How to Measure AI Referral Traffic in Google Analytics 4 (GA4)

Why both layers? Google's channel is a good start, but its definition names five assistants, not Perplexity, and Google doesn't publish the full list it matches. In SE Ranking's 2026 data, Perplexity and Claude together sent 9.85% of all AI referral traffic from January to April (7.23% plus 2.62%). That's the slice most likely to sit in Referral. ChatGPT's own tag can push some visits into Unassigned, and clicks that lose their referrer land in Direct.

AI referral traffic is small, which is why it's easy to get wrong. In April 2026, AI assistants sent about a third of one percent of all US website visits in the same study of 101,574 sites. Misfile a few hundred sessions and your AI channel can look a third smaller than it is.

This guide is the full GA4 setup: a copy-paste rule set with a tested regex, the exact steps, where visits still get lost, a worked example, and three ways to build the comparison. For the wider picture (citations, prompt panels and control tests), read our guide to measuring GEO. For a short definition, see AI referral traffic in our glossary.

Where AI Referral Traffic Lands in GA4 Today

An AI click can end up in four places: Google's AI Assistant channel, Referral, Unassigned or Direct. Which one depends on the assistant, on how the link is built, and on whether the browser passes a referrer.

What Google's AI Assistant channel does

On 13 May 2026, Google added an AI Assistant channel to the default channel group. When a session's referrer matches Google's list of assistants, GA4 sets the medium to ai-assistant and the campaign to (ai-assistant), and files the session under the new channel. The channel definition describes its sources as "ChatGPT, Gemini, Deepseek, Copilot, or Grok." The launch note gives ChatGPT, Gemini and Claude as examples.

Three limits follow from how it works:

• It needs a referrer. A click that arrives with no referrer can't match, so it stays in Direct.
• The list is private. Search Engine Journal noted at launch that Google hasn't published it. The source-category file on Google's help page dates from February 2023 and has no AI hosts.
• Google's own AI is excluded. Clicks from AI Overviews and AI Mode count as Organic Search, and GA4 can't split them out.

Google also hasn't said whether the new channel re-sorts sessions from before 13 May. Custom channel groups, which you'll build below, are documented as retroactive.

How each assistant's AI referral traffic shows up by default

Treat the last column as a starting guess. Open Reports, then Acquisition, then Traffic acquisition, set the primary dimension to Session source / medium, and search for each host to see where your own sessions sit. Our engine guides cover each referrer in depth, starting with ChatGPT, Perplexity and Copilot.

Four rows need a note. Microsoft is moving the Microsoft 365 Copilot web app in September and October 2026. Copilot Search, launched in Bing in April 2025, lives on bing.com, so its clicks look like ordinary Bing visits. SE Ranking found DeepSeek's referrals "dropped to essentially zero across all regions" from September 2025. And Phind shut down in January 2026.

Why ChatGPT visits can land in Unassigned

OpenAI's publisher FAQ says ChatGPT adds utm_source=chatgpt.com to the links in its answers. That tag cuts both ways.

It helps because the tag rides in the URL, not in the Referer header. A click that loses its referrer can still be credited to chatgpt.com, as long as nothing on your site strips the query string.

It hurts because OpenAI documents a source but no medium. Google's processing rules map UTM values straight onto source and medium, and use the referrer only when no campaign fields are set. So a tagged click can be recorded as chatgpt.com / (not set). No default channel rule matches that pair, so the session drops into Unassigned, a pattern Analytics Mania describes for AI links. Other analysts report ChatGPT sessions as chatgpt.com / referral, so check which one your property shows.

Google doesn't say how its new ai-assistant rewrite interacts with the tag. The safe answer is a rule on Source, which catches chatgpt.com whatever medium GA4 records.

A Copy-Paste Rule Set for AI Referral Traffic

This rule set sorts AI referral traffic into one channel. It uses two conditions joined with OR, so it catches everything Google already recognizes, plus the assistants Google doesn't name.

The rules

^((www\.)?perplexity\.ai|chatgpt\.com|chat\.openai\.com|claude\.ai|gemini\.google\.com|copilot\.microsoft\.com|(www\.)?copilot\.com|(copilot|m365)\.cloud\.microsoft|(www\.)?meta\.ai|chat\.mistral\.ai|chat\.deepseek\.com|grok\.com|(www\.)?you\.com|(www\.)?phind\.com)$

What each part of the regex catches

• ChatGPT: chatgpt.com, and chat.openai.com for older links, but not openai.com's blog, docs or forum.
• Perplexity: with or without www., since the bare domain redirects to the www host.
• Copilot: both consumer hosts and both addresses of the Microsoft 365 Copilot web app.
• The rest: Claude, Gemini, Meta AI, Mistral (its assistant is now called Vibe, at the same address), DeepSeek, Grok, You.com, and Phind for old data.

