Your ChatGPT traffic may already be in GA4, buried under Referral beside random blogs and tool spam. I start by fixing the filing system, not by chasing more traffic.
By the end of this tutorial, you will have a custom channel group that separates AI assistants, a saved landing-page report, and a way to ask your assistant which pages deserve attention through groas. You will also know what this setup cannot recover. Better reporting is useful. Pretending it captures everything is not.
Before starting: get access and choose your reporting window
Have these ready:
Editoraccess to your GA4 property. Viewer access cannot create channel groups.- About 20 minutes for the setup.
- A reporting window with enough traffic to inspect. I use 90 days where available.
- Key events configured for the actions you want to measure.
GA4 now includes a built-in AI Assistant channel. Check it, but do not stop there. This tutorial creates separate rows for ChatGPT, Perplexity, Gemini, Claude and Copilot so you can inspect each assistant rather than one combined bucket.
Custom channel groups apply retroactively. You can reclassify recorded traffic; you cannot recreate attribution signals GA4 never received. That distinction matters throughout the setup.
Step 1: Find the AI sources your property actually records
Open Reports > Acquisition > Traffic acquisition. Set the primary dimension to Session source, then search for these names one at a time:
chatgpt
openai
perplexity
gemini
claude
copilot
deepseek
grok
Write down the source values that appear. Common examples include chatgpt.com, legacy chat.openai.com, perplexity.ai, gemini.google.com, claude.ai, copilot.microsoft.com, chat.deepseek.com and grok.com. These are the domains practitioner guides repeatedly identify, but your own report should decide the list.
For the five-assistant setup below, the combined source pattern is:
chatgpt\.com|chat\.openai\.com|perplexity\.ai|gemini\.google\.com|claude\.ai|copilot\.microsoft\.com
Do not build a rule that matches anything containing ai. That is a substring, not an attribution strategy. It can collect unrelated tools and .ai domains along with the assistants you meant to measure.
Expected result: a short list of recognizable assistant sources already present in your property.
Common mistake: copying a domain list without checking whether your source values match it. If a source differs, use the value your report actually records rather than silently excluding it.
Step 2: Create one channel group with separate assistant rows
Go to Admin > Data display > Channel groups. Open the three-dot menu beside Default Channel Group, choose Copy to create new, and name the copy:
AI split by assistant
Work inside that copy. Standard properties allow up to two custom channel groups, with up to 25 channels in each, so keep the assistant rows together rather than spending a group on each one.
Add a channel named AI: ChatGPT. Set its condition to Source → matches regex, then paste:
chatgpt\.com|chat\.openai\.com
Repeat for the other four channels:
| Channel name | Source pattern |
|---|---|
AI: Perplexity | perplexity\.ai |
AI: Gemini | gemini\.google\.com |
AI: Claude | claude\.ai |
AI: Copilot | copilot\.microsoft\.com |
GA4 regex uses full matching by default, and an unescaped dot is a wildcard. These patterns match the listed source values. If Step 1 revealed a different value, adjust the relevant pattern deliberately.
Do not add a condition requiring the default channel to be Referral. A recognizable assistant source can arrive through a tagged URL without a referrer and fall outside that bucket. The source is the signal this group needs to catch.

Drag all five assistant channels above the existing channels that could otherwise claim those sessions, including Referral, Organic Search and AI Assistant if present. Apply the ordering and save the group.
GA4 assigns traffic to the first channel it matches. A perfect rule underneath an earlier matching rule does nothing useful.
Expected result: one saved group containing five named assistant channels, followed by the existing channels.
Common mistake: leaving the new rows below Referral and concluding that AI traffic is zero. Check channel order before debugging the regex.
Step 3: Confirm the group reclassifies historical sessions
Return to Reports > Acquisition > Traffic acquisition. Select your new group’s session-scoped dimension in the primary-dimension picker, then use the same date range you inspected in Step 1.
Look for the assistant sources you already found. Their sessions should now appear under the corresponding AI: rows rather than whichever broader channels previously held them.
Do not expect every assistant row to contain traffic. A rule creates a category, not visitors. If your property recorded no Claude sessions, an empty Claude row is not a setup failure.
If an expected row is empty, check in this order:
- Confirm that the report uses
AI split by assistant, not the default group. - Confirm that the date range includes the source sessions from Step 1.
- Check the new channels’ position in the group.
- Compare the recorded source value with the regex, character by character.
- Confirm that no Referral-only condition slipped into the rule.
Expected result: historical sessions with matching sources appear in the assistant rows. You do not need to wait for new traffic to test the classification.
Common mistake: expecting a custom group to repair Direct sessions that contain no assistant source. It cannot. Fix the classification you have evidence for; leave the rest unclaimed.
Step 4: Build the landing-page and conversion view
Open Explore > Blank exploration. Name the exploration so you can find it again, then import these dimensions:
Session sourceLanding page + query string- The session-scoped dimension for
AI split by assistant
Import these metrics:
SessionsEngaged sessionsKey eventsSession key event rate
Set up the free-form table:
- Put
Landing page + query stringinRows. - Put your custom channel group’s session dimension in
Columns. - Put the four metrics in
Values. - Add a filter on that same channel-group dimension:
begins with→AI:. - Set the reporting dates and save the exploration.
This landing-page layout connects assistant traffic with the pages people actually reach. The acquisition report tells you volume. This view gives you somewhere to act.
Keep Landing page + query string while investigating. ChatGPT can append utm_source=chatgpt.com to destination URLs, so the query string can help explain how a visit arrived.
