ChatGPT Is Sending You Shoppers. Your Analytics Is Probably Filing Some of Them Under Direct

Traffic from AI assistants to retail sites grew almost 700 percent in the 2025 holidays, and it converts well. Here is how those clicks reach GA4, where they get misfiled, what the October 2026 ChatGPT tagging bug showed, and a measurement setup that survives the next change.

Shoppers now ask ChatGPT, Perplexity, Gemini and Claude which product to buy, then click through to a store. Adobe measured the shift across U.S. retail sites: during the 2025 holiday season, traffic to retail sites from generative AI tools rose 693.4 percent year over year, and visitors arriving that way converted 31 percent more than visitors from other sources. Adobe also notes the base is still modest. For most stores this is a small channel growing quickly, which is exactly the kind of channel that gets mismeasured.

The hard part is not counting the visits you can see. It is knowing how many you cannot. This post covers how AI assistant clicks reach your analytics, where they get misfiled, what happened in October 2026 when a large share of them vanished from reports for a week, and how to build a measurement that survives the next change.

Three ways an AI click can arrive

A visit from an assistant carries one of three signals: a referrer header naming the assistant’s domain, a UTM tag added to the link, or nothing at all. Which one you get depends on the assistant, the surface (web or app) and sometimes the week.

Where an AI assistant click lands in GA4 Three paths. A click with a referrer from a known assistant lands in the AI Assistant channel. A click with utm_source set but no medium can land in Unassigned. A click with no referrer and no tag lands in Direct. Shopper clicks a link in an assistant Referrer: chatgpt.com, perplexity.ai, gemini… utm_source=chatgpt.com but no utm_medium No referrer, no tag (in-app browser, bugs) AI Assistant channel (visible) Unassigned (visible if you look) Direct (hidden)
The same assistant can send traffic down all three paths. Only the top one is reported as AI traffic by default.

GA4’s default channel group now includes an AI Assistant channel. Google defines it as traffic from sources like ChatGPT, matched when the medium is ai-assistant or when the referrer is on Google’s list of assistants. It excludes Google’s own AI Overviews and AI Mode, which stay inside organic search.

The middle path is the one most stores miss. OpenAI appends utm_source=chatgpt.com to cited links but no utm_medium, and a UTM hit without a medium can be grouped as Unassigned rather than Referral. In practice chatgpt.com can appear under several channels in the same report, so filter by session source, not by channel, when you want the full ChatGPT count.

What each assistant passes

Assistant and surfaceReferrerUTM tagTypical landing in GA4
ChatGPT, webUsually chatgpt.comOften utm_source=chatgpt.comAI Assistant or Unassigned
ChatGPT, mobile appsAlmost neverDepends on the linkDirect unless tagged
Perplexityperplexity.aiNoAI Assistant
Geminigemini.google.comNoAI Assistant
ClaudeRarelyNoMostly Direct
Compiled from Terminus and Seer Interactive testing. Behaviour changes without notice; verify on your own data.

The week ChatGPT traffic “fell by half”

On October 9, 2026, ChatGPT referral sessions dropped sharply across many sites. LLM Pulse measured a fall of about 46 percent between October 9 and 11 while total traffic stayed flat and Perplexity and Claude referrals held steady. Adobe’s Cedric Huesler traced it to links opening with no referrer and no UTM tag, so the visits were filed under Direct. Search Engine Roundtable later updated its report: Huesler confirmed it was a bug, and ChatGPT was tagging its links again by October 15.

Nothing about shopper behaviour changed that week. Only the label did. A store reading its channel report at face value would have concluded that AI shoppers disappeared, and possibly cut content work aimed at them. The lesson is general: a channel figure is the product of demand and of a tagging convention controlled by someone else. When the convention changes, the figure moves.

Catching this kind of break is a classic anomaly detection problem: a series that drops while a related series (Direct sessions on the same landing pages) rises by a similar amount. A simple weekly alert on both series would have flagged it within days.

A measurement that survives the next change

1. Count by source, then group

Build a custom channel group with an AI channel matched on session source by regex (for example chatgpt|openai|perplexity|gemini|claude|copilot|deepseek) and place it above Referral and Unassigned. Google’s documentation notes that traffic goes into the first channel whose definition it matches, so order matters, and custom groups apply to historical data as well.

2. Record the source on the order

Browser analytics loses visitors to consent and blockers; your order table does not. WooCommerce has shipped built-in Order Attribution since version 8.5. It stores origin, source type, UTM values and device on each order, and it is session based and last click, so treat it as a second opinion rather than the truth. On Magento, capture the landing URL and referrer into order attributes at checkout. Either way you can compute AI-sourced revenue from the ledger instead of from a sample.

3. Estimate what is hidden in Direct

People rarely type a deep product URL. Direct sessions that land on a specific product or comparison page, from a new visitor, on mobile, are disproportionately untagged clicks from apps, messages and assistants. Track the ratio of “Direct to deep page” sessions to all sessions over time. A baseline that rises while tagged AI traffic grows is a reasonable upper-bound signal of hidden assistant traffic. It is an estimate, so label it as one.

4. Ask the customer

A one-question post-purchase survey (“Where did you first hear about us?”) with an “AI assistant (ChatGPT, Perplexity, etc.)” option catches the visits no tag can. Response rates are partial, but the share answering “AI assistant” is a direct check on whether your tagged numbers are in the right range.

Which numbers to watch

MetricSource of truthWhy it matters
AI sessions by assistantGA4, custom channel groupTrend and mix; sensitive to tagging changes
AI-sourced orders and revenueOrder attribution in the store databaseNot affected by consent or blockers
Conversion rate and average order valueStore orders divided by GA4 sessionsTests whether the Adobe pattern holds for your store
New customer shareStore databaseAssistants may introduce you to people who never saw your ads
Landing pages for AI sessionsGA4Shows which pages assistants cite, and so what content earns citations
Direct to deep page ratioGA4Early warning when tagging breaks
Survey share answering “AI assistant”Post-purchase surveyIndependent check on the tagged numbers

A session source dimension filtered to the assistant domains is enough to put this channel on a weekly report next to your store’s order attribution.

Sources