How to measure conversions from AI traffic in GA4

Get the source dimension right first

GA4 offers several source dimensions and they answer different questions. Session source is the source of the session in which the event happened. First user source is where that person originally came from, days or weeks earlier. Session default channel group is the grouped version of the first.

For the question most people are actually asking — did this AI visit lead to something — Session source is the right dimension. First user source answers a different and often more flattering question, and quietly using it is one of the easier ways to overstate a channel.

Pick one, write down which, and use it consistently. A figure computed on Session source in one month and First user source in the next is not a trend.

Build the report

The fastest route is an exploration. Reports, then Explore, then a free-form report. Set the row dimension to Session source and add Landing page plus query string as a second dimension. Add Sessions, Key events and Total revenue as values.

Filter to sources matching your AI expression — chatgpt\.com, perplexity\.ai and the rest. What you get is a table of assistant by landing page with conversions and revenue attached, which is the artefact worth having.

Note that GA4 renamed conversions to key events. If you are following an older guide that references a conversions metric, that is why the field is not where the guide says it is.

Why the attribution model changes your answer

GA4 applies an attribution model to credit conversions across touchpoints. Under a data-driven model, a conversion that involved an AI visit and two later visits from other sources is split between them, and the AI channel receives a fraction.

That is defensible, and it means the number you see is not the count of conversions whose session began at an AI assistant. If you want that simpler figure, an exploration filtered on Session source gives it to you, because it counts events within qualifying sessions rather than distributing credit.

Neither figure is wrong. They answer different questions and they will differ, sometimes by a great deal. Reporting one while describing the other is how a number stops being trustworthy.

The gap you must state

Every conversion figure here covers only sessions that arrived carrying a referrer or a campaign tag. Assistants that strip the referring page send visitors who convert exactly as often as anyone else, and their conversions are attributed to Direct.

So the honest form of the sentence is: this is the revenue attributable to AI-referred sessions we could identify, and an unknown additional amount sits in Direct. That sentence is longer and less satisfying than a single number, and it is the version that survives the follow-up question.

If you want to bound the unknown rather than ignore it, compare the conversion rate of your identified AI traffic with the conversion rate of Direct traffic over the same period, and watch whether Direct's landing page mix shifts towards the deep pages assistants tend to cite. Neither is proof. Both beat silence.

What to report and how

Report three things together: identified AI sessions, key events from those sessions, and revenue. One without the others invites the wrong conclusion, because a channel with few sessions and excellent conversion deserves a completely different response from a channel with many sessions and none.

Break it down by landing page rather than by assistant alone. Which pages AI traffic lands on is more actionable than which assistant sent it, because the page is the thing you can change.

And carry the caveat into the report itself rather than into a footnote nobody reads. A figure whose limits are stated beside it is one people trust; a figure whose limits emerge later under questioning is one they stop trusting entirely.

Setting up so the question is answerable at all

None of this works without key events configured. A property tracking pageviews and nothing else has no conversions to attribute, and the exercise ends before it starts.

Configure the events that correspond to the outcomes you care about — a form submission, a signup, a purchase — and mark them as key events in Admin. Give them at least a few weeks before drawing conclusions, because a key event created today has no history behind it.

If you sell online, make sure revenue is being passed with the purchase event. Total revenue is empty without it, and an empty revenue column is usually a tagging gap rather than a channel that earns nothing.

Finally, decide once whether your reporting uses attribution-model credit or session-filtered counts, and hold to it. The two produce different numbers for the same month, and mixing them across periods produces a trend that reflects your method rather than your business.

Reporting it so the figure survives

State the source dimension, the period and the attribution basis beside the number. Three short clauses turn a contestable figure into one that answers its own follow-up questions.

Show sessions, key events and revenue together. Any one of them alone invites a conclusion the other two would have prevented.

And name the gap in the same breath. Attributable AI revenue with an unknown additional amount in Direct is a complete and defensible sentence, and it is the version that holds up when somebody senior starts asking how you know.

And keep the exploration itself rather than rebuilding it monthly. A saved report that anyone can open is what turns a one-off analysis into something the organisation actually uses.

The saved report also settles a question that otherwise recurs: whether two people quoting different numbers were measuring different things. If both opened the same report, they were not, and the conversation moves on quickly.