The ^ and $ change nothing inside GA4, where "matches regex" is a full match by default. They make the same pattern safe in tools that match partially, such as BigQuery.

Why just above Referral, and not at the top? Paid clicks. If you buy ChatGPT ads (OpenAI opened a beta self-serve Ads Manager in May 2026, per SE Ranking) and tag them with utm_source=chatgpt.com, a source rule at the top would count them as organic AI referral traffic. Below the paid channels, a paid medium wins first. Build those links with our UTM link builder and a medium such as cpc.

Tested against 54 hostnames

The regex was run through Python's re module and through Google's RE2 library, whose syntax GA4 uses. It matched all 20 assistant hosts and none of the 34 others, in both full-match and partial-match mode.

Why not paste Google's sample regex?

Google's help page offers a sample that starts ^.*ai|.*\.openai.*|.*chatgpt.*. Run it against the same 54 hosts under GA4's full-match rule, and it counts 15 of the 34 non-assistant hosts as AI. It catches every source that ends in "ai", so a link from a competitor on a .ai domain, like the fictional loopcraft.ai, becomes AI referral traffic. It also catches gemini.com, bombardier.com, bard.edu, gptzero.me, copilot.money and OpenAI's help, docs and forum hosts.

It misses, too: chat.deepseek.com, grok.com, you.com, phind.com and m365.cloud.microsoft all fall through. Exact hostnames take longer to type, but they don't guess.

Set Up the AI Referral Traffic Channel, Step by Step

Google documents the path on its custom channel groups help page. You need the Editor role or higher on the property.

1. In GA4, open Admin. Under Data display, click Channel groups.
2. Click Create new channel group. GA4 starts you with a copy of the default group. Name it, for example "AI view".
3. If the copy already contains Google's AI Assistant channel, open and edit it. If not, click Add new channel and name it "AI assistants".
4. Click + Add condition group. Choose Source, then matches regex, and paste the regex.
5. Add a second condition joined with OR: Medium, exactly matches, ai-assistant. Click Save channel.
6. Click Reorder, drag "AI assistants" directly above Referral, and click Apply.
7. Click Save group.

Step 6 is the one people skip. GA4 files each session under the first channel whose rules it meets, in the order you set. Leave the AI channel below Referral and any assistant visit with a referral medium never reaches it.

Limits and rules to know

Check that it works

Open Reports, Acquisition, Traffic acquisition. Switch the primary dimension to your new group (a session custom channel group dimension) and add Session source as a secondary dimension. You should see one AI assistants row, with each assistant under it. Then run a monthly leak check:

• Referral: filter Session source with a loose pattern such as .*(gpt|chat|copilot|assistant).*. Any AI host your regex misses is a candidate to add.
• Unassigned: chatgpt.com rows should be gone. If not, check the regex, and check that you're viewing the new group rather than the default one.
• Direct on deep pages: nobody types a comparison page from memory. A rise in Direct landings there, with no campaign to explain it, hints at lost referrers.

Where AI Referral Traffic Still Gets Lost

A channel group can only re-sort sessions that carry a source. A visit with no referrer and no tag is Direct before your rules run.

The browser sends a hostname at most

Browsers default to a strict-origin-when-cross-origin referrer policy, which sends only the origin (the scheme and host) to another site, per MDN. On 28 September 2026, the live assistant web apps in the table above served that policy, a similar one (origin or origin-when-cross-origin), or none, which leaves the default in place. So GA4 can see claude.ai, but never the conversation. You can count AI referral traffic by assistant, not by prompt.

Five ways the referrer disappears

1. Apps and in-app browsers. Seer Interactive warns that "traffic from in-app browsers, copy-pasted links, or other 'dark' paths will still land in direct." The vendors don't document what their desktop and mobile apps send.
2. Copy and paste. A link pasted into a new tab carries no referrer.
3. HTTPS to HTTP. Under the default policy, a browser sends no Referer to a less secure page. If an answer cites an old http:// address of yours, the click arrives without one, even after your redirect to HTTPS.
4. Redirects that drop the query string. ChatGPT's tag survives a lost referrer only if every redirect on your side keeps the ?utm_source= part.
5. Links that opt out. A link marked rel="noreferrer", or a page with a no-referrer policy, sends nothing.

