How to find AI shopping traffic in GA4 after ChatGPT's tracking glitch

If your ChatGPT traffic looked like it fell off a cliff in the second week of October, it probably didn't. For about a week, ChatGPT links stopped carrying the tag that tells analytics tools where a visit came from, and those visits were counted as direct. It's fixed now. But it was a useful fire drill, because it showed how much of what brands call "AI traffic" depends on one parameter that a platform can drop without notice.

This guide is the setup we'd want in place before the holiday peak. It assumes you have Google Analytics 4 on your own store and sell through retailers too, which is the usual CPG situation. It's about getting honest numbers. Whether those numbers should drive your strategy is a separate argument we made in our piece on why AI traffic is small.

What happened in October

Search Engine Land reported on October 20 that ChatGPT had stopped appending utm_source=chatgpt.com to many outbound links. Cedric Huesler of Adobe compared about 500 pages where ChatGPT had accounted for at least 5% of visits, for October 9 to 13 against the week before. ChatGPT visits fell by about 31,000 a day. Direct visits rose by about 26,000 a day. Two-thirds of the pages that lost ChatGPT traffic gained direct traffic in the same period.

Show the data

Channel	Change in visits per day
ChatGPT (tagged referrals)	−31000
Direct	+26000

Source: Cedric Huesler (Adobe), reported by Search Engine Land, October 20, 2026. Figures are approximate, as reported. Search traffic to the same pages fell 7% over the period.

Huesler contacted OpenAI on October 14, and the company confirmed it was a known bug. A fix began rolling out early on October 15, and by 9 a.m. Pacific that day tagged ChatGPT clicks were back to about 95 to 100% of the level at the same hour a week earlier, according to the same report. Referrer information is still missing from most ChatGPT clicks, so the UTM tag is doing almost all the work.

Huesler's advice is sensible and we'd repeat it: annotate October 8 to 15 in your reports, don't change attribution models in response, and keep using utm_source=chatgpt.com as your main identifier.

Step 1: Build an AI channel group

GA4's default channel groups put most assistant traffic in Referral, mixed in with everything else. A custom channel group fixes that. Our measurement tip has the short version; here's the longer one.

1. In GA4, go to Admin, then Data display, then Channel groups, and create a new group by copying the default.
2. Add a channel called "AI assistants." Set the condition to Source matches regex, with a pattern like chatgpt|openai|perplexity|gemini\.google|copilot|claude.
3. Move the new channel above Referral in the list. GA4 assigns each session to the first channel whose rule it matches, so order matters.
4. Save, then open Traffic acquisition and switch the report to your new group.

Expect a small number. Marie Haynes' study of 100 sites, covered by Search Engine Land, found ChatGPT sent sessions equal to 0.62% of Google organic, and all assistants together 0.73%. Haynes also found those visitors converted at 13.8% against 3.02% for Google organic on the 55 sites tracking conversions, which is why the channel is worth separating even when it's tiny.

Remember what this channel can't see. Google AI Overviews and AI Mode clicks arrive as Google organic. An AI Overview is usually the most-read AI answer about a CPG product, and it won't show up here.

Step 2: Don't treat direct traffic as hidden AI

When AI numbers look small, someone on the team will suggest that the real AI traffic is hiding in direct. October gave that idea a little support: for a week, it literally was. But as a general rule the evidence points the other way.

Kevin Indig looked at 12 GA4 properties for Growth Memo. Eight showed growth in direct traffic from August to September, and much of it turned out to be bots. One site recorded 409,000 direct sessions in a period when only 198 were engaged.

Measure	Change
Total direct sessions	+140.9%
Engaged direct sessions	-1.0%
Share of direct sessions with an AI trace	0.003%

Source: Kevin Indig, Growth Memo, October 19, 2026.

In a separate analysis of four brands, Indig ran 632 tests and couldn't establish a reliable link between AI visibility and later branded search or direct traffic. He mentions a counterpoint worth knowing: Ramp's Eric Wu identified 30 to 40% of Ramp's direct traffic as coming from answer engines, by isolating direct visits to long blog URLs nobody would type and comparing their behavior with search visitors. Self-reported attribution confirmed those users came from AI.

The practical step is a bot check before anyone draws conclusions. In GA4, break direct traffic down by day, landing page, country, browser, browser version, operating system and screen size, the cuts Indig suggests. Spikes on one day, one odd screen resolution, or one country you don't ship to are almost always automated. Compare engaged sessions, not total sessions. If you also want to know which AI crawlers are hitting your pages, tip 31 covers the server logs.

Step 3: Ask customers

Analytics will always undercount AI influence for a CPG brand, because a shopper who asks an assistant for a snack recommendation will usually buy it at Target or on Amazon, not on your site. The cheapest way to see that influence is to ask.

Add a "How did you hear about us?" question to your post-purchase page or survey, with "ChatGPT or another AI assistant" and "Google (search or AI answer)" as separate options. Indig recommends this kind of self-reported attribution alongside analytics, and it's what confirmed Wu's finding at Ramp. Capture the shopper's own words where you can, including which assistant they used. It's imperfect, since people misremember, but it captures purchases that never touched your site's tracking.

If you sell mostly through retail, add the same question to any warranty registration, recipe sign-up or sampling form you run.

Step 4: Tag the links you control

More AI shopping will run through agents and carts that you don't control. You can still tag what passes through you.

• Universal Cart and agent handoffs. When Google hands a Universal Cart over to your store, it uses a link you provide. As we covered in our Universal Cart piece, adding UTM parameters to that link is how you'll tell those visits apart from ordinary Google traffic.
• Product feeds. Tag the landing page URLs in feeds sent to assistants and marketplaces, using a consistent utm_source per platform, so a feed visit doesn't land in direct.
• Platform reports. If you're on Shopify, the Agentic section reports sales, orders, online store sessions and conversion for each AI channel, and its sales figure combines referral sales with sales completed through direct checkout inside the assistant. Those direct checkout sales never touch GA4, so this report is the only place you'll see them.

Step 5: Put it on one page

The point of all this is a single monthly view, not five dashboards. We'd keep it to a short table:

• AI assistants channel in GA4: sessions, engaged sessions, conversions.
• Shopify Agentic sales and orders by channel, if you use it.
• Survey answers naming an AI assistant, as a share of responses.
• Direct traffic, engaged sessions only, with any bot spikes noted.
• Annotations for known breaks, starting with October 8 to 15.

Next to it, keep a record of what the assistants actually say about your top products, because a sale that happens at a retailer after a ChatGPT answer won't appear anywhere above. Our tip on checking ChatGPT shows the manual way. Some teams pay for tracking instead: XENA Intelligence, for one, puts shopper-style questions to ChatGPT, Perplexity, Gemini and Amazon's assistant and scores each product on how it appears, with the answers available in its SearchPanel.

What we'd skip

Two things we wouldn't do this quarter. Don't switch attribution models because one platform's tags broke for a week; you'll lose comparability for nothing. And don't build a forecast on AI referral growth. Haynes found ChatGPT referrals stopped growing exponentially in August 2025, and the October glitch shows how one platform change can move the line. Measure carefully, annotate the breaks, and judge the channel on a full quarter.