Buyer's guide

Every vendor's number is right.
None of them is comparable.

Nine questions that separate a measurement from a dashboard. Ask them of a tool, an agency, or of us — our own answers are on this page, including where we are the weaker option.

Why this is hard to evaluate

There is no console to check the vendor against.

The problem in one line Search engines publish your real impressions and clicks, so a vendor's claims can be checked against the engine itself. Answer engines publish nothing, so an AI visibility score is a vendor asking its own questions and counting its own mentions — which means whoever wrote the questions decided the result.

That is not an accusation of bad faith. It is the structure of the category. It means the only thing you can usefully compare between two vendors is the method, and the questions below are the ones that expose it.

The nine questions

What are the best AI visibility and brand monitoring tools?

Each one has an answer that should worry you and an answer that should not. Ours is on the third row, marked.

Who wrote the prompts, and can I change them?

01
Why it matters A prompt list is the whole measurement. A vendor that writes its own and will not show you can raise your score by editing the questions.
Worrying answer "Our proprietary prompt set covers your category."
Ours You approve the panel before anything runs, it is versioned, and a changed panel restarts the trend line rather than quietly continuing it.

Is any of it held back as a control group?

02
Why it matters Without a holdout, a rise caused by a competitor's press release looks exactly like a rise caused by work you paid for.
Worrying answer "We benchmark you against competitors instead." That measures the market, not the work.
Ours Twenty percent, drawn at random and stratified by buying stage, excluded from every pipeline. The method, and its limits →

Which engines, and through which interface?

03
Why it matters Driving the consumer apps with a browser script is against most engines' terms, breaks without warning, and makes your measurement a liability rather than an asset.
Worrying answer Vagueness here, or a claim to measure an engine that has no API for what they say they are reading.
Ours Official APIs only, which is a real limitation: it is what the engine will say, not a log of what buyers were shown. No third party has the latter.

Do you store the raw response, or only the score?

04
Why it matters If only the score is kept, nobody can ever re-check a claim, and a scoring change rewrites your history silently.
Worrying answer "We keep the metrics." Metrics without the text underneath are an opinion with a decimal point.
Ours The response text, the engine, the model version and the timestamp are stored per question, so any number in a report can be opened.

What happens to the trend when a model version changes?

05
Why it matters Engines ship new models without notice. A chart that does not mark those dates will show you a cliff and call it performance.
Worrying answer Silence, or a trend line with no version annotations anywhere on it.
Ours The model version is recorded with every run and marked on the chart, and the holdout absorbs the shift because it moves too.

How much does the number move when nobody does anything?

06
Why it matters If the natural week-to-week noise is larger than the lift being claimed, the lift is not a result. Any serious vendor has measured this.
Worrying answer "It is stable." Nobody who has run the same prompt twice in two sessions believes that.
Ours The held-out set is that measurement, every week, for your company specifically rather than as a vendor average.

Who does the work the dashboard implies?

07
Why it matters Measurement is close to a commodity. The expensive half is filing correction requests, chasing registries and writing pages, and a tool does none of it.
Worrying answer "The platform surfaces the opportunities." That is a task list billed as a solution.
Ours We do it, and we sell no seats and no dashboard login — which is the wrong answer for you if a tool your team drives is what you actually want.

What will you refuse to do?

08
Why it matters Hidden text, pages written to instruct a model rather than inform a reader, cloaking, invented reviews and bylines on people who do not exist all work briefly. You own the liability afterwards, not the vendor.
Worrying answer An agent that writes and publishes on its own, with no named human approving what went out.
Ours All five are refused, and nothing is published, edited or sent without an approval record naming a person. It is enforced in the code, not in a policy document.

What does a failed month look like in your report?

09
Why it matters A vendor who cannot describe a bad month in advance will not show you one. Ask what the report says when the work did nothing.
Worrying answer A story about leading indicators and brand lift.
Ours Both sets flat means nothing happened, in those words. Three months of that and you should stop paying us.

Shortcut

If you only ask three.

Ask two, six and nine. A vendor who answers those three cleanly has usually got the rest right, and a vendor who dodges them has nothing underneath.

Can I see the questions

If the panel is proprietary, the number is unfalsifiable by construction.

What is the noise floor

A lift smaller than the week-to-week drift is not a lift.

Describe a failed month

If they cannot, you will never be told you are in one.

Questions

What people ask next.

Why can't I just compare AI visibility scores between vendors?

Because each vendor writes its own prompts, runs them on its own schedule against its own set of engines, and counts a mention by its own rule. Two scores for the same company can differ by a wide margin and both be computed correctly. Compare the method, not the number.

Is a control group realistic for a small company?

It needs a panel of roughly fifty questions or more before the held-out fifth is large enough for the difference to mean anything. Below that, a holdout is theatre and the honest report is a before-and-after with the caveat attached. Say so rather than reporting a lift.

Should the tool and the work come from the same vendor?

There is a real conflict either way. One vendor grading its own homework needs a control group to stay honest; two vendors means the agency disputes the tool's numbers every month. We chose one vendor plus a holdout, and the holdout exists precisely because we grade our own work.

What does a monthly AI visibility tool typically cost?

Measurement alone is close to a commodity and priced like software, often a per-seat monthly fee. Doing the work the dashboard implies is the expensive part, because it is correction requests, pages and outreach done by people. Be clear which one you are buying before comparing prices.

Related reading

The rest of it.