How much each app loses isn't well documented yet. Rankbox is testing referrer loss app by app and will publish the results.

Some AI clicks do arrive, just not as AI. Google's AI Overviews and AI Mode stay in Organic Search, Copilot Search looks like Bing organic, and Meta doesn't document the referrer its answers send inside Facebook, Instagram and WhatsApp. Treat Meta AI numbers in GA4 as partial.

Fixes that do move visits out of Direct

• Keep query strings through every redirect: trailing slashes, HTTP to HTTPS, locale switches.
• Show engines your HTTPS addresses only. Canonical tags, sitemaps and internal links should all use https://.
• Tag links you control inside AI products, such as a custom GPT, an app listing or a product feed, with your own source and medium from the UTM link builder.
• Read your server logs for demand analytics can't see. Hits from ChatGPT-User or Perplexity-User mean a user's request fetched your page. Our AI crawler log analyzer counts them per URL.

Worked Example: Plannora Before and After the AI Channel

Plannora is a made-up project management app at plannora.io. The numbers are illustrative, not real data, but they add up.

September: the same 30 days, two channel groups

Where the 1,000 came from:

• 720 from Google's AI Assistant channel: ChatGPT 610, Gemini 70, Copilot 40.
• 230 from Referral: Perplexity 150, Claude 70, Mistral 8, Meta AI 2.
• 50 from Unassigned: tagged chatgpt.com / (not set) sessions.

That's 720 + 230 + 50 = 1,000. AI's share of sessions rises from 0.72% to 1.0%, so Plannora's AI referral traffic is 39% bigger than the default view showed (280 ÷ 720).

Direct doesn't move, and it can't. Those 18,000 sessions carry no source for any rule to read.

October: the fix that does move Direct

Plannora's server redirected /pricing to /pricing/ and dropped the query string on the way. A ChatGPT click from an app that lost its referrer arrived with no tag, so GA4 filed it as Direct. In October the redirect was changed to keep the query string.

With the total flat, 55 sessions moved from Direct into the AI channel. The channel group made AI visits visible. The redirect fix made more of them countable.

Conversion rates, and when to trust them

On paper, Plannora's AI referral traffic converts 36% better (3.4 ÷ 2.5 = 1.36). Don't put that on a slide yet.

At 1,000 sessions, the 95% margin of error on 3.4% is about ±1.1 points: 1.96 × √(0.034 × 0.966 ÷ 1,000). The true rate could be anywhere from 2.3% to 4.5%, a range that includes organic's 2.5%. A standard two-proportion test gives p ≈ 0.07, short of the usual 0.05 bar.

Pool three months at the same rates, 3,000 sessions and 102 signups, and the margin shrinks to about ±0.65 points, or 2.8% to 4.0%. Now the gap clears organic (p ≈ 0.002). To detect a gap this size reliably (80% power), plan on roughly 2,700 AI sessions.

Per-engine rates are weaker still. Perplexity's 150 sessions with, say, 8 signups give 5.3%, with a margin of ±3.6 points. Report engines by volume and trend, and keep conversion rates at the channel level.

Compare AI Referral Traffic With Organic Search

The same comparison can live in three places. GA4 is fastest, Data Studio is best for sharing, and BigQuery is best for auditing.

In GA4: one exploration with three tabs

1. Open Explore and start a Free form exploration.
2. Add dimensions: your session custom channel group, Session source, and Landing page + query string.
3. Add metrics: Sessions, Engaged sessions, Engagement rate, Key events and Session key event rate.
4. Tab 1, the comparison: rows by channel group, filtered to AI assistants and Organic Search.
5. Tab 2, the engine split: rows by Session source, filtered to AI assistants.
6. Tab 3, the pages: rows by landing page, filtered to AI assistants, sorted by key events.

Google defines session key event rate as the share of sessions with at least one key event, and engagement rate as the share of engaged sessions: 10 seconds or longer, a key event, or two or more page views. If you mark several key events, compare the one that matters, such as the signup. In standard reports, the Key events metric has a drop-down that narrows it to a single key event.

