How to Measure a Channel That Produces No Analytics

Every other channel you run reports on itself. Search Console shows impressions and position. Ads report spend and conversions. Social reports reach. You can be wrong about what the numbers mean, but the numbers arrive.

AI visibility reports nothing. When an assistant recommends your business, the visitor frequently arrives by typing your name directly, and your analytics records it as direct traffic with no explanation. When an assistant recommends a competitor instead, absolutely nothing happens on your side.

So the measurement has to be built rather than collected. Here is what that actually looks like.

**The core method: a standing prompt set**

Write down the questions your customers actually ask, in their words rather than your marketing vocabulary. Fifteen to thirty is usually enough to be representative without becoming a chore.

Mix the types. Category questions — "best X in Dubai". Comparison questions — "should I use X or Y". Qualification questions — "is X safe / worth it / regulated". Direct questions — "tell me about [your company]". Each type exercises a different part of how the model assembles an answer, and businesses often perform very differently across them.

Then run that exact set, unchanged, on a fixed schedule. Monthly is enough for most businesses. The value comes entirely from the set staying constant — a prompt you reword is a prompt you cannot compare over time.

**What to record**

For each prompt, on each assistant, capture:

Whether your business was named at all. This is the primary metric and it is binary.

Which competitors were named, and in what order. Order is not a ranking in any formal sense, but it is stable enough to be informative.

What was said about you, verbatim. Not a summary — the actual text. You are looking for two things: whether the description is accurate, and whether it reflects how you position yourself or how someone else describes your category.

Whether any sources were cited, and which. Perplexity shows these directly. Others sometimes do. Where citations are visible, they tell you which sources your category actually draws on, which is more actionable than the answer itself.

Anything factually wrong. Prices, services, credentials, locations, regulatory status. This deserves its own column because it needs fixing regardless of visibility.

**Run it per assistant, and in each language**

Do not blend the results into a single score. ChatGPT, Perplexity, Gemini, Claude and Copilot draw on different sources and weight them differently, and a blended number hides exactly the platform where you are weakest.

The same applies to languages. In the UAE specifically, running the set in Arabic as well as English usually produces a noticeably different picture, and in most categories a much thinner one.

**What "share of voice" means here, and what it does not**

You can compute a share-of-voice figure — the proportion of prompts where you were named, or your share of total mentions across the set. It is a useful trend line and a legible number for a board deck.

Be honest about its limits. It is a sample, not a census. It reflects your prompt set, so it can be gamed by choosing flattering prompts. It has no denominator in real query volume, so it does not tell you how many people actually asked. And answers vary between runs even with identical prompts, so small movements are noise.

Treat it as a directional indicator over quarters, not a KPI you manage weekly.

**The leading indicators that move first**

Because visibility itself is slow and lumpy, track the things that change before it does:

Structured data coverage — how much of your site a model can parse without interpretation. Citation source presence — whether you appear in the sources your category's answers draw on. Description accuracy — how closely what the model says matches reality. Crawler access — whether the relevant bots can reach you at all, which is binary and worth re-checking after any infrastructure change.

These move within weeks and they are causally upstream of the thing you actually want.

**Tooling, honestly**

There are commercial tools that automate this, and they save real time on the mechanical part — running prompts across platforms on a schedule and storing results. What none of them can do is choose your prompt set or interpret what the answers mean for your business, which is most of the value.

A spreadsheet and a disciplined monthly hour gets you most of the way. Buy a tool when the manual version has proven the channel matters to you, not before.

**The uncomfortable part**

Your first measurement will probably be worse than you expect, and you will not be able to tell how long it has been that way. There is no historical data to look back on, because nothing was recording.

That is the argument for starting the record now rather than after you have done the optimisation work. A baseline you took before you changed anything is the only way you will ever know whether the changes did anything.

Prêt à développer votre entreprise à Dubai ?

Obtenez un audit gratuit et découvrez comment nous pouvons vous aider à dominer votre marché local.