Is ChatGPT traffic hiding in your "direct" numbers? A GA4 check

Open the Traffic acquisition report in GA4 and Direct is probably one of your biggest channels. Some of it is genuine: people typing your address, using bookmarks or coming back to buy again. But some of it is ChatGPT traffic Google Analytics could not attribute, because the visit arrived with no referrer and no campaign tag. In early October 2026 that gap briefly became visible, when ChatGPT links lost their tracking tag and direct visits rose on the same pages.

This guide is for UK marketing managers, ecommerce owners and in-house analysts. It walks through a GA4 check you can finish in about 30 minutes, then shows how to estimate how much of Direct may be AI and how to report it honestly.

Quick Answer: Is ChatGPT traffic hiding in your “direct” numbers?

Probably some of it, but you can’t prove exactly how much. ChatGPT adds utm_source=chatgpt.com to many links it cites, and GA4 now has an AI Assistant channel, yet visits from apps, copied links and stripped referrers still land in Direct. Measure tagged AI visits first, then look for direct sessions from new users landing on deep pages, and report those as an estimated range.

Key Takeaways

• ChatGPT tags many of the links it cites with utm_source=chatgpt.com, so a good share of its traffic is already visible in GA4 under that source.
• Since 13 May 2026, GA4’s default channel group has included an AI Assistant channel, but Google doesn’t publish the full list of assistants it recognises.
• Visits from apps, copied links and links opened without a referrer still arrive as Direct, and no report can label them with certainty.
• A regex exploration and a custom channel group take minutes to set up and give you a clean, retroactive view of AI referral traffic.
• Direct sessions from new users landing on deep pages, especially when they move in step with AI referrals, are your best signal of hidden AI visits, so report them as a range.

Where does the ChatGPT traffic Google Analytics misses go?

GA4 works out where a session came from using campaign tags in the URL (the utm_ parameters) and the referrer, which is the page the visitor clicked from. With neither, the session source is (direct) and the visit lands in Direct.

How ChatGPT tags its links

OpenAI’s Publishers and Developers FAQ, as archived on 4 October 2026, said ChatGPT automatically adds utm_source=chatgpt.com to referral URLs, so sites that allow OAI-SearchBot to crawl them can track the traffic in tools such as Google Analytics. OpenAI hasn’t said when the tag started. In June 2025, PPC Land reported that tags had been extended to the previously untagged links in ChatGPT’s “More” sources list. The tag isn’t on every link, and OpenAI can change the behaviour without notice. Because it sits in the URL, it survives when the referrer is lost.

When the referrer goes missing

• Apps. A link opened from a phone or desktop app has no referring web page. If the link has no tag, GA4 records Direct.
• Copy and paste. Someone pastes a link from an answer into WhatsApp or another browser. The full link keeps the tag; a tidied-up address doesn’t.
• Links opened with no referrer. A website can tell the browser not to send one, and privacy settings or extensions can trim it too.
• AI browsers and agents. Referrer behaviour when an assistant’s built-in browser or agent opens your page isn’t consistently documented, so test it rather than assume. OpenAI’s help centre says its Atlas browser has now been retired.
• Remembered recommendations. Someone reads your name in an answer and types your address a day later. AI influenced that visit, but no tag can catch it.

The October 2026 tagging bug

In a report on 14 October 2026, Search Engine Roundtable said ChatGPT links appeared to have lost their UTM tag around the GPT-6 release a week earlier, and were being opened with noopener and noreferrer settings, so there was no tag and no referrer. Adobe’s Cedric Huesler, quoted in the report, saw ChatGPT visits fall by about 31,000 a day and direct visits rise by about 26,000 a day across roughly 500 pages where ChatGPT normally sent at least 5% of visits. The report was updated on 15 October 2026: Huesler said it was a bug and the tags were back. OpenAI hasn’t published an explanation.

If your data covers early October 2026, pages that lost chatgpt.com sessions and gained direct sessions that week are pages ChatGPT sends people to, which makes the bug a useful natural experiment.

What does the GA4 AI Assistant channel catch?

On 13 May 2026, the Google Analytics release notes announced AI Assistant traffic measurement. When the referrer matches a recognised AI assistant, GA4 sets the medium to ai-assistant and the campaign to (ai-assistant), and files the session under a new AI Assistant channel. Google’s default channel group documentation gives ChatGPT, Gemini, DeepSeek, Copilot and Grok as examples and says the channel excludes Google’s AI Overviews and AI Mode. On 11 June 2026, Google added a Source Group dimension that groups sources for common platforms, including AI sources such as ChatGPT and Perplexity, and is filled in retroactively.

