Guide · Measurement
Measure visits from AI sources
Review recorded AI referral visits, channel definitions and customer actions without treating every unexplained visit as AI traffic.
Use the analytics you already have
Use the authorised analytics already available for the website. Establish what it records before proposing another tracking tool or changing a customer form.
A mention in an answer doesn't create a measured visit, and a link in an answer may get no clicks.
Ask who can change the analytics settings and when they last did. A filter added long ago can remove traffic before any report shows it.
Server logs are a second record. They can hold the referring address a browser sent with each request, whether or not the analytics script loaded. They count requests rather than people, and they include crawlers, so filter them before you compare.
Check how sources are grouped
Google Analytics groups traffic by the source and medium values it records. Check the definitions and any custom rules on your account.
Keep a dated list of the source addresses you include. Review unclear rows before you label them AI traffic.
Expect new assistants and new addresses to appear over time. Review the source list on a schedule, not only when something looks odd.
Know where Google Analytics puts AI visits
Google Analytics has a default channel called AI Assistant. Google describes it as visits from sources such as ChatGPT, Gemini, Deepseek, Copilot or Grok.
The same page says this channel excludes Google's AI Overviews. Visits from those count under Organic Search instead.
So an AI Assistant figure leaves out visits from Google's own AI features in Search. A low figure there says nothing about those visits either way.
Analytics sets the medium to ai-assistant when the referrer matches its list of AI assistants. You can't edit that list or the default channel group.
Google says these definitions may change as the market changes. Note the date you read them beside any report that relies on the channel.
Look at the session source as well as the channel. The source shows which address was recorded, while the channel only shows the group it fell into.
Break the channel down by landing page too. It shows which of your pages visitors from assistants arrived on, which may not be the homepage.
Build your own AI group with care
Google Analytics also lets you create custom channel groups, and its help says they can be applied to your reports retroactively.
Traffic goes into the first channel whose rules it matches, in the order you set. Google's help suggests placing an AI channel above Referral for that reason.
Its example uses a broad pattern to match assistant addresses, and it says to update the pattern when those addresses change. Test any pattern against your own list of sources first.
A loose pattern can catch unrelated sources whose names happen to contain the same letters. Check each matched source by hand before you report the group.
Compare the custom group with the default AI Assistant channel for the same period. A large gap means one includes sources the other doesn't, so find out which.
Pick an action worth measuring
Choose a completed enquiry, booking or other business action you can verify. Keep it apart from a visit or page view.
Check the event fires once in a test run. Write down what it means before you report a count.
Pick one main action per report. Several actions mixed into one total make any change hard to explain.
A page view and a completed enquiry answer different questions. Report them separately and state the period each figure covers.
Use the tools and permissions your business already has. This review doesn't need private conversation text or changes to a visitor's consent choices.
Agree the reporting period and time zone with whoever reads the figures. A calendar month in one tool and the last thirty days in another won't match.
Small numbers move a lot. A rise from four visits to eight looks like a doubling but may be ordinary variation. Report the counts, not only the change, when totals are low.
Test the action before you trust the count
In Google Analytics, an action worth reporting is an event marked as a key event. Google says you first create or identify the event, then mark it.
DebugView shows the events Analytics collects from a device in real time, once debug mode is on. Google suggests turning it on for your own device while you troubleshoot.
Submit a test enquiry and watch the event arrive once. Then check it doesn't fire on a page view, a failed form or a second click.
Count the event, not a thank-you page view. A visitor who refreshes the thank-you page can otherwise be counted twice.
Check the event on a phone and a desktop, and in more than one browser. A form that works in one place can fail quietly in another.
Keep a note of each test, with the time and what you did. That lets you take test events out of figures meant to describe customers.
Keep missing data visible
Referrer information can be missing. A report shows the traffic it can identify, not every visit AI influenced.
If analytics isn't connected or an event isn't measured, say so. Don't report the gap as zero.
If analytics stopped collecting for part of a period, mark those days. On a chart, a gap in collection looks just like a quiet week.
Know why some visits arrive without a source
Referring information travels with the visitor's browser, and the page holding the link has a say in what is sent.
The Referrer Policy specification from the World Wide Web Consortium lets a page choose how much to pass on. Its policies range from the full address, to the site's address alone, to nothing at all.
So a visit from a link in an AI answer may arrive with a full address, a bare site name or no source. The linking page's policy and the visitor's browser both play a part.
Treat a missing source as missing. Don't move it into an AI group because the timing looks right, and don't drop it from the total either.
