Google Analytics can now show visits that arrive from recognized AI assistants. It cannot tell you about every customer who saw your business in an answer, never clicked, and called later. To get a useful picture, compare the website record with what new customers tell you when they contact the business.

That distinction matters. A visit is evidence that someone reached the site from a recognized source. A customer saying ChatGPT led them to you is evidence from the sales conversation. Either can exist without the other. Neither should be stretched into a claim that cannot be checked.

Start with the AI Assistant channel in GA4

You can see the sessions that GA4 recognizes as coming from chatbot platforms in the new AI Assistant channel, under Reports > Acquisition > Traffic acquisition. This channel appears only if one of the supported platforms refers a visitor to your website. The AI Assistant channel differs; it appears only when there's a recognized referrer.

Google added the channel to its Default Channel Group in May 2026. When a recognized AI assistant supplies a referrer, Analytics assigns `ai-assistant` as the medium and `(ai-assistant)` as the campaign. Google's examples include ChatGPT, Gemini, Deepseek, Copilot and Grok. The session source can then identify the referring platform.

In the report, change the main dimension to Session source if you want to see the platforms separately. Then add Landing page as a secondary dimension or open a landing-page report with the AI Assistant channel selected. The useful questions are ordinary ones:

The key events section shows the actions your business has decided are important to what it does. If form submissions or bookings are not marked as key events, fix that before drawing conclusions from the channel. A traffic count without the actions that matter to the business is only a count of visits.

Google's own AI Overviews and AI Mode in Google Search do not appear in GA4's new AI Assistant channel. Instead, they continue to appear in Organic Search. That separation matters when you're reading the report. A rise or fall in the AI Assistant row does not describe every visit influenced by an AI-generated answer.

  • Which assistant sent the visit?
  • Which page did the visitor land on?
  • Did the session produce a form submission, booking or other action the business has marked as a key event?

A click report is only part of the evidence

A referral click proves that a session reached your site from a recognized source. It does not prove that an assistant recommended your business to a visitor and it does not show what prompt the assistant had when it did so. It doesn't count the answers that produced no click, nor does it identify phone calls made without visiting your site.

Consent choices and missing referrer information can also leave gaps. Those limits do not make the report useless. They define what the report can honestly support. It can show identifiable visits, their landing pages and the actions recorded during those sessions. It cannot reconstruct a customer's entire decision from one source label.

When traffic appears as Direct in GA4, it means that Analytics can't find any clear referral source for the session. Some AI-influenced visits may be hidden there, along with many other visits. Reclassifying Direct as AI would replace an unknown with a guess.

There is usually no need to build a custom AI channel now that GA4 has a built-in one. A custom channel can still be useful if the built-in channel is unavailable in the property or the business has a documented reason to include more sources. Custom rules also have an order, so a broader rule placed first can capture traffic before a later AI rule sees it. Keep the default report unless there is a specific measurement problem to solve.

Ask one question when a new lead arrives

Website analytics can't identify every AI-influenced lead. Phone-only enquiries and people who never click need a separate intake record.

Where possible, add a "How did you hear about us?" field to your web forms and ask the question by phone as well. What's important is that you preserve the customer's own wording. For a record like this, you can't claim perfect accuracy: it is a directional signal at best.

Do not make the field a long interrogation. A short optional list with choices such as Google, ChatGPT or another AI assistant, social media, referral, advertising and other can reduce the work. Include an open field or let staff record the exact answer when someone says more. The words "ChatGPT showed me your emergency service page" are more useful than a generic AI checkbox.

Staff need one consistent place to put the answer. That could be the contact record, booking system, call sheet or a plain spreadsheet. Record the date, lead, stated source and outcome. Do not paste private customer details into an analytics note.

Compare the website record with the lead record

Once you have a record of web sessions and a lead record, reconcile them. One record will show sessions recorded as visits to the website through AI Assistant. The other will show calls, forms and bookings where the customer named ChatGPT or another assistant as a reason for contacting you.

Together they create four ordinary evidence states:

This isn't an attribution model. It is a check between two incomplete records. The first state gives the strongest combined evidence, although it still may not reveal the original question or prove that one answer caused the sale. The second may be an ordinary non-converting visit. The third is the case analytics alone misses: a person may have called from the answer, returned later through another route or never visited the website. The fourth means no evidence was recorded, not that AI had no influence.

The comparison is deliberately modest. It prevents two common errors at once: reporting every AI referral as a customer and reporting no AI effect because the referral report is quiet.

  • An AI-labeled visit also becomes an identified lead
  • An AI-labeled visit appears, but there is no identified lead
  • A customer names an AI assistant, but there is no identifiable website visit
  • Neither record contains AI evidence

Review a short monthly record

A local business does not need a new dashboard for this. Once a month, keep four lines together:

Small counts are still observations. Report them as counts, preserve the pages and customer wording behind them, and avoid percentages that suggest a level of certainty the evidence does not have. If one useful service page repeatedly receives identifiable visits but no enquiries, inspect that page. If callers name ChatGPT while GA4 remains quiet, improve the intake record before changing the website. If neither record shows activity, there may simply be nothing worth acting on yet.

The goal is not to award credit to a channel at any cost. It is to find out whether there is enough evidence to improve a page, repair measurement or leave the work alone. Start by adding the source question to the form or call script your staff already use, then check the AI Assistant row and those answers together at the end of the month. If an SEO company owns the reporting, ask it to preserve the evidence behind the numbers.

  • AI Assistant sessions, separated by session source
  • Landing pages reached from those sessions
  • Forms, bookings or other key events recorded from those sessions
  • Leads who named ChatGPT or another assistant when they contacted the business

Sources

  1. Google Analytics: What's new in Google Analytics
  2. Google Analytics: Default channel group
  3. Google Analytics: Traffic acquisition report
  4. Google Analytics: Mark events as key events