How to measure AI search when GA4 shows you almost nothing

Every AEO conversation eventually arrives at the same fair question: how would I know if this is working? With Google, the answer is comfortable — rankings, impressions, clicks, all sitting in a dashboard. With ChatGPT, Perplexity, Gemini and Google's AI overviews, the honest answer is that your analytics will show you a fraction of the story, and the fraction it shows will be the least interesting part.

That is not a reason to stop measuring. It is a reason to measure something else. Here is the reporting we set up for Dubai clients, and how you can build most of it yourself.

Why AI search barely registers in your analytics

An assistant answers the question inside its own interface. If the answer is good enough, nobody clicks anything — and a visit that never happens cannot appear in GA4. When someone does click through, the referrer sometimes survives and sometimes doesn't: links get pasted into a message, opened on a different device, or forwarded to whoever actually signs off on the decision.

Then there is the most common path of all. A parent asks an assistant which nurseries in her area offer an early-years programme, sees three names, and the next morning searches one of those names on Google. In your analytics that is branded organic search. It was AI search that put you on the shortlist.

So the first rule of AI-search reporting: your referral number is a floor, never a total. Judge the channel on whether you are being named, not on how many sessions it can be credited with.

Start with a citation baseline, not a traffic number

A baseline is a fixed list of questions your buyers genuinely ask, run against the assistants that matter, with the results written down. Ten to thirty questions is plenty for an SMB. The discipline is in choosing them properly:

• Category questions — "best British curriculum nursery in Al Barsha", "top HR training providers in Dubai". Hardest to win, most valuable.
• Comparison questions — "X vs Y for a mid-year start". These are where comparison pages earn their keep.
• Problem questions — "what should I ask on a nursery tour?", "how long does a PMP course take?" Easier to be cited on, and they build the association between your name and the topic.
• Brand questions — "what does [your company] charge?", "is [your school] any good?" You should always want to know the answer to these.

Record three things per question: were you named, what exactly was said about you, and which sources the answer cited. That third column is the one people skip and later wish they hadn't — it tells you which pages, directories and third-party listings the model actually trusts on your topic.

If you want a starting point without building the spreadsheet, our free AI visibility check runs the first pass for your own business. Our plans carry the baseline forward as monthly tracking — three queries on Starter, five on Growth, ten on Premium, with the full breakdown on the pricing page.

The five numbers worth reporting every month

• Citation rate. The share of your tracked questions where you are named at all. One number, tracked over months, on a query set that does not change.
• Answer accuracy. Of the answers that do mention you, how many get the facts right — services, locations, age groups, opening hours, fee ranges. A confident wrong answer costs you more than silence; we covered the repair process in when AI gets your business wrong.
• Source mix. Which URLs the assistants lean on. If your own site never appears and a directory always does, you have a content gap, not a visibility problem.
• AI referral sessions. The floor number from analytics. Useful for direction of travel, not for board-level claims.
• Self-reported enquiries. What people tell you when you ask how they found you. Imperfect, unbeatable.

Setting up the analytics side (about thirty minutes)

In GA4, create a custom channel group that catches the assistant hosts — chatgpt.com, perplexity.ai, copilot.microsoft.com, gemini.google.com and their variants — so those sessions stop being scattered through "referral" and "direct". Add new hosts as they appear; the list keeps growing. Then put those sessions next to your conversion events, because a small number of AI referrals that convert well is a much better story than a large number of visitors who bounce.

UTM tags will not save you here. You do not control the link an assistant produces, so there is nothing to tag. What you do control is your enquiry form. Add one optional free-text field — "How did you hear about us?" — and read the answers monthly. When people start typing "ChatGPT recommended you", that is the channel reporting itself, and it is worth more than any dashboard.

Also watch branded search in Google Search Console. A steady rise in people searching your name, without a campaign to explain it, usually means you are being named somewhere you cannot see.

What progress actually looks like

Nobody can promise a citation, and any agency that guarantees one is guessing on your budget. What we can describe is the order things tend to move in. Accuracy improves first, because correcting the facts about your business is largely within your control. Citations on problem and brand questions follow, since those have less competition. Category questions — the commercially valuable ones — move last and least predictably.

That ordering matters for expectations. Two months in, the right question is not "are we number one for the big query" but "is what the assistants say about us now correct, and are we appearing on the easier questions". If the answer to both is yes, the foundations are working. For why the models pick who they pick, see how AI search engines choose which websites to cite, and for the broader timeline, how long SEO takes in Dubai.

Four reporting mistakes we see constantly

• The lucky screenshot. One flattering answer, framed as proof. Assistant responses vary between sessions, accounts and locations — a single result is an anecdote.
• Tracking questions nobody asks. Query sets built from keyword tools rather than from what your sales team actually hears on the phone.
• Changing the query set every month. It feels thorough and it destroys the only thing a baseline is for. Keep the core fixed; add to it, rarely.
• Reporting citations with no commercial line. Citations are the leading indicator. Enquiries are the point. Put them on the same page.

Common questions

If you already publish good content and still cannot tell whether AI search is doing anything for you, the gap is usually measurement rather than marketing. Our AEO service starts with the baseline for exactly that reason — you cannot improve a position you have never recorded.