Referral traffic in Google Analytics: how to track AI and direct visits in GA4

Referral traffic in Google Analytics is traffic from another site or app that GA4 can identify as the source of a visit. It is useful because it shows who is sending people to your site, which pages they land on, and whether those visits turn into engagement, carts, purchases, or pipeline.

But referral traffic is no longer just an affiliate, PR, partner, or publisher report. AI assistants now send shoppers and buyers to websites too. Some of those visits arrive with a visible referrer, such as ChatGPT, Perplexity, Gemini, Claude, or Copilot. Others arrive with no usable source data and fall into Direct.

That distinction matters for ecommerce teams. A shopper might ask an assistant to compare products, click a recommended product detail page, and arrive ready to buy. If the app or browser does not pass a referrer, GA4 may treat the visit as (direct) / (none). The visit still happened. The source signal did not.

This guide shows how to read referral traffic in GA4, separate referral from Direct, track known AI assistant visits, segment likely assistant-driven Direct visits, and connect the results to product and channel performance without pretending attribution is more certain than it is.

What referral traffic means in Google Analytics#

In GA4, referral traffic is traffic that reaches your site from another website or app when that outside source is visible to Analytics. A visitor clicks a link on another domain, arrives on your site, and GA4 records a source and medium for the session.

The most useful dimensions are:

• Session source / medium: the sending source and traffic medium, such as example.com / referral.
• Session source: the source alone, such as example.com.
• Session medium: the medium alone, such as referral, organic, cpc, or ai-assistant when GA4 identifies supported AI assistant traffic.
• Session default channel group: GA4's grouped channel label, such as Referral, Direct, Organic Search, Paid Search, or AI Assistant.
• Landing page + query string: the first page in the session.
• Page referrer: the referring page URL when GA4 has it available.

A normal referral report might show traffic from review sites, partners, affiliates, press mentions, marketplaces, forums, and tools that link to your site. For a commerce brand, the next question is not only "who sent traffic?" It is "which products or categories did that traffic reach, and did those visits create revenue?"

Referral traffic vs direct traffic in GA4#

Referral traffic and Direct traffic are easy to confuse because Direct often contains visits that were not truly direct in the everyday sense.

Google Analytics describes (direct) / (none) as traffic with no clear referral source. Some of that traffic comes from people who typed your URL or used a bookmark. Some of it comes from links where the original source was lost.

The technical reason is simple. Analytics depends on signals such as referrers and campaign tags. The HTTP Referer header can tell a server which page sent a request. Referrer-Policy can reduce or omit that information. Apps, webviews, copied links, redirects, privacy tools, and missing UTMs can all break the trail.

So Direct is partly a channel and partly an unknown-source bucket. Treat it as a signal to investigate, not a clean label for brand loyalty.

How to find referral traffic in GA4#

Start with the standard acquisition report.

1. Open Reports.
2. Go to Acquisition.
3. Open Traffic acquisition.
4. Change the primary dimension to Session source / medium.
5. Use the table search or filter to find referral.
6. Add Landing page + query string as a secondary dimension.
7. Compare sessions, engaged sessions, key events, add-to-cart events, purchases, revenue, and revenue per session.

That view answers the first set of questions: which sources sent referral traffic, where did those users land, and whether the sessions were useful.

For deeper analysis, create an Exploration:

1. Go to Explore.
2. Create a blank exploration.
3. Add dimensions such as Session source / medium, Page referrer, Landing page + query string, Device category, and Session default channel group.
4. Add metrics such as Sessions, Engaged sessions, Key events, Add-to-carts, Purchases, Total revenue, and Average purchase revenue.
5. Filter to referral sources, AI assistant sources, or Direct segments depending on the question.

For ecommerce, always add the landing page or page type. A referral that lands on a product detail page is different from one that lands on a blog post, homepage, store locator, or account page.

Clean up referral data before you analyze it#

Referral data is only useful if the report is not polluted by sources that should not get credit.

Common unwanted referrals include:

• payment gateways
• checkout domains
• authentication providers
• your own domains and subdomains when cross-domain tracking is wrong
• password reset or account-management flows
• support or review widgets that send users away and back

Google Analytics has a List unwanted referrals setting for domains that should not start a new referral session. MeasureSchool's GA4 referral guide also recommends checking unwanted referrals before acting on referral data.

This matters because ecommerce conversion credit can be fragile. If a customer leaves your site to pay through a third-party gateway and returns to your confirmation page, you do not want the payment provider to appear as the source of the purchase. If your checkout, login, or subdomain setup is wrong, a referral report can tell a false story.

Also check the basics:

• owned email, SMS, influencer, affiliate, QR, and PDF links use consistent UTM parameters
• redirects preserve query strings and campaign tags
• paid campaigns use the expected source and medium
• internal links do not carry UTMs that overwrite the real source
• site migrations or consent-banner changes did not disrupt tagging

Clean what you can before modeling what you cannot.