ChatGPT traffic analysis in three checks

ChatGPT sends about three quarters of all AI referral traffic in SE Ranking's data, so give it its own monthly read.

• Homepage share. SE Ranking found that 60% of AI-referred visits land on homepages, against 17% from organic search, and that homepages were the biggest driver of ChatGPT's May 2026 jump, which analysts tie to brand names in answers becoming clickable. Value ChatGPT at the channel level, not only through the posts it links.
• Tag health. Count chatgpt.com sessions by medium. Any (not set) rows left outside your AI channel mean the source rule isn't matching.
• Quality against organic. Compare engagement rate and session key event rate over the same dates, for the same key event.

In Data Studio (formerly Looker Studio)

Google renamed Looker Studio back to Data Studio on 16 April 2026. You have two options there.

Use your channel group. Connect the property with Google's Analytics connector. After you create or edit the group, refresh the data source's fields and the custom channel group appears as a dimension, as Analytics Mates shows.

Rebuild it as a calculated field. This works even if someone deletes the group in GA4:

CASE
  WHEN REGEXP_MATCH(Session source, r"^((www\.)?perplexity\.ai|chatgpt\.com|chat\.openai\.com|claude\.ai|gemini\.google\.com|copilot\.microsoft\.com|(www\.)?copilot\.com|(copilot|m365)\.cloud\.microsoft|(www\.)?meta\.ai|chat\.mistral\.ai|chat\.deepseek\.com|grok\.com|(www\.)?you\.com|(www\.)?phind\.com)$")
    OR Session medium = "ai-assistant" THEN "AI assistants"
  WHEN Session default channel group = "Organic Search" THEN "Organic Search"
  ELSE "Everything else"
END

Data Studio's REGEXP_MATCH is a full match that uses RE2, and the r before the pattern saves you from escaping every backslash twice. Build the page from four scorecards, a weekly time series split by the new field, and tables by source and landing page. Google's connector notes say GA4 segments and comparisons aren't available through the connector.

In BigQuery: the raw export

GA4's export schema holds three traffic-source records, and each answers a different question.

Custom channel groups aren't in the export, so rebuild the rule in SQL. BigQuery's REGEXP_CONTAINS is a partial match that uses RE2, which is why the regex keeps its ^ and $.

-- AI assistants vs Organic Search, last 28 full days
WITH events AS (
  SELECT
    user_pseudo_id,
    (SELECT value.int_value FROM UNNEST(event_params) WHERE key = 'ga_session_id') AS session_id,
    event_name,
    (SELECT COALESCE(value.string_value, CAST(value.int_value AS STRING))
       FROM UNNEST(event_params) WHERE key = 'session_engaged') AS session_engaged,
    LOWER(COALESCE(session_traffic_source_last_click.cross_channel_campaign.source,
                   session_traffic_source_last_click.manual_campaign.source)) AS source,
    LOWER(COALESCE(session_traffic_source_last_click.cross_channel_campaign.medium,
                   session_traffic_source_last_click.manual_campaign.medium)) AS medium
  FROM `your-project.analytics_123456789.events_*`
  WHERE _TABLE_SUFFIX BETWEEN
    FORMAT_DATE('%Y%m%d', DATE_SUB(CURRENT_DATE(), INTERVAL 28 DAY))
    AND FORMAT_DATE('%Y%m%d', DATE_SUB(CURRENT_DATE(), INTERVAL 1 DAY))
),
sessions AS (
  SELECT
    user_pseudo_id,
    session_id,
    MAX(source) AS source,
    MAX(medium) AS medium,
    LOGICAL_OR(session_engaged = '1') AS engaged,
    LOGICAL_OR(event_name IN ('sign_up', 'generate_lead')) AS converted  -- your key events
  FROM events
  WHERE session_id IS NOT NULL
  GROUP BY user_pseudo_id, session_id
)
SELECT
  CASE
    WHEN REGEXP_CONTAINS(source, r'^((www\.)?perplexity\.ai|chatgpt\.com|chat\.openai\.com|claude\.ai|gemini\.google\.com|copilot\.microsoft\.com|(www\.)?copilot\.com|(copilot|m365)\.cloud\.microsoft|(www\.)?meta\.ai|chat\.mistral\.ai|chat\.deepseek\.com|grok\.com|(www\.)?you\.com|(www\.)?phind\.com)$')
      OR medium = 'ai-assistant' THEN 'AI assistants'
    WHEN medium = 'organic' THEN 'Organic Search'  -- close to GA4's Organic Search
    ELSE 'Everything else'
  END AS channel,
  COUNT(*) AS sessions,
  ROUND(COUNTIF(engaged) / COUNT(*), 3) AS engagement_rate,
  COUNTIF(converted) AS converting_sessions,
  ROUND(COUNTIF(converted) / COUNT(*), 4) AS session_key_event_rate
FROM sessions
GROUP BY channel
ORDER BY sessions DESC;

A second query audits what actually arrives. It reads the referrer and the collected tag on each session_start event, which carries page_referrer per Google's event list.

