Most AI-assisted enquiries will never show up as AI traffic. They arrive as direct visits, branded search, or no visit at all. To measure them you need three things: referrer capture where it exists, a “how did you hear about us” field on the form, and a prompt log that records what the engines say about you. Reported and self-reported sources stay in separate columns.
The short version
- A citation does not have to create a visit. Judging AI visibility by sessions will tell you it is failing while it works.
- Some engines pass a referrer and some do not. Capture it where it exists and accept the gap.
- One optional form question recovers more attribution than any analytics configuration.
- Keep tracked sources and self-reported sources apart. Adding them together double counts.
Why does AI traffic barely appear in analytics?
Three reasons, and only one is fixable.
First, the answer often satisfies the question, so nobody clicks. Second, when someone does follow up, they frequently open a new tab and search your brand name, which lands in analytics as organic branded search or direct. Third, referrer passing is inconsistent across the assistants, and some strip it entirely.
The result is that the channel reports as smaller than it is. We have watched this on our own site: a page holding position one in Google produced roughly 330 impressions and a single click over 28 days, while enquiries arrived saying they had found us through an AI answer. On a clicks report that page looks broken. It was working.
What can you actually capture?
| Signal | What it tells you | Its limit |
|---|---|---|
| Referrer from an assistant | A visit came from a named AI surface | Only when that surface passes a referrer, which not all do |
| Landing page plus no referrer | A deep internal page entered directly, which rarely happens naturally | Suggestive, never proof. Bookmarks look the same |
| Branded search volume | People are looking you up by name | Cannot tell you what prompted it |
| Form question | What the person believes led them to you | Self-reported, and some will skip it |
| Prompt and response log | Whether the engines name and cite you at all | Measures visibility, not revenue |
| AI crawler hits in server logs | Which engines fetch your pages, and which pages | A fetch is not a citation and not a visit |
How do I set this up in GA4?
Start with the conversion itself. Track confirmed form submissions, not clicks on the submit button, or your whole model is built on a number that includes failures. Treat call-link and WhatsApp-link clicks as intent signals and label them that way, because a tap is not a conversation.
Then preserve the context. Capture the landing page, the referrer and any campaign parameters into hidden fields on the form, so they arrive in your CRM attached to the enquiry rather than sitting in an analytics report you cannot join to it. The enquiry record is where attribution has to live if you want to know which sources produce clients rather than leads.
Finally, add the question: “How did you hear about us?” as an optional free-text or short-list field. It is unfashionable and it recovers more of this channel than any amount of tagging.
How do you connect visibility to enquiries?
Keep two records and compare their shapes over time rather than trying to join them row by row.
The first is the prompt log: a fixed set of buyer questions, run on a schedule across the engines, recording the exact prompt, the platform, the date, the full response, which brands were named and which URLs were cited. Keep the runs where you were not mentioned, because the rate is the measure.
The second is the enquiry log with its self-reported source field. If the share of responses naming you rises over a quarter, and the share of enquiries mentioning an AI assistant rises in the same quarter, that is a reasonable inference. It is not proof of causation, and saying so plainly is what separates a real report from a dashboard.
What should the monthly report contain?
Three visibility measures kept separate: the share of responses that name you, the share that cite a URL of yours, and the share that place you in the first three recommended providers. Then the enquiry side: total enquiries, how many self-reported an AI source, and how many of those became qualified conversations.
Keep the prompt set unchanged between months. The most common way these reports mislead is that the questions quietly improve in your favour, and then the trend is measuring the question rather than the business.
A mistake worth avoiding: creating a channel in analytics called “AI” and treating its number as the size of the channel. It will capture only the assistants that pass a referrer, which makes it an arbitrary fraction, and once it exists on a dashboard people start optimising against it. If you build the segment, label it “AI referrals we can see” so nobody mistakes it for the total.
If you want this measured properly rather than estimated, our free AI visibility audit gives you a baseline wave with the raw responses, including the runs where you were not mentioned. Request the free audit, or read how we set up conversion tracking.
Frequently asked questions
Can I see ChatGPT traffic in Google Analytics?
Partly. Some AI surfaces pass a referrer and will appear as a referral source, and others pass nothing, so those visits land in direct. Treat whatever you can see as a floor rather than the total, and use a form question to recover the rest.
Does a citation in an AI answer bring traffic?
Not reliably. Being used as a source means the engine read your page, not that the reader visited it. Measure citations as a visibility outcome in their own right, and measure enquiries separately, or you will keep concluding that one proves the other.
How do I know whether AI visibility work is paying off?
Compare two waves of an unchanged prompt set, then look at whether qualified enquiries moved in the same period. Keep the two measures in separate columns and state that the link is an inference. Anyone offering you a single attributed revenue figure for this channel is guessing.
Should I add a “how did you hear about us” field?
Yes, optional and short. It is the single most effective thing you can do for attribution on this channel, because it captures the journeys that leave no technical trace at all. Report it separately from tracked sources so the two never get added together.