How to measure your company's citations in AI answers

"Do the AIs mention us?" isn't a question with a single answer. The same question asked twice to the same assistant can produce two different answers, and two people testing on the same day from two different cities can see different things. Measuring, then, means building a repeatable test and recording its conditions, not running a single check and drawing a conclusion from it.

The second problem is that "being cited" refers to at least three different things, which have different value and need to be counted separately.

What to record along with the answer

A measurement without its conditions can't be compared to anything, not even to itself the following month. Every test needs to be recorded together with what surrounds it.

• The platform and the mode used: different assistants, and different modes within the same assistant, don't behave the same way.
• Whether the answer was built by searching the web at that moment or not. It's the most important distinction of all, because it changes where the information comes from.
• The language of the question and the country you're testing from.
• Whether the conversation was new or continuing an earlier one: what was written before influences what comes after, and a session with a personal history can't be reproduced by anyone else.
• The date, the time, and the exact question, copied word for word.
• The full answer, saved in its entirety. Summaries made by hand lose exactly the details you'll need to reread later.

Three outcomes that get called by the same name

Before counting, you need to decide what you're counting, and the definition should be written down once and used every time.

The last row is the one that matters most and gets looked at the least. Appearing with a wrong description — a service you don't offer, a location that isn't yours, an industry that isn't yours — is a different problem from not appearing at all, and it's addressed in a different way.

A stable set of questions, repeated over time

Comparison only makes sense with constant questions. If you test something different each time, the differences you observe are differences between questions, not between periods.

1. Set a list of questions and don't change it. If you need to add more, add them without removing the old ones, and note when each one was added.
2. Repeat the same question several times within the same measurement session, in new conversations. If your name comes up once out of three tries, the data point is "one out of three," not "we appear."
3. Always run the tests under the same conditions: same mode, same language, a clean conversation.
4. Also record which competitors appear and which sources are cited. In the first rounds of measurement, that's the most useful information you'll come away with.
5. Repeat at a regular interval, one you can actually keep up: a monthly measurement that gets done is worth more than a weekly one that gets abandoned after three rounds.

Reading the sample for what it is

The log describes a set of questions, asked in a certain way, over a certain period, from a certain location. It isn't a market share and it isn't a ranking.

• Proportions calculated on a small number of tests move on their own: from one month to the next, one more or fewer answer changes the number without anything having actually changed.
• It isn't a measure of how many people saw those answers: none of these counts tells you how many times that question was actually asked by someone.
• The useful comparison is with yourself over time, with constant questions and conditions, and secondarily with the competitors that appear in the same answers.

The honest way to report it is a sentence that includes its own limits: on this set of questions, in this language, with this mode, on these dates, the name appears in this proportion of the tests. Put this way, it's less impressive and far more useful, because next time it can be repeated and compared.

From measurement to decisions

A log is useful if it produces two lists. The first: the questions on which you never appear, ranked by how much they matter for your business. The second: the sources cited instead of you — industry directories, publications, comparison sites, competitors' pages — because they show where the information is being drawn from.

From there, the work is no longer about measurement. It's about what's written and where, how clear and verifiable it is, and whether the information about your company is consistent across the places it appears. The measurement tells you where you're starting from and, later on, whether something has changed.

It's worth adding one word of caution. If customers in your industry don't use these tools to choose a vendor, measuring your presence remains an exercise in curiosity. Before the first measurement, it's worth asking which decision the result would actually change: if there isn't one, your time is better spent elsewhere.

What this guide doesn't cover

This guide covers building the measurement: what to record, how to repeat it, how to read it. Which questions to choose — how to identify the ones your customers would actually ask, and how to avoid measuring questions nobody asks — is the earlier step and has its own dedicated guide. Comparing yourself with a competitor who appears in your place, and what to look at to understand why, is also covered separately.