How AI assistants change referral traffic reporting#

AI assistants create two kinds of measurement signals.

The first is visible traffic. If an assistant passes a referrer or GA4 recognizes the source, the visit may appear as a referral, a known AI source, or GA4's AI Assistant channel. MarTech reported in May 2026 that GA4 now classifies supported AI assistant traffic with medium ai-assistant and an AI Assistant channel when the source is detectable.

The second is hidden or under-attributed traffic. A link from an assistant app, mobile webview, copied answer, privacy-controlled browser, or branded-search follow-up may arrive without a visible AI source. GA4 may classify it as Direct, Organic Search, or another last-click channel.

Retailgentic's Dark Agentic Commerce Traffic article put a name on this ecommerce-specific problem: answer engines can send high-intent shoppers to product detail pages while analytics credits the visit somewhere else. Catalog's dark traffic in agentic commerce guide goes deeper on the full DACT model.

The practical takeaway for GA4 reporting is this: known AI referrals are the floor, not the whole channel.

The floor is still worth measuring. Adobe reported that generative AI traffic to U.S. retail sites increased 1,200% from July 2024 to February 2025, with AI-referred retail visitors showing stronger engagement indicators than non-AI traffic sources. Search Engine Land reported a 2026 Previsible dataset of 6.77 million LLM-driven sessions across 166 GA4 properties where ChatGPT accounted for 92.4% of trackable LLM referral traffic. That data also noted ecommerce AI traffic often lands on product pages.

Those are not your numbers. They are reasons to build the report before the channel is large enough to surprise you.

Build a known AI referral view in GA4#

Start with what GA4 can see.

1. Check the AI Assistant channel#

In Traffic acquisition, set the primary dimension to Session default channel group and look for AI Assistant. If it exists, compare it with Referral, Direct, Organic Search, Paid Search, Email, and Organic Social.

Then switch to Session source / medium to see which tools are actually sending traffic.

2. Create a custom AI assistant report or channel group#

GA4's native classification may not catch every AI source, and coverage can change. A custom view is still useful.

Common AI sources to include:

• chatgpt.com
• chat.openai.com
• perplexity.ai
• gemini.google.com
• claude.ai
• copilot.microsoft.com
• meta.ai
• you.com
• poe.com

A starter regex for GA4 filters or a custom channel group:

(chatgpt\.com|chat\.openai\.com|perplexity\.ai|gemini\.google\.com|claude\.ai|copilot\.microsoft\.com|meta\.ai|you\.com|poe\.com)

If you create a custom channel group, place the AI channel above generic Referral so AI sources are grouped before GA4 falls through to the broader referral rule.

3. Analyze quality by landing page#

For known AI traffic, do not stop at sessions. Add:

• landing page
• page type
• product category
• device category
• engaged-session rate
• add-to-cart rate
• purchase rate
• revenue per session
• product or item revenue

The question is not only "did ChatGPT send traffic?" The better question is "which product pages did AI tools send people to, and did those visits behave like qualified shoppers?"

Segment likely assistant-driven Direct traffic#

You cannot prove that a single Direct session came from an AI assistant if the source was never sent. You can build a useful likely segment.

Start with a Direct segment:

• Session source / medium equals (direct) / (none)
• user is new, first-time, or not recently active
• landing page is a product page, category page, collection page, comparison page, buying guide, or other high-intent page
• landing page is not the homepage, login, account, support, careers, checkout return, or store locator
• session includes meaningful engagement, add-to-cart, checkout, purchase, or another high-intent event

Then compare it with a baseline period. The baseline should be before the AI visibility change you are trying to measure. That might be before a product started appearing in answer engines, before a content launch, before an AI shopping feature launch, or before known AI referrals began showing up in GA4.

Layer in cross-checks:

• known AI assistant referrals for the same period
• answer-engine visibility for the same products and queries
• Search Console branded and product-name queries
• non-branded organic traffic
• paid, email, affiliate, influencer, and SMS campaign calendars
• product launches, promotions, PR, and marketplace activity
• device and browser patterns
• server logs, when available, to confirm whether requests had referrers

This segment should be labeled carefully. Call it "likely assistant-influenced Direct" or "Direct product-page investigation segment." Do not call it exact AI traffic.

Tie visits to product and channel performance#

A useful GA4 report for AI-era referral traffic has three layers.

1. Channel layer#

Compare:

• known AI Assistant / AI referral traffic
• Referral traffic
• Direct traffic
• likely assistant-influenced Direct segment
• Organic Search
• branded Organic Search, if you separate it
• Paid Search and Paid Social
• Email, SMS, affiliate, and influencer campaigns

This keeps AI in context. A visible AI channel may be small but high quality. Direct may be large but mixed. Organic may still close the purchase after AI created demand.

2. Product and page layer#

Break traffic down by:

• landing page
• page type, such as product detail page, category, collection, comparison, blog, or homepage
• product category
• item name, item ID, SKU, or variant, if ecommerce events are implemented cleanly
• canonical product URL

For a commerce team, page type is often the biggest clue. A Direct session landing on /products/wide-fit-trail-running-shoe-blue-size-10 is less likely to be typed manually than a Direct session landing on /.

