ChatGPT sent you a customer. Your analytics says Direct.

A local business does not need a perfect AI attribution model to tell whether AI assistants are sending people to its site. It needs a defensible answer to one question: is this working, or am I guessing?

Why AI visits end up in Direct

A visit is attributed to a source in one of two ways: the browser tells your site where the click came from (the referrer), or the link carries a tag such as utm_source. When neither arrives, GA4 usually has no source to attribute, so the session may appear as Direct or Unassigned. Search Engine Land reported on October 15, 2026 (we cover the details in this week’s Update) that ChatGPT links appear to have dropped their tag and now suppress referral information, with ChatGPT-labelled visits down about 31,000 a day and Direct up about 26,000 a day across roughly 500 pages. INDEPENDENT RESEARCH OpenAI has not confirmed the change is intentional. INDEPENDENT RESEARCH

The problem predates that change. Practical Ecommerce reports that ChatGPT often adds a noreferrer attribute, which suppresses referral information and sends some visits to Direct or Unassigned. INDEPENDENT RESEARCH Clickport adds a useful qualification: a missing referrer does not always mean GA4 will show Direct, because a URL tag can preserve the source and GA4 can retain an earlier one. INDEPENDENT RESEARCH So what lands in Direct is a mix: some AI clicks, some typed addresses, and a lot of everything else.

The four layers

No single number proves that a lead came from an AI assistant. Four partial signals, considered together, can inform a decision. Think of them as a floor, a hint, a testimony and a trend.

Layer 1: build the floor

In GA4, create a custom channel group so AI assistants are easier to distinguish from Referral. Go to Admin, then Data display, then Channel groups, create a new group, add a channel named AI assistants, and set the condition to Source matches regex with a pattern like chatgpt|openai|perplexity|gemini|copilot|claude. Place it above Referral so it takes precedence. The menu names can differ slightly by interface version, so check them against what you see. DHM INTERPRETATION

Read the result as a minimum. After the October change, a drop in tagged visits does not mean AI sent fewer people; it may mean fewer visits are labelled. Annotate the date so nobody misreads the drop later.

Layer 2: read Direct by landing page

Direct sessions on your homepage are mostly people who typed your name, bookmarked you, or clicked from an untracked app. Direct sessions on a deep page, such as a long guide or a city service page, are different: few people would type that address. In GA4 Explorations, build a table of Direct sessions by landing page and look for deep pages receiving them. This is inference, not proof, and email apps, messaging apps, PDFs and QR codes also produce Direct visits. DHM INTERPRETATION

Layer 3: ask every lead

This layer remains useful when platforms change how they tag links. Add one required question to every form: “How did you hear about us?” with options for Google search, Google Maps, an AI assistant such as ChatGPT or Gemini, a referral, social media and other. Put the same question in your phone script and write the answer into your CRM or a spreadsheet. Self-reported source has its own bias, since people forget or credit the last thing they saw, but it is the only layer that records what the customer actually did.

Layer 4: watch the demand signals

People who hear a business name from an assistant may search for it next. Track branded queries in Google Search Console, along with calls, direction requests and website taps in your Business Profile performance report. These are signals to watch, but they do not show that an assistant caused the search. A real example from our own Search Console, July 16 to October 13, 2026: the homepage query with the most clicks was a variant of our own name, “digital handshake mmc”: 5 clicks from 23 impressions at an average position of 5.8, with the next two queries earning one click each. DHM DATA That is branded demand arriving through a name variant. It is the kind of signal to watch, though we cannot say an assistant caused it.

What our own numbers show

In GA4 between July 16 and October 15, 2026, our site recorded 86 sessions. Of those, 54 (63%) were Direct, 22 were Google organic, and none carried a source of chatgpt.com, Perplexity or Gemini. DHM DATA Direct sessions engaged 26% of the time and organic sessions 45%. At this size, the layers do their job as a checklist rather than as a measurement: our sample is too small to separate a typed address from an AI click. That is the honest position at our size. DHM DATA

What not to do

• Do not add up Direct sessions and call the total AI traffic. Direct is a mix, and its AI share is unknown.
• Do not credit a Direct spike to AI with no corroboration. Look for the same movement in branded search, profile calls and the source question.
• Do not pay for an “AI visibility” report before you have a source question. A tool can show what an assistant says. Only a customer can tell you what they did. Being named in the first place is a separate job, covered in how to get ChatGPT to recommend your business and in our note on entity confusion.
• Do not rebuild your reports every time a platform changes its tags. Move your measurement to things you control: your forms, your calls, your CRM.

None of this changes how local SEO works, and none of it guarantees AI mentions. We do not guarantee rankings, and the same rule applies to assistants. What it changes is whether you can tell, with a straight face, what is working. Rankings still need their own measurement, which is why we scan them as a geo-grid. The facts behind every platform’s record of you should also stay identical, as in our guide to business info consistency.