ChatGPT referral traffic in GA4: how to track it

We haven’t run a published test for this post. So what follows is the setup, its blind spots and a protocol we would run. It contains no Top SEO Experts results.

What does ChatGPT referral traffic look like in GA4?

It should show up as a session source containing chatgpt.com. OpenAI’s help center says publishers who allow OAI-SearchBot can track referral traffic from ChatGPT, and that ChatGPT automatically includes utm_source=chatgpt.com in referral URLs. That is documented behavior.

Two things the page does not tell you matter for setup. It does not say what medium value GA4 will show. It also does not cover every OpenAI surface, such as mobile apps or other products.

What does Google document about the medium?

Google does document one medium for AI tools. Its Default channel group page says GA4 sets the medium to ai-assistant when the referrer matches Google’s list of AI assistants. That list is Google’s, so check whether your chatgpt.com visits land in it.

The table below separates what the sources say from what they leave open.

Google’s own UTM guidance explains why the medium is worth checking. Google Analytics help says to always use utm_source, utm_medium and utm_campaign together, and that missing parameters produce “(not set)” values in reporting. A link carrying only a source tag may behave differently from a fully tagged campaign link.

Google does not say how a chatgpt.com visit that carries a utm_source tag is classified. We did not check a live property for this post, so we won’t state one. Open your own report and read what your property records, then write it down with the date.

What should you check before building reports?

Run three checks on the site and the property. Each one can quietly zero out your data:

1. Confirm robots.txt does not block OAI-SearchBot. OpenAI’s FAQ says to make sure you aren’t blocking it if you want content in ChatGPT summaries and snippets.
2. Confirm GA4 is collecting on every page, including the thank-you pages for forms and booking.
3. Confirm a form submission and a click-to-call are recorded as events.

Our guide to AI crawlers covers the first check in detail, including which bots to allow and why.

How do you find AI referral traffic in GA4 step by step?

Start with the standard report and a text search, and build a custom channel group only after you’ve seen the real source strings. The order is ours, not a Google procedure, and it exists so each step rests on data you have actually seen.

1. Open Reports, then Acquisition, then Traffic acquisition.
2. Set the date range to at least the last 90 days, so small sources are not hidden.
3. In the Session default channel group view, look for a row called AI Assistant. That is Google’s built-in channel for AI tools.
4. Change the primary dimension from Session default channel group to Session source / medium.
5. Use the search box above the table to look for chatgpt, then perplexity, gemini, copilot and claude. Note every exact source string that appears.
6. Add Landing page as a secondary dimension, so you can see which pages those visits entered on.
7. Record the strings, the medium values and the date in a notes file.

Step 5 is the one people skip, pasting a regex from the internet before seeing what their own property records. Source strings may differ from the names above. Google notes that UTM values are case sensitive, so one source in two spellings becomes two rows.

Does GA4 already have an AI channel, and when do you need a custom one?

GA4 already has one. Google’s Default channel group page lists an AI Assistant channel for visits from sources like ChatGPT, Gemini, Deepseek, Copilot and Grok. The channel matches the medium ai-assistant.

That page also tells you where AI Overviews and AI Mode traffic goes. The AI Assistant channel excludes both, and the Organic Search description includes them. So Google-side AI clicks will not appear in the AI Assistant channel.

Google says default channel groups cannot be edited. So a custom channel group is for a different job: your own list of sources, or sources the default channel misses. Create one in Admin, add a channel named “AI assistants,” and give it a rule that matches the source strings you recorded.

• Default AI Assistant channelBuilt in. Matches the medium ai-assistant. Cannot be edited. Excludes AI Overviews and AI Mode.
• Custom channel groupBuilt in Admin. Uses your own rule and the source strings you recorded. Fills gaps in the default.

A starting pattern is below. It’s a draft to edit against your own data, not a documented list of values:

chatgpt|openai|perplexity|gemini\.google|copilot\.microsoft|claude\.ai

GA4 shows a “matches regex” option in several condition menus. We have not cited a Google page for regex syntax in this post. So test the rule in the preview and check the results row by row.

Four rules keep the pattern honest as your data grows:

• If a bare word could match something unrelated, anchor it. A pattern like copilot alone can also catch other products with that name. Use the full hostname.
• If a source string appears in your report that is not in your pattern, add it only after you have seen it at least twice. One stray row is not a pattern.
• If a string appears with a medium that suggests paid placement, keep it out of the AI channel and review it separately.
• If you cannot tell what a source is, leave it in its default channel and add a note.

Channel rules are evaluated in an order, and we recommend placing your custom AI channel above Referral and checking the order in the interface. That is our practice, not documented behavior, so verify that your test visits land where you expect.

After saving, check the channel against a known visit. Click a link from a real ChatGPT answer in a fresh browser window, then watch the Realtime report for the source string.

Why do some AI visits show up as Direct or branded search?