The GA4 AI Assistant channel has gaps:

• Google doesn’t publish the full list. Perplexity isn’t a named example, so check whether perplexity.ai shows as ai-assistant or referral in your property.
• It relies on the referrer. A visit with utm_source=chatgpt.com but no referrer may show a medium of (not set) and miss the channel.
• The release note doesn’t say older data is reclassified, so expect earlier AI visits under Referral.
• It can’t see visits with no referrer and no tag, which is the whole “direct” problem.

Which AI assistants show up as referrers?

These are the session sources you are most likely to see, as of October 2026. Confirm them against your own data.

How do you run the 30-minute GA4 check?

You need Analyst access to build an exploration and Editor access to create a channel group. Use the last three months of data and keep a note of what you find.

1. Minutes 0 to 5: check the AI Assistant channel in your standard reports.
2. Minutes 5 to 12: build an exploration that isolates AI referrers with a regex.
3. Minutes 12 to 17: create a custom channel group so the view sticks.
4. Minutes 17 to 25: inspect direct landing pages for signs of hidden AI visits.
5. Minutes 25 to 30: compare with Search Console and write down your estimate.

Step 1: Check the AI Assistant channel

Go to Reports, then Acquisition, then Traffic acquisition. Keep Session default channel group as the primary dimension and add Session source / medium as a secondary dimension. Search the table for “ai-assistant”, then for “perplexity” and “chatgpt”. If you find perplexity.ai / referral or chatgpt.com / (not set), the default channel is missing some AI visits. That tells you how much ChatGPT traffic Google Analytics already recognises.

Step 2: Find AI referral traffic with a regex exploration

Go to Explore and start a blank exploration. Import the dimensions Session source, Session source / medium, Landing page + query string, Date and New / established, and the metrics Sessions, Engaged sessions, Key events and Total revenue. Put Session source in Rows and the metrics in Values. Add a filter on Session source, choose “matches regex” and paste:

.*(chatgpt|openai\.com|perplexity|gemini\.google|copilot\.microsoft|claude\.ai|deepseek|grok\.com|meta\.ai|mistral\.ai).*

Here is what each part does:

• .* at each end means “anything or nothing”. GA4’s “matches regex” must match the whole value, so these allow for “chat.” or “www.” in front. With “partially matches regex” you can drop them.
• The brackets and | hold a list of alternatives; | means “or”.
• chatgpt catches chatgpt.com and any source tagged plainly as chatgpt.
• openai\.com catches older chat.openai.com visits. The backslash makes the dot literal, and plain “openai” would also match a site like openairtheatre.com.
• perplexity and gemini\.google catch Perplexity with or without www, and Gemini without matching Google Search.
• copilot\.microsoft, claude\.ai, deepseek and grok\.com catch Copilot, Claude, DeepSeek and Grok.
• meta\.ai and mistral\.ai cover Meta AI and Mistral’s Le Chat as a precaution; remove them if they never appear.

We tested the pattern against common source values: it matched every assistant hostname above and none of google, bing, mail.google.com, facebook.com, dailymail.co.uk or t.co. GA4 regex is case sensitive, which rarely matters because sources are normally lower case. Google’s help page offers an example AI regex too, but as written in October 2026 it matches any source ending in “ai”, such as notion.ai, and leaves out DeepSeek and Grok, so test any pattern against your own source list.

Now swap Session source for Landing page + query string. This is your list of pages AI assistants send people to; export it for Step 4. To have checks like this run each week, see our guide to AI agents for marketing data analysis.

Step 3: Build a GA4 custom channel group for AI

A custom channel group puts AI into your everyday reports, and Google says custom groups can be applied retroactively, so your trend line goes back before May 2026.

1. Go to Admin, then Data display, then Channel groups, and create a new group called something like “Channels with AI”.
2. Keep the existing channels and add one called “AI assistants”.
3. Set the condition to Source, “matches regex”, and paste the pattern from Step 2.
4. Reorder so AI assistants sits below the paid channels and above Organic Search and Referral. The first matching channel wins, so paid ChatGPT ads stay paid and organic AI visits leave Referral.
5. Save, then switch the primary dimension in Traffic acquisition to your new group.

