with noreferrer, apps that send nothing) and sometimes by change, as in October 2026. In one published sample, 35.7 % of detected AI sessions had a tag and no referrer, so a referrer-only rule misses them.
Last checked · How we check
On this page
Four ways an AI click loses its label
- noreferrer on the link. In Ahrefs' test (May 2025), inline links on paid ChatGPT accounts were opened with
noreferrer, and Gemini's Deep Research used it while "researching websites". A link opened that way passes no referrer at all. - An app is not a website. The same test found that the Perplexity desktop app sent no referrer. A table based on it reports the same for the ChatGPT mobile app and the Copilot Windows app. Apps have no web page to name as the referrer.
- A referrer policy that shortens the referrer. ChatGPT was reported to use
strict-origin-when-cross-origin. That policy cuts the referrer to the origin: the hostname stays, the path goes. On its own it does not turn a visit into Direct, but it hides which conversation the click came from. - Redirects. Clickport reports that a redirect on the way to the page can drop the tag, and that attribution inside the analytics tool then decides the rest.
Which of these applies depends on the assistant and the surface. The referrer list records what each hostname was observed to do, with the date of the test.
What the split looks like in one sample
Clickport published figures for 30 days up to 23 April 2026: 2,619 detected AI sessions, which it puts at 0.7 % of all sessions on the sites it measures. It is a self-selected sample of its own customers, so it shows how the signals combine, not how the web behaves.
| Signals on the AI session | Share of the 2,619 sessions |
|---|---|
| Referrer and a source tag | 41.7 % |
| Tag only, no referrer | 35.7 % |
| Referrer only, no tag | 22.6 % |
Shares by assistant in the same sample: ChatGPT 76.1 %, Perplexity 13.8 %, Gemini 6.5 %, Copilot 1.8 %, Claude 0.5 %. See the published numbers for how this compares with other studies.
What each assistant sent in the last test we know of
These results come from a test in May 2025, more than a year before this page's check date. They describe that day, not today. We have not re-tested yet, and the October 2026 change shows how quickly this moves.
| Assistant and surface | Referrer | Hostname | Note from the test |
|---|---|---|---|
| ChatGPT, web | Partly | chatgpt.com | Free accounts: links worked; paid: inline links with noreferrer; source links had ?utm_source=chatgpt.com |
| ChatGPT, mobile app and Atlas | No (per a secondary table) | (Direct) | Atlas was not checked in the original text |
| Perplexity, web | Yes | perplexity.ai | Desktop app: none; mobile not tested |
| Claude | Yes | claude.ai | All tested links |
| Gemini, web | Mostly | gemini.google.com | Deep Research "researching websites": noreferrer |
| Copilot | Varied | copilot.microsoft.com | Months without a referrer; Windows app: none |
| Grok | No when standalone; yes on X | grok.x.com | Hostname after a third-party table |
| DeepSeek | No, except web-search citations | Not named | |
| Meta AI | Yes | www.meta.ai | |
| Mistral Le Chat | Yes | chat.mistral.ai | |
| Google AI Mode | Was a noreferrer bug; reported fixed | google.com | Counted as google / organic today |
Check your own site
- List the landing pages of your Direct traffic and look for deep pages that no one would type by hand.
- Compare Direct with the AI channel before and after a date you suspect. The October 2026 incident is one such date.
- Look at the share of new users in Direct against your earlier AI sessions; a rise points at relabelled visits.
- Click a link from the assistant yourself and open the browser's developer tools:
document.referrershows what the page received. - Paste what you find into the referrer checker.
Follow-up questions
Is Direct traffic from AI recoverable?
Not after the fact from the referrer: it was never sent. Server logs face the same gap. What helps is a tag on the link where the assistant adds one, and comparing Direct with the AI channel over time.
Why do studies disagree about the AI share?
They count different things: referrer hits only, referrer plus tag, or platform visits. The published numbers page explains the differences.
Does a tag help if the referrer is missing?
Yes, for tools that read utm_source as the source, and no for tools that classify only by referrer. The UTM explainer shows where the ChatGPT tag comes from.
Sources
Read on . Links lead to the original pages.