Because GA4 can only classify what the click carries. Google defines Direct as visits from a saved link or a typed URL, with the source exactly “(direct)” and the medium “(not set)” or “(none)”. A visit with no referrer and no tag has nowhere else to go.

Four things hide AI influence from a referral report. One is documented, one is measured by Pew, and two are our hypotheses.

Can a later branded search hide the AI answer?

This is our hypothesis, not a measured pattern. The visit appears as branded organic search, and the AI answer never gets credit. Our post on branded search growth covers how to read that signal.

What if a click arrives with no source information?

GA4 would then record it as Direct. Whether any given app or browser does this is theory until you test it, and we have not tested it.

Where do Google’s AI features show up?

This is documented. Google’s AI features page says sites shown in AI features are included in overall search traffic in Search Console. It describes no separate report for them.

What if the click never happens?

Pew Research Center’s July 2025 analysis looked at the browsing of 900 U.S. adults in March 2025. It found that people clicked a regular result in 8% of visits to pages with an AI summary, against 15% without one. Only 1% of visits clicked a link inside the summary.

• Regular click, no AI summary15%
• Regular click, AI summary shown8%
• Link inside the AI summary1%

Source: Pew Research Center

View the data as a table

Item	Value
Regular click, no AI summary	15%
Regular click, AI summary shown	8%
Link inside the AI summary	1%

Pew studied Google AI summaries, not ChatGPT, and the sample is U.S. adults across all kinds of searches. It does not tell you what share of AI-assisted buyers click. It does tell you that a referral report will undercount exposure.

Google’s AI features page also says AI feature clicks are “higher quality.” Google gives no data on that page. So test it against your own conversions before believing it.

Do the Google and Bing webmaster tools show AI traffic?

Search Console covers Google, and Bing Webmaster Tools covers Microsoft’s side. Neither is a ChatGPT report.

Google counts AI feature appearances in its normal Web search data. Bing’s AI Performance report counts citations, not visits.

Bing introduced that report in public preview on its Webmaster Blog. It shows:

• Total Citations
• Average Cited Pages
• Grounding queries, which are a sample of the phrases the AI used to retrieve content
• Page-level citation counts

Microsoft states its limits plainly:

• Total Citations does not show where a citation appears in an answer.
• Average Cited Pages does not indicate ranking or importance.

Together, the sources cover different ground:

Our Copilot visibility post covers Microsoft’s side, and our guide to measuring AI visibility covers the wider picture.

How do you tie AI referral visits to calls and form leads?

Mark the actions that matter as key events in GA4, then look at AI sessions by landing page and by event. Google describes a key event as one that measures an action particularly important to your business, and any collected event can be marked as one.

For a service business, the events that count are phone calls, form submissions and booking requests. Pageviews are not outcomes.

1. Mark key eventsMark calls, form submissions and booking requests as key events in GA4.
2. Read AI sessions by page and eventLook at AI sessions by landing page and by event.
3. Ask how people heard of youAdd "How did you hear about us?" to the form and call script.
4. Check the landing pagesKeep facts and contact details current, with a clear next step.

Why are phone calls the hard part?

Phone calls are the hard part. A click on a phone number can be tracked as an event. But many calls start from a dialed number or a Google Business Profile listing.

Our call tracking guide explains how to connect calls to organic sessions.

Once a lead reaches your CRM, our posts on SEO attribution and attribution inside ServiceTitan pick up the trail.

What does the math look like?

Some arithmetic, with invented numbers. A hypothetical remodeler’s property records 40 sessions from AI sources in a quarter.

Three of those hypothetical sessions include a form submission, and 37 do not. That is 3 divided by 40, or 7.5%.

• With a form submission3 sessions
• Without a form submission37 sessions

Source: Hypothetical example

View the data as a table

Item	Value
With a form submission	3 sessions
Without a form submission	37 sessions

Do not read much into 7.5%. One conversion more or fewer moves it by 2.5 points. The numbers are invented to show the math, not a benchmark.

What else can you add?

Add one human question. Put “How did you hear about us?” on the form and in the call script, with an option for AI assistants such as ChatGPT. It is crude, and it is the only method that catches people who read an answer and then typed your name.

Finally, check the landing pages. A service page that attracts AI clicks is telling you which question it answers well. Keep its facts and contact details current, and give it a clear next step.

Our guide to service page SEO shows what a converting page needs, and our AI search optimization service starts from this kind of evidence.

How do you run a first-party test you can trust?

Run the same short prompt set on a schedule, record everything with dates, and treat the results as a sample. This is the protocol we would use. We have not published a result set here, so treat it as a method, not a finding.

Why does repetition matter?

Repetition matters because of the research. A SparkToro and Gumshoe.ai study ran over November and December. It was reported by Search Engine Journal on 30 January 2026.

The study found that ChatGPT and Google’s AI returned the same brand list in under 1% of repeated runs.