A standard property can hold two custom channel groups (five on GA4 360), so agree with colleagues before you use one.

Step 4: Which direct traffic in GA4 looks like dark AI traffic?

Add a second tab to your exploration filtered to Session default channel group exactly matches Direct. Exclude your homepage and pages people bookmark, such as basket, account and login. Put Landing page + query string in Rows, New / established in Columns and Sessions in Values, then look for:

• Direct visits to deep pages. Nobody types the address of a 2,000-word guide or a specific comparison page.
• Mostly new users. Google’s help pages say sessions that start with a direct entrance are credited to the user’s earlier campaign details where GA4 has them, so what stays in Direct is largely visits with no known source in that browser.
• Overlap with your AI landing pages. If the same pages appear in your Step 2 export, that is a strong signal.
• Spikes that line up with brand mentions. Chart your top ten pages by Date and compare the peaks with tagged AI sessions, with when your brand began appearing in AI answers, and with mentions in widely cited articles or forum threads.

Rule out the ordinary causes too: untagged email links, QR codes, radio ads, PDFs, links shared in Teams or Slack, staff visits and campaigns launched without UTMs. A Leeds dental practice whose “composite bonding cost” page jumps the week a local newspaper covers it should look at the newspaper first, not ChatGPT.

How do Search Console and server logs help?

Search Console

Google announced a branded queries filter for the Search Console Performance report on 20 November 2025 and made it available to all eligible sites from 11 March 2026. If branded clicks rise alongside direct visits to deep pages, people may be hearing about you in AI answers and then searching your name. Keep Google’s own AI features separate: Google says AI Overviews and AI Mode clicks are counted in Search Console’s Web search type, and in GA4 they arrive as Google organic. Our post on Google’s AI Mode official sites test covers that side.

Server logs

Server logs record requests whether or not visitors accept analytics cookies, which matters in the UK, where consent rules mean GA4 may not see everyone. OpenAI’s crawler documentation explains that ChatGPT-User visits pages in response to user actions, while OAI-SearchBot crawls for ChatGPT search; Perplexity documents Perplexity-User and Anthropic documents Claude-User for the same purpose. Rising user-triggered fetches of a page, alongside rising direct sessions to it, is about the strongest supporting evidence you will get. Check suspicious requests against published IP ranges where they exist, such as OpenAI’s, because user agents can be faked.

How much of your direct traffic could be AI?

You can’t measure hidden AI visits directly, but you can bracket them:

1. Take total direct sessions for the month.
2. Keep only new-user sessions landing on deep pages. This is your unexplained pool.
3. Narrow it to pages that also received tagged AI referral traffic that month.
4. Find the same figure for a baseline: the same pages six months earlier, the same month last year for seasonal businesses, or similar pages with no AI referrals.
5. Your low estimate is the growth above the baseline; your high estimate is the whole AI-page pool. Express both as a share of Direct.

A worked example with illustrative figures: a UK online homeware store records 20,000 direct sessions in September. 6,000 are new users on deep pages, and 2,400 of those land on pages that also received tagged AI visits. The same pages had 1,500 such sessions in March. The low estimate is 900 sessions (4.5% of Direct) and the high estimate is 2,400 (12%), on top of 1,100 confirmed AI sessions in the custom channel group. Multiply by your average order value, say £60, only if you label the result an estimate.

Leave early October 2026 out of any baseline, because the tagging bug pushed ChatGPT visits into Direct for about a week. Use the same method every month: the trend will tell you more than any single number.

What should you report to stakeholders?

Separate what you measured from what you estimated. A modest, clearly labelled figure earns more trust than a big one that falls apart under questions.

A one-line summary might read:

AI assistants sent at least 1,100 sessions in September. We estimate a further 900 to 2,400 arrived as Direct, based on new-user visits to pages AI also links to.

For the visibility line, NSP’s SEO Rank & AI Visibility Tracker is a paid tool with a 14-day trial that tracks Google rankings and AI mentions together; logging manual prompt checks consistently works too. If you plan to hand this routine to an AI agent, our guide to autonomous AI marketing agents explains where human review still belongs.

What can’t this check prove?