Note how many visits had no source in each period. If that number jumps, look for a change in your own set up before anything else.
Look at the recorded source
Look at source and medium in your analytics, not just the top level channel name. Check how your system groups sources.
Google Analytics channel rules define broad categories. If you want a separate AI referral group, document how you build it.
Keep recognised referring domains apart from unclassified visits. An unknown source isn't evidence that an assistant sent the visit.
Check the referring addresses one level down. One assistant can send visits from more than one address, so a list built on one address can miss the rest.
Read Search Console and Analytics separately
Google says visits from AI Overviews are included in the Performance report in Search Console, within the Web search type. They form part of the overall search traffic there.
Search Console counts clicks from Google to your site. Google Analytics counts sessions and events on your site. They measure different things, so don't add them together.
Use Search Console to see what Google Search as a whole sent you, AI features included. Use the AI Assistant channel for visits from other assistants.
Write both figures with their source and period, side by side. A reader can then see two views of different traffic rather than one blended total.
A simple monthly note helps. Record the AI Assistant count, your custom group, the unknown sources and the enquiries, each with where the figure came from.
Google's newer guide also documents a Generative AI performance report in Search Console. Check which report your property provides and its coverage.
Follow the route from visit to action
Check that a test action shows up as expected in your permitted analytics set up. Keep test events out of customer outcome figures.
When you report enquiries from a recorded source, keep the time window and the attribution definition. They decide which visits are linked to an action.
Don't assume a citation caused a visit. The saved answer and the website event are separate records unless you have evidence linking them.
Attribution settings decide how credit for an action is shared between visits. Write down the model your property uses before you compare one period with another.
If you take enquiries by phone, ask callers how they found you and record the answer as given. A stated source is useful, but it's a customer's memory, not a measurement.
Explain the gaps beside the figures
State the connection, measurement window and known exclusions beside the figures. A missing measurement is not evidence that nobody arrived or enquired.
If analytics is disconnected, show not connected. If a working integration recorded no event, describe the period and the event you checked.
A missing integration isn't a measured zero. Keep the difference visible so a reader knows whether there was anything to count.
Record the date of any change to tags, consent banners or analytics settings. A step change on that date probably reflects the set up, not the traffic.
If your analytics depends on consent, visitors who decline may not be counted. Say so beside the figures, so nobody reads them as every visit.
A worked example
A website has a visit record but no enquiry record.
| Detail | Observation | Review or next step |
|---|---|---|
| Visit | Analytics records a recognised referring source. | Count the visit under your documented grouping. |
| Action | No completed enquiry event is available. | Report the action measurement as unavailable. |
A recorded visit isn't an enquiry, and an unavailable event isn't a measured zero.
Common questions
Is direct traffic evidence of an AI visit?
A visit with no usable source can't tell you where the visitor came from. Leave it unclassified. Don't assign it to an assistant because an answer mentioned you.
Can a citation and an enquiry be connected?
Only if your measurement can establish the link. Otherwise they are separate observations. Report the recorded source and attribution method for the enquiry, and don't join two records because their dates are close.
Is direct traffic evidence of an AI visit?
A visit with no usable source can't tell you where the visitor came from. Leave it unclassified. Don't assign it to an assistant because an answer mentioned you.
Can a citation and an enquiry be connected?
Only if your measurement can establish the link. Otherwise they are separate observations. Report the recorded source and attribution method for the enquiry, and don't join two records because their dates are close.
Check your understanding
What evidence do you need for a recorded referral?
Read the answer and explanation
- A business mention in an answer.
- A website visit recorded under a stated attribution method.
- A crawler access log.
Answer. A website visit recorded under a stated attribution method.
A referral is a visit. A mention or a crawler request can't stand in for the visit record.
Answer and explanation
A website visit recorded under a stated attribution method.
A referral is a visit. A mention or a crawler request can't stand in for the visit record.
Your next steps
Sources and method
A Carqo review method. Provider statements link to the relevant documentation.
- Google Analytics channel definitions Link checked 4 November 2026
- Google Analytics help on custom channel groups Link checked 4 November 2026
- Google Analytics help on key events Link checked 4 November 2026
- Google Analytics help on DebugView Link checked 4 November 2026
- World Wide Web Consortium, Referrer Policy specification Link checked 29 October 2026
- Google guidance on AI features Link checked 4 November 2026
- Google Search Console help on impressions, position and clicks Link checked 4 November 2026
- Google Analytics help on campaign links Link checked 29 October 2026
- Google, optimizing your website for generative AI features on Google Search Link checked 4 November 2026