How to measure AI visibility: a working method for mid-market teams

You cannot query an AI assistant for your position the way you query a rank tracker for a keyword. An assistant returns prose. Sometimes it names four vendors, sometimes one, sometimes none, and the wording differs between two identical asks. So measurement here is not retrieval, it is sampling: you fix a set of questions, you ask them repeatedly on a schedule, you read the answers with a parser instead of your eyes, and you turn thousands of sentences into a small number of figures you can put in a report.

Four decisions decide whether that figure is worth anything: which prompts you ask, which engines you ask them of, how often you ask, and whom you compare yourself to. Get any of the four wrong and you will produce a number that moves without an explanation, which is worse than no number, because someone will act on it. The rest of this page is those four decisions in order, followed by the test that matters most once the data starts arriving, which is whether a move came from the market or from a model.

One thing belongs at the top rather than in a footnote. LLM answers are non-deterministic and they change when a provider ships a model update. A week over week move in any AI visibility metric may reflect a model update rather than a real change in how the market sees you. Every credible version of this measurement carries that caveat next to the number, not three pages away from it.

Why your existing numbers do not transfer

Analytics can tell you a session arrived. It cannot tell you the answer that preceded it, whether you were named, whether you were named first, or whether the assistant described you as the enterprise option when you sell to mid-market. Impressions and clicks describe a results page. AI visibility describes a paragraph. They are different objects and they need different instruments, which is why the work below starts with writing questions rather than with pulling a report.

Most disappointing AI visibility programmes fail here, and they fail in the same two ways.

The first is vanity prompts. If your set is dominated by questions that contain your brand name, you will measure a high presence rate and learn nothing, because an assistant asked about you will talk about you. Branded prompts are worth a handful of slots to check the accuracy of your description. They should not drive the headline.

The second is prompts nobody asks. Generated prompt sets drift toward tidy category phrasing that no buyer types. Read the set before you freeze it. If a line does not sound like something a person in a deal would say out loud, cut it or rewrite it. The test is simple: could you imagine this sentence arriving in your inbox from a prospect who had not spoken to you yet.

A useful shape for a set of fifty to a hundred prompts:

• Category discovery, the plain "best tool for X" and "how do I solve Y" questions, the largest group.
• Comparison and alternatives, which is where competitor sets surface most reliably.
• Constrained buying questions that carry a segment, a size, a budget or an industry, since these are where mid-market brands get sorted into or out of the shortlist.
• Objection questions, the risk, price, migration and integration asks.
• A small branded block to check how you are characterised rather than whether you appear.

Two disciplines apply throughout. Write the prompt set once, then leave it alone for at least a quarter, because every edit resets the comparability of the trend. And keep a note of the date you froze it and the date of any change, because the first question an analyst will ask about a step change in your chart is whether you changed the questions. If you cannot answer that from a record, you cannot defend the chart.

It is also worth deciding, at this stage, which prompts you intend to win. A frozen set is not just an instrument, it is a shortlist for the content team. Marking five or six prompts as commercially decisive turns a monitoring habit into a brief, and it gives you something to review against in a month.

Coverage is a trade-off, not a checklist. Each additional engine multiplies query volume, cost and the number of different failure modes you have to handle. Each engine you leave out narrows what your number describes.

Sample where your buyers actually ask, and be honest in your reporting about the boundary. A number computed from two engines is a number about two engines. That is a legitimate measurement as long as the label says so. A number that claims to cover every assistant but quietly samples one is not.

Two practical constraints shape this decision more than most teams expect. First, access is granted, not owned: providers set usage policies and rate limits for automated querying, and continuous commercial monitoring sits squarely in the territory those policies govern. Build the system behind an adapter layer so a single engine can be switched off without taking the measurement with it. Second, answers with citations and answers without them are different evidence. Some assistants show the sources they drew on, which tells you which pages earned the mention. Others give you prose only, which tells you the outcome but not the route. Do not average those two things together and call the result one insight.

Non-determinism is handled with volume and repetition, not with cleverness. A single ask of a single prompt is noise. The same prompt asked every week for eight weeks, alongside fifty-nine others, produces a distribution you can reason about.

