ChatGPT Sent Your Customers Last Week. Your Analytics Called Them Direct.

Hi folks. Let me start with a small mystery from the week of October 8.

On Thursday, analytics people started noticing that ChatGPT's referral traffic was wobbling. By Friday, October 9, it had fallen off a cliff. LLM Pulse, which tracks AI referral data from the GA4 accounts of the sites it monitors, found ChatGPT referral sessions down about 46% from October 9 to 11, compared with the same days in the two prior weeks. Total traffic on those same sites? Down about 2%. More than eight in ten sites lost over 30% of their ChatGPT referrals (LLM Pulse: https://llmpulse.ai/blog/chatgpt-referral-traffic-fell/).

So did people suddenly stop clicking on ChatGPT's links? Not quite. Hold that thought, because the answer is more interesting than the question, and it has a very practical lesson for anyone who runs a service business.

Where the visitors went

Cedric Huesler at Adobe looked at roughly 500 pages where ChatGPT had been driving at least 5% of visits. Comparing October 9 to 13 with the week before, ChatGPT visits on those pages fell by about 31,000 a day. Direct visits rose by about 26,000 a day on the same pages. Direct went up on two out of three of the pages that lost ChatGPT traffic. Search traffic, meanwhile, dipped only about 7% (Search Engine Land: https://searchengineland.com/chatgpt-appears-to-drop-utm-tracking-making-ai-referral-traffic-harder-to-measure-496552).

Read those numbers side by side and the picture comes into focus. The people didn't vanish. They changed name tags. Many links in the new GPT-6 experience arrived without the tracking parameter (utm_source=chatgpt.com) and without a referrer, so analytics had no idea who sent them and filed them under "Direct," the junk drawer of web traffic.

Is it fixed? Huesler updated his LinkedIn post on October 15 to say it was a bug and that ChatGPT is tagging its links again, and Barry Schwartz of Search Engine Roundtable noted the tags seemed to be returning (Search Engine Roundtable: https://www.seroundtable.com/utm-parameters-missing-chatgpt-42299.html). OpenAI has not publicly commented, as far as I can find, and LLM Pulse's own write-up stops short of calling it a bug. So the careful version is: it looks like a tagging glitch, it looks like it's being repaired, and nobody official has confirmed either. A few other caveats worth having in your back pocket. LLM Pulse doesn't say how many sites are in its sample, and Huesler's 500 pages are ones that already got heavy ChatGPT traffic, which means they are mostly not local service businesses. And LLM Pulse sells AI visibility tracking, so it has a horse in this race.

Why a plumber or a landscaper should care

Nothing about this glitch is aimed at contractors. But it exposes something that is true every ordinary week: the number in your analytics for "AI" is a floor, not a count.

Consider how a homeowner actually uses these tools. She asks an assistant who can handle a failing water heater this weekend. The assistant names two or three companies. She might click a link, or she might just remember a name and type it into Google later, or call the number she saw. In analytics, the first path might show up as ChatGPT, the second as Branded Search, and the third as nothing at all. This week's glitch simply made the first path invisible too.

I think about a landscaper I talk to who works residential jobs around DC. When we've talked, his single biggest issue is finding a consistent source of new leads. That's the question underneath all of this. It's one thing to want more leads. It's another to know, when one shows up, which door it walked through, because that's what tells you where to put the next dollar and the next Saturday. If a new door opens (AI assistants) and your instruments can't see it, you can't tell whether it's a trickle or a flood.

The instrument that doesn't break

Here is the unglamorous good news. There is a measurement tool that no platform update can touch: asking.

"How did you hear about us?" takes a few seconds on a phone call, one dropdown on a web form, and one line in a job-booking script. It has well-known flaws. People forget, or say "Google" for everything, or credit the neighbor who mentioned you six months ago. But its flaws don't change overnight, and they don't change because a chatbot shipped a new interface.

The upgrade for 2026 is small. Make sure your answer list includes something like "ChatGPT or another AI assistant," and keep "Google Maps," "Google search," "Facebook," "Angi or another directory," "Referral (who?)," and "Saw your truck" as separate choices. If AI assistants aren't on the menu, those customers get filed under "Other," or "Google," and the channel stays invisible to you.

Then use the self-reported answer and your analytics together, as two imperfect witnesses. If both say AI is sending you work, you can believe it. If analytics says AI is flat and callers keep mentioning it, analytics is probably undercounting (as it did last week). If callers never mention it and analytics shows a few visits, it's likely curiosity, not customers.

A ten-minute check you can do today

If you have Google Analytics 4, go to Reports, then Acquisition, then Traffic acquisition. Set the dimension to "Session source / medium" and search for "chatgpt." Compare October 9 to 15 with the week before. Then look at "(direct) / (none)" for the same dates.

If ChatGPT dipped and Direct rose, you probably saw this glitch with your own eyes. If nothing moved, that's useful too. It likely means AI isn't a meaningful source for you yet, or that your site gets too few visits to see it in a week.

Next, open your CRM, your call-tracking tool, or even the notebook by the phone, and look at how you record "lead source." If there's no AI option, add one. That's the whole project.

What this looks like when it's built in

This is also an example of why we like to put a thin data layer under the tools a business already uses. The thing we've built for clients like Black River Design and Build pulls the CRM, accounting, estimating, and ad data into one place, with one agreed definition for each metric, so questions like "which lead sources become paid jobs" have one answer, not four. A channel like AI search is exactly where it earns its keep, because the self-reported source from the CRM can sit next to the web analytics number, and a gap between them becomes something you can see rather than something you feel.

That doesn't require anything fancy to start. A spreadsheet with lead date, lead source (as told by the customer), and whether the job closed will tell you a surprising amount in 60 days.

One more thought

I find this episode oddly comforting. We're all being told that AI is changing how customers find businesses, and it is. But the measurement of it is still held together with tags and parameters that can fall off for a week without anyone noticing. Meanwhile, the oldest tool in marketing, a person asking another person a friendly question, kept working the whole time.

So here's my question for you: when a new customer calls, does anyone ask how they found you, and where does the answer go? I'd love to hear what your list of answer choices looks like. And if you'd like a second pair of eyes on where your leads actually come from, our Revenue Audit is a good place to start. No pressure, and no tracking parameters required.

Sources

LLM Pulse, "ChatGPT referral traffic fell" (Oct 13, updated Oct 16, 2026): https://llmpulse.ai/blog/chatgpt-referral-traffic-fell/

Search Engine Land, on Adobe's Huesler analysis (Oct 15, 2026): https://searchengineland.com/chatgpt-appears-to-drop-utm-tracking-making-ai-referral-traffic-harder-to-measure-496552

Search Engine Roundtable, UTMs returning to ChatGPT links (Oct 14, updated Oct 15, 2026): https://www.seroundtable.com/utm-parameters-missing-chatgpt-42299.html