3. Outcome layer#

Use outcome metrics that match the page:

• engaged sessions
• add-to-cart events
• checkout starts
• purchases
• total revenue
• item revenue
• revenue per session
• average order value
• key events or lead events for non-transactional pages

A simple product-channel matrix can make the report easier to read:

When you report the result, use ranges:

• Known AI floor: sessions and revenue from visible AI Assistant / AI referral sources.
• Likely dark estimate: Direct product-page sessions above baseline that match the segment and survive campaign/technical checks.
• Influence signals: branded search, product-name search, visibility changes, and revenue movement that align with AI visibility but cannot be assigned to a single click.

That structure keeps the analysis useful without pretending GA4 can reconstruct source data it never received.

Why product data readiness matters#

Referral traffic analysis tells you what reached the site. Product data determines whether AI assistants can understand products well enough to recommend them and link to the right pages in the first place.

For ecommerce brands, assistant-readable product data should include:

• canonical product URLs
• clean product titles and descriptions
• product category and collection context
• variant-level price and availability
• sizes, colors, materials, compatibility, fit, use case, and other attributes
• product identifiers
• shipping, returns, and warranty policies
• review and rating signals
• current images and specifications

This matters for measurement too. If assistants link to canonical product detail pages, your GA4 landing-page patterns are easier to read. If product data is fragmented across duplicate URLs, stale feeds, marketplace copies, inconsistent variants, and missing attributes, both recommendation quality and measurement quality get worse.

That is where Catalog fits. Catalog is the product data layer for AI commerce. It helps brands make product data structured, current, and machine-readable so AI shopping surfaces can understand products without replacing the existing storefront.

For the broader visibility side of this work, see Catalog's guides to AI visibility for ecommerce, how to make products show up in ChatGPT, and agentic commerce. The measurement loop is the other half: once assistants can understand and recommend your products, GA4 needs clean enough product, page, and channel reporting to show what happens next.

Common mistakes to avoid#

Treating every Direct session as AI traffic#

Direct traffic has many causes. Some of it is real direct demand. Some is untagged marketing. Some is dark social. Some may be AI-assisted. Your job is to isolate a likely segment and explain the assumptions.

Counting visible AI referrals as the whole channel#

If GA4 shows 200 AI Assistant sessions, that is useful. It is not proof that AI influenced only 200 sessions. App traffic, copied links, privacy settings, and branded-search follow-ups can hide part of the journey.

Forgetting payment and auth referral cleanup#

Referral reports can be polluted by checkout, payment, and login flows. Clean those before reporting referral revenue or AI referral quality.

Comparing channels without product-page context#

Sitewide channel averages hide the story. AI and likely dark assistant traffic may concentrate on product pages, category pages, comparison pages, or buying guides. Report by page type and product group.

Waiting for perfect attribution#

Perfect attribution is not coming soon. The better operating model is to measure known traffic, model likely hidden traffic, compare it with visibility and product performance, and keep improving the product data that makes AI commerce measurable.

FAQ#

Is ChatGPT traffic referral traffic in GA4?#

Sometimes. If the visit arrives with a visible ChatGPT referrer or GA4 classifies it as AI Assistant traffic, you can measure it as known AI traffic. If the visit comes from an app, copied link, webview, or another context that strips source data, it may appear as Direct or another channel.

Does GA4's AI Assistant channel solve dark AI traffic?#

No. It helps with detectable AI assistant traffic, but it does not solve visits where no referrer or campaign data reaches GA4. Keep using Direct product-page segments and visibility cross-checks for the likely dark portion.

What is the best GA4 segment for likely AI-driven Direct traffic?#

Start with (direct) / (none) sessions from new or first-time users landing on product, category, comparison, or buying-guide pages. Exclude homepage, login, account, support, checkout return, and careers pages. Then require meaningful engagement, add-to-cart, checkout, purchase, or another high-intent event, and compare the segment against a baseline.

Should ecommerce teams use UTMs for AI traffic?#

Use UTMs anywhere you control the link, including owned content, email, SMS, affiliate, influencer, QR, PDF, and partner links. You usually cannot control links generated by third-party AI assistants, so UTMs will not solve dark AI traffic by themselves. They do reduce avoidable Direct traffic and make the remaining unknown bucket easier to interpret.

The takeaway#

Referral traffic in Google Analytics still matters. It shows which outside sources send people to your site and how those visits perform. But in an AI-commerce journey, the referral report is only one part of the picture.

Measure visible AI assistant traffic. Clean up referral pollution. Segment likely assistant-driven Direct visits. Compare everything by product page, channel, and revenue outcome. Then use answer-engine visibility and structured product data to explain why those visits happened.

The goal is not perfect attribution. The goal is a measurement loop that is honest enough to act on.