Daily sampling multiplies provider cost without changing what you can act on, because the work that moves AI visibility, publishing, positioning, earning citations and third-party coverage, does not land daily. Monthly sampling is too slow to connect a move to a cause, and it also means a single bad sampling day can define the month. Weekly is where most mid-market teams land, and it maps onto the reporting rhythm they already have.

Whatever cadence you choose, hold three things constant: the prompt set, the competitor set, and the time of week you run. Then record what you could not hold constant. Model versions change underneath you. Providers go down. A batch fails halfway. If you do not log these, you will eventually explain a model update as a competitor campaign in front of your leadership team, and nobody in the room will be able to tell you that you are wrong.

One more rule that sounds pedantic and saves you twice a year: an unsampled week is not a zero. If a run did not complete, report no score, not zero percent. A zero in a trend line is a claim that nobody named you, which is a very different statement from "we did not look".

This is the strongest argument against buying or building anything in this category, and it deserves a straight answer rather than a rebuttal.

It is true. Ask the same question twice and the wording, the order and sometimes the list of brands will differ. Ask it after a provider updates a model and the difference can be larger still. No tool, including ours, can make an LLM deterministic or promise that a given answer will repeat.

What a measurement can do is stop pretending each answer is a fact and start treating it as an observation. Three things make the aggregate usable:

• Volume. Fifty to a hundred prompts per run means no single answer sets the number. Individual variance averages out; a real shift in how you are represented does not.
• A constant instrument. The same prompts, the same competitor set, the same cadence. When only the answers change, the change is about the answers.
• Recorded conditions. Model version stored on every answer, so a discontinuity that coincides with a provider update can be identified as exactly that.

What you get from this is a directional, caveated market signal, not a precise instrument. Report it that way. A team that says "our share of AI voice moved from the high twenties to the mid thirties across sixty prompts on two engines, and no model version changed in that window" is making a defensible claim. A team that reports 34.2 percent with a decimal point and no methodology is making an indefensible one.

Once you have four or five weeks of data, most of your analysis time goes on one question: did the market change, or did the model? This is the part of the method that no dashboard can do for you, and it is where the difference between a habit and a reporting line gets decided.

Work through it in order.

First, check the model version recorded against the answers in the current run versus the previous one. If the version changed, treat the step as suspect until proven otherwise, and expect the level to have reset rather than trended. A reset level is not a failure of the measurement, it is a new baseline, and it should be annotated as one.

Second, check whether the move is broad or narrow. A real competitive shift usually concentrates: a competitor starts appearing in the comparison prompts, or you drop out of one use case cluster. A model-driven shift tends to move everything at once, including prompts where nothing about your market has changed. Sort your prompts by how much each one moved before you look at the total.

Third, read the answers. This is the step teams skip and the only one that produces a cause. Pull the prompts where you were named last week and not this week, and read what the assistant said instead. You are looking for the new source it leaned on, the framing it adopted, the competitor it substituted. Two dozen answers read properly will beat another metric every time.

Fourth, check your own denominator. If a competitor rebranded, merged or is now referred to by a different name, your mention extraction may be missing them, which inflates your share without anything real having happened. The same goes for your own brand if it has a common word in it.

Only after those four checks is it reasonable to attribute the move to your own work, or to someone else's. And when you do attribute it, write down the attribution and the date, because next quarter you will want to know which of your explanations held up. That record, more than any single figure, is what turns AI visibility measurement into something a leadership team can rely on.

The measurement is half the job and the reporting is the other half. The sibling pages in this pillar carry the parts this overview deliberately leaves alone: the tracker page covers the single weekly score and the caveat that travels with it, the benchmarking page covers the named competitor view answer by answer, the sentiment page covers classification and rank inside each answer, the weekly report page covers the digest and its drift alerts, the SEO reporting page covers fitting all of this into a monthly report an analyst cannot dismantle, and the brand manager page answers the question that usually starts this search, which is how ChatGPT describes you when nobody from your company is in the room.