One run therefore proves nothing. The protocol below is sized for a small business with an afternoon to spare:

1. Write 10 prompts a customer might type, such as “emergency AC repair near me” or “who installs standing seam metal roofs in Tampa.” Keep them fixed.
2. Choose the engines you care about, for example ChatGPT, Perplexity and Google AI Mode.
3. Run each prompt 3 times per engine, in a logged-out or clean browser, on the same day.
4. For each run, log the date, engine, prompt, whether your site was cited, which URL, and whether the facts were correct.
5. Click each citation to your site and watch Realtime in GA4 to see the source string it produces.
6. Repeat monthly. Never compare a month with a different prompt set.

The infographic below shows the loop in short form. After it, a spreadsheet with these columns holds everything the protocol produces.

1. Write fixed promptsPick 10 prompts a customer might type and keep them the same.
2. Choose the enginesPick the AI tools you care about, such as ChatGPT or Perplexity.
3. Run each prompt 3 timesUse a clean browser, on the same day, for each engine.
4. Log every runNote the date, engine, prompt, citation, URL and fact check.
5. Click and watch GA4Click each citation and read the source string in Realtime.
6. Repeat monthlyNever compare a month that used a different prompt set.

Here is the spreadsheet layout:

Three runs is a floor, not a statistically sound sample. The method gives you a pattern to read and a record of what GA4 shows for real clicks.

It does not give you a visibility percentage you can defend. Our post on GEO tracking limits explains why even paid tools struggle there.

What is still unproven about AI referral traffic?

A lot of it. Each item below is either undocumented in the sources we could open or untested by us.

• Perplexity referral behavior. We could not open a Perplexity document on its referrers or UTM tags, so we make no claim. Check your own data.
• Which channel a chatgpt.com visit lands in. OpenAI documents the source parameter, and Google documents an ai-assistant medium for referrers on its own list. We did not check how a tagged chatgpt.com visit is classified in a live property.
• Apps and browsers that strip referrers. A plausible cause of Direct traffic, and unmeasured by us.
• Whether AI visitors convert better. Google calls AI feature clicks higher quality but publishes no data on that page. So any claim of better conversion, ours included, stays a hypothesis until your own numbers show it.
• How much AI exposure leaves no click. Pew’s data suggests many visits end without one, but it covers Google summaries only.
• Whether you can isolate Google AI Mode clicks in GA4. Google documents none.
• Whether llms.txt or any file changes AI referrals. Experimental at best. We have no evidence either way.
• Retroactive reporting. We did not verify whether a new channel group reprocesses past data.

What should a monthly AI referral check look like?

About ten minutes, on a fixed day, using the same saved views every time. The point is a dated record you can compare, not a dashboard you admire. This is our routine, offered as our view and not as a Google procedure.

1. Open Traffic acquisition for the last 28 days, with Session source / medium as the dimension. Write down the AI-looking rows and their session counts.
2. Add Landing page as a secondary dimension and note which pages the AI sessions entered on.
3. Check the key events for those sessions: form submits, booking requests, phone-number clicks.
4. Log any new source string, with the date you first saw it.
5. Compare the count to the previous month, as raw numbers, not percentages.

Then apply a rule about the channel group. Under about 30 AI sessions a month, skip it and keep the saved text search, since you should be reading each visit anyway. Past that, the group saves time and earns its upkeep.

Check the basics first if the numbers look oddly low. Our website SEO checker flags crawl and indexing problems that stop pages from being cited at all, and fixing those comes before measuring anything.

Report counts instead of growth rates until volume is steady. Judge the channel by leads, not visits. A jump from 4 sessions to 9 is a doubling, and it is also noise.

Frequently asked questions

Get Your Organic Growth Plan

Sources

• OpenAI Help Center, “Publishers and developers FAQ”: https://help.openai.com/en/articles/12627856 (accessed 22 Oct 2026)
• Google Analytics Help, “Default channel group”: https://support.google.com/analytics/answer/9756891 (accessed 23 Oct 2026)
• Google Analytics Help, “Best practices for creating custom campaigns”: https://support.google.com/analytics/answer/10917952 (accessed 22 Oct 2026)
• Google Analytics Help, “Key events”: https://support.google.com/analytics/answer/9267568 (accessed 22 Oct 2026)
• Google Search Central, “AI features and your website”: https://developers.google.com/search/docs/appearance/ai-features (accessed 22 Oct 2026)
• Bing Webmaster Blog, “Introducing AI Performance in Bing Webmaster Tools (Public Preview)”: https://blogs.bing.com/webmaster/2026/2/Introducing-AI-Performance-in-Bing-Webmaster-Tools-Public-Preview/ (accessed 22 Oct 2026)
• Pew Research Center, “Google users are less likely to click on links when an AI summary appears in the results,” 22 July 2025: https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/ (accessed 22 Oct 2026)
• Search Engine Journal, “AI recommendations change with nearly every query: SparkToro,” 30 Jan 2026: https://www.searchenginejournal.com/ai-recommendations-change-with-nearly-every-query-sparktoro/566242 (accessed 22 Oct 2026)