• It can’t prove any single direct visit came from an AI assistant. It shows patterns, not journeys.
• Messaging apps, email apps, documents and offline activity also send direct visits to deep pages.
• GA4 only sees visitors your consent setup lets it record, so every figure is partial.
• Assistants change how they open links without notice, as October 2026 showed, so a sudden shift may be a tracking change rather than a change in demand.
• Much AI influence never produces a click. Our comparison of GEO vs SEO explains why visibility in answers matters even when traffic is flat.

If AI assistants are already sending people to your guides, the next job is making sure they describe you accurately. Our post on AI agents for GEO and AEO covers that work.

How NSP Global Services Can Help

Our Google Analytics services team can run this check, build the channel group and exploration, and set up a monthly report that separates confirmed AI traffic from estimates. If your property was set up in a hurry during the move from Universal Analytics, our GA4 migration and setup service can audit tagging, consent and channel rules first, because the check is only as good as the data underneath it.

To grow the AI traffic itself, our AI visibility services help your brand appear accurately in ChatGPT, Gemini, Perplexity and Copilot answers. Get in touch to talk through your data.

Conclusion

Some of your direct traffic is almost certainly AI traffic. Since May 2026, the ChatGPT traffic Google Analytics can see, the part that arrives with a tag or a referrer, has been easier to separate. The rest needs detective work: deep landing pages, new users, timing that matches AI referrals, and evidence from Search Console and server logs.

Spend 30 minutes on the check this week. Report what you measure as a floor, what you infer as a range, and how you got there. That honest split will hold up better than any headline number while assistants keep changing how they link to you.

Frequently Asked Questions

Does ChatGPT show up as a referral in Google Analytics?

Often, yes. ChatGPT adds utm_source=chatgpt.com to many links it cites, so GA4 records the session source as chatgpt.com. Since May 2026, visits with a recognised AI referrer can also appear under GA4’s AI Assistant channel with the medium ai-assistant, while older visits sit under Referral. Visits from apps or copied links that carry neither a tag nor a referrer still land in Direct.

What does utm_source=chatgpt.com mean in GA4?

It is a tracking tag ChatGPT adds to many of the links it cites. GA4 reads it as the session source, so you can filter or group those visits. It doesn’t mean you paid for the click: if you run ChatGPT ads, tag them with your own campaign values and report them separately. The tag isn’t on every link, and OpenAI can change the behaviour without notice.

Why has my direct traffic gone up in GA4?

Direct rises for many reasons: untagged email links, QR codes, offline advertising, staff visits, a consent banner change, broken tracking or campaigns launched without UTMs. AI assistants are another cause when visits arrive from apps or copied links. Check which landing pages gained direct sessions, whether the visitors were new users and whether those pages also receive tagged AI traffic before deciding AI is responsible.

Can I see Google AI Overviews or AI Mode traffic separately in GA4?

Not reliably. Clicks from AI Overviews and AI Mode arrive as Google organic traffic, and Google’s documentation says the GA4 AI Assistant channel excludes them. Search Console counts these clicks within the Web search type of the Performance report, but it doesn’t split them out. Treat them as part of your organic search reporting rather than your AI assistant figures.

How do I track Perplexity traffic in GA4?

Search your Traffic acquisition report for “perplexity” with Session source / medium as a secondary dimension. If it appears as perplexity.ai / referral, the default AI Assistant channel isn’t catching it in your property. Add Perplexity to a custom channel group using a regex condition on Source, and place that channel above Referral. Custom channel groups apply retroactively, so you also get its history.

Is GA4’s AI Assistant channel retroactive?

Google’s 13 May 2026 release note doesn’t say older data is reclassified, so expect AI visits before then to sit under Referral in the default channel group. Two retroactive options exist: custom channel groups, which Google says can be applied to reports retroactively, and the Source Group dimension added in June 2026, which Google says is populated retroactively.

How accurate are estimates of hidden AI traffic?

They are informed ranges, not measurements. You can’t prove an individual direct visit came from an AI assistant, so use a consistent method: new-user direct sessions to deep pages that also receive tagged AI visits, compared with a baseline. Support the estimate with Search Console branded clicks and server log fetches, and present a low and a high figure with the method stated.

Do AI crawlers inflate my Google Analytics numbers?

Usually not. GA4 automatically excludes known bot traffic, and crawlers such as OAI-SearchBot or PerplexityBot typically request page HTML without running your analytics tags. User-triggered fetchers such as ChatGPT-User show up in server logs rather than GA4 for the same reason. Agents that drive a full browser are a grey area, so watch for odd engagement patterns in direct traffic.