-- Audit: referrer hosts and collected tags on session starts
SELECT
  NET.HOST((SELECT value.string_value FROM UNNEST(event_params)
            WHERE key = 'page_referrer')) AS referrer_host,
  collected_traffic_source.manual_source AS tagged_source,
  collected_traffic_source.manual_medium AS tagged_medium,
  COUNT(*) AS session_starts
FROM `your-project.analytics_123456789.events_*`
WHERE _TABLE_SUFFIX BETWEEN '20260901' AND '20260927'
  AND event_name = 'session_start'
GROUP BY referrer_host, tagged_source, tagged_medium
ORDER BY session_starts DESC
LIMIT 200;

Scan it for assistant hosts. A tagged source of chatgpt.com with a null tagged medium is the tag-without-medium case from earlier. An assistant's referrer host with no tag is what your source rule has to catch.

A third query uses traffic_source to count users whose first visit came from an assistant. Run it on daily tables only.

-- New users first acquired by an AI assistant
SELECT traffic_source.source AS first_source, COUNT(DISTINCT user_pseudo_id) AS new_users
FROM `your-project.analytics_123456789.events_*`
WHERE _TABLE_SUFFIX BETWEEN '20260901' AND '20260927'
  AND event_name = 'first_visit'
  AND (traffic_source.medium = 'ai-assistant'
       OR REGEXP_CONTAINS(traffic_source.source, r'^((www\.)?perplexity\.ai|chatgpt\.com|chat\.openai\.com|claude\.ai|gemini\.google\.com|copilot\.microsoft\.com|(www\.)?copilot\.com|(copilot|m365)\.cloud\.microsoft|(www\.)?meta\.ai|chat\.mistral\.ai|chat\.deepseek\.com|grok\.com|(www\.)?you\.com|(www\.)?phind\.com)$'))
GROUP BY first_source
ORDER BY new_users DESC;

Swap in your own project, dataset and key event names, and expect small gaps against the GA4 interface: Google says the interface estimates sessions from unique session IDs.

Does AI Referral Traffic Convert Better? What Studies Say

These numbers don't agree because they don't measure the same thing. Ahrefs is one site with a strong brand, and Clarity measures sign-up prompts on publisher sites. Amsive and Adobe conflict most directly: Amsive compares AI with organic search across many sites and finds no reliable gap, in B2B or B2C, while Adobe finds a large premium against all non-AI traffic, including email and paid search, in retail only. A year earlier, the same Adobe series showed AI visitors converting worse.

So there's no standard premium for AI referral traffic to plug into a forecast. Report your own AI channel against your own Organic Search, for one key event, over enough sessions. Amsive's stricter sample required at least 50 AI sessions and 5 AI conversions per site, and even that is a low bar.

The Weekly AI Referral Traffic One-Pager

A one-page report you can rebuild in GA4 or Data Studio in an hour: six cards, each with a fixed dimension and one question.

Read cards 1, 2 and 5 every week. Cards 3 and 4 move slowly, so judge them monthly, once the channel has enough sessions to trust. Recheck the regex each quarter and log every rule change in the notes card, since a new rule changes past numbers too. For the monthly view across crawls, citations, clicks and conversions, use the report template in our GEO measurement guide.

Where Rankbox Fits in This Setup

Rankbox doesn't track AI citations or analyze your analytics today. Everything above runs on GA4, Data Studio and BigQuery. To see how often assistants name you before anyone clicks, you need a tracker; our ChatGPT rank tracker roundup compares them.

What Rankbox does is the step before the click. Answer-Space Research maps the questions your buyers ask ChatGPT, Perplexity and Google, and the Citation-Ready Writer turns them into source-backed articles built to be quoted. Your AI assistants channel then shows whether those pages earn AI referral traffic that converts. See plans and pricing.