But do not confuse URL detail with session attribution. Removing the query string from the displayed landing-page dimension does not erase the session source or merge non-AI traffic into an AI-filtered report. It groups URL variants together. Keep the detail when diagnosing; simplify the view when that detail becomes clutter.
Expected result: separate assistant columns showing landing pages, session volume and key-event performance.
Common mistake: treating key-event counts as conversion rates. A page with more events may simply have more traffic. Read the rate alongside the count, and check that the configured key events represent actions you care about.
Step 5: Mark the traffic this setup cannot identify
Before sharing the report, add a plain-language limitation: these are identifiable AI-referred sessions, not all visits influenced by AI.
GA4 needs a usable signal to attribute a visit to an assistant. App opens and copied-and-pasted links can arrive without a referrer. A source tag can preserve attribution when the referrer is absent; without either signal, this channel group has nothing to match.
Source-only tagging also explains why the rules must not require Referral. A ChatGPT URL tagged only with utm_source=chatgpt.com can appear as chatgpt.com / (not set). The recognizable source is still useful even when the medium is not.
Keep three other boundaries clear:
- A citation without a click is not a website session.
- Crawlers such as GPTBot that do not run your tag are not measured as these visitor sessions.
- Google’s AI Overviews and AI Mode remain in Organic Search. Do not force them into this assistant table.
Tagging behavior can change, too. Profound documented a sharp interruption in ChatGPT’s UTM tagging and warns against turning its observations into a multiplier for your revenue reports. Restored tags do not repair visits already recorded without them.
Expected result: a report labeled as an attributable slice of traffic, with its limits visible.
Common mistake: multiplying the measured sessions by a guessed correction factor. Treat the count as a floor, not an invitation to invent the missing total.
Step 6: Ask your assistant through groas to read the report
Once the group works, connect GA4 through groas and give your assistant the group name. The repetitive job is pulling the same sessions, pages and key events every week without changing the question halfway through.
Start with:
Through groas, use the GA4 channel group “AI split by assistant” to show AI referral sessions by assistant for the last 90 days. List the top landing pages, sessions, key events and session key event rate. Keep the reporting window consistent and flag missing data rather than estimating it.
Then compare visits with citations:
Compare those landing pages with the pages AI answers cite for our tracked prompts. Where do the two lists differ? Separate pages that are cited without recorded visits from pages that receive AI visits but do not appear in the tracked citations.
That comparison makes why ChatGPT and Perplexity cite some pages and not others a question about your own pages rather than somebody else’s theory.

Visits and citations do different jobs. A page can appear in ten answers and receive zero clicks. Another can receive ChatGPT visits while Perplexity cites a competitor.
Through groas’s AI mention and citation workflow, your assistant can compare the GA4 sessions with the AI visibility data groas collects around the clock. That saves the two-CSV export and the ritual squinting afterward.
Expected result: a shortlist of visited pages, cited pages and mismatches worth investigating.
Common mistake: treating your tracked prompts as every answer an assistant gives. Compare the two datasets without pretending either is complete.
Step 7: Fix one landing page before chasing more visits
Return to the exploration. Find a page that receives AI traffic but performs poorly on the action you want visitors to take. Compare it with the same page’s organic traffic over the same dates.
For example, suppose ChatGPT sends 400 sessions to a pricing page with a 0.8% session key event rate, while organic traffic converts on that page at 2.4%. That is a reason to inspect the page, not proof that attribution needs another dashboard.
I start with three checks:
- Does the opening answer the question that brought the visitor there?
- Is supporting proof visible within one scroll?
- Is the next action obvious without hunting through navigation?
Change one thing on a page AI visitors already reach. Start with the CTA if it is unclear, or the opening if it fails to deliver on the likely question.
Ask your assistant to propose the edit first. Review it, then explicitly approve publication through groas. Changes are logged with their reasons and can be undone in one step. Autopilot stays off until you enable it, but that setting is not a substitute for a clear approval instruction.
Expected result: one deliberate page change with a recorded reason, not a site-wide rewrite prompted by a small table.
Common mistake: declaring the channel a failure because one landing page converts badly. Inspect what the visitor encounters before buying more traffic to send there.
Step 8: Verify the setup, then repeat the useful part monthly
Reopen Reports > Acquisition > Traffic acquisition and select AI split by assistant. Confirm that known assistant sources appear in the correct rows for your historical reporting window. Then reopen the saved exploration and confirm that its filter includes only AI: channels.
That is the verification: known sources classify correctly, and the landing-page report carries that classification through to key events. A populated table alone is not enough if it is the wrong table.
Once a month:
- Record sessions and key events by assistant.
- List the top five landing pages and their session key event rates.
- Note which tracked cited pages receive no recorded visits.
- Pick one page to improve.
- Keep the same attribution caveat beside the numbers.
For an outside check, use the monthly routine combining GA4, Search Console branded clicks, a how-did-you-hear-about-us form and a 30-to-50-prompt tracker. Those signals add context; they do not manufacture missing GA4 sessions.
After the first page edit, wait two weeks and rerun the first groas prompt. Compare the same metrics, while remembering that a before-and-after movement is not proof by itself.
The first change once the setup works: fix the CTA on a high-traffic AI landing page that converts below its organic rate. Do that before adding more assistants to the regex.
Build the group. Save the exploration. Fix the page. Some traffic really is invisible to GA4. The traffic already sitting under Referral does not need another attribution debate. It needs a better report.

