How do I choose an AI visibility agency?

What evidence should I demand?

Three artefacts, and all three should already exist before you are a client.

• A published method you can recompute. Not a description of a method. The formula, the weights, the question set and the rule that decides whether a mention counted. If you cannot redo their arithmetic, you cannot audit whether a number moved for a real reason or because the formula quietly changed.
• The raw data behind every statistic they quote at you. A percentage with no denominator and no date is not evidence. If a supplier tells you some share of businesses are invisible to AI, ask how many businesses, which industries, scanned on what date, and whether you can have the file.
• The answers themselves, as text. A score is a summary of something. Ask to see the something. A supplier who can show you what an engine actually said, question by question, with the competitor names in it, is working from observation. One who can only show you a dial is working from a dial.

This is why our own study ships its CSV and JSON under CC BY 4.0, and why the scoring formula sits on a public page rather than inside a sales deck. That is not generosity. It is the minimum condition for any of the numbers meaning anything, and you should hold every supplier in this category to it, us included.

Which questions actually separate one supplier from another?

The ones about their own numbers, not yours.

Anybody can describe a process. Far fewer can answer these on the spot.

1. What is your own AI visibility score, and what did the engines say about you? A firm selling this that has never measured itself is selling something it does not use. If the score is mediocre, a straight answer about why is more reassuring than a perfect one.
2. Show me the exact questions you will ask on my behalf. Then check that none of them contain your business name. A question that names you cannot test whether an engine names you unprompted, and quietly including the brand is the most common way a result gets flattered.
3. Who decides that a mention counted, a rule or a model? If a language model is asked to grade the output of a language model, there is a second layer of variation in the number that nobody can see.
4. What is my number today, before any work starts? With no baseline written down before the invoice, no improvement can be claimed later.
5. Which of my gaps do you think is costing me the mention, and why that one? This is the judgement question. Listen for whether the reasoning refers to what the engines actually said about you, or to a generic checklist.
6. Will the same questions be re-asked after the work, and when? Different questions after the work is not a re-measurement.
7. What do I own at the end, and what happens if the number does not move? Both answers belong in writing.

Commercial terms are a separate checklist and it is already published on the cost page. Take that one into the quote conversation and this one into the capability conversation.

What are the red flags?

Six, and the first is disqualifying on its own.

• A guarantee of a ranking, a placement or a position. No supplier controls what a model generates. Anyone promising a specific outcome inside somebody else's model is either misunderstanding the product or misrepresenting it, and neither is a good start.
• A score with no published formula. An unauditable number can be adjusted quietly. If they will not show you how it is calculated, treat every future report as marketing.
• Statistics with no denominator or date. This category is full of borrowed percentages with no population attached. Ask for the source and the sample size once, politely. The reaction tells you a great deal.
• Ranking language. Covered on its own below, because it is the most revealing tell of the six.
• A retainer with no re-measurement. If nobody re-asks the questions, nobody can honestly tell you it worked.
• Monitoring sold before a diagnosis. Watching a number nobody is trying to change is a subscription, not a strategy.

One more that is easy to miss: a case study with no dates. The engines change underneath everybody. A result from an unnamed month is not a result you can weigh.

And one that sounds like rigour: a dollar value attached to a single AI mention. Nobody has credibly measured that, us included, so a figure quoted at you has been invented rather than sourced. How much a ChatGPT mention is worth sets out the method for building your own number instead.

Why is ranking language such a tell?

Because there is no ranking to be in.

Generative engines do not maintain a stable ordered list that a business occupies a slot in. They compose an answer, and the composition varies. In our scan of 300 Australian businesses, across the 137 with complete records from both engines, Gemini scored the same business higher than ChatGPT 77% of the time, ChatGPT was higher 5% of the time, and the average gap between the two engines for the same business on the same day was 25.3 points out of 100.

Two systems, one business, one day, a quarter of the scale apart. A supplier promising you position three has imported a mental model from classic search that does not apply here. That does not make them dishonest, but it does mean their reporting will describe something that is not happening. The background is why AI answers change every time you ask.

What should the first month look like?

A baseline, a written question set, a ranked gap list, then work, then a re-scan on the identical questions.

If the first month is all strategy documents and no measurement, you have bought a report. If it is all work and no measurement, you have bought activity. The sequence matters more than the volume, because without the first and last steps nothing in the middle can be evaluated. Knowing whether the work is working covers what a genuine improvement signal looks like against ordinary variation.

Be realistic about timing too. Nothing here moves in a week, and a supplier who implies otherwise is setting your expectations badly on purpose.

A baseline also reads better when you know what normal looks like in your trade. The per-industry never-named rates from the same scan, with each denominator attached, are on which Australian industry has the worst AI visibility.

What if I end up hiring someone else?

Then take this page with you and hold them to it.

We would rather be one of several suppliers that can survive these questions than the only firm that gets asked them. The checklist works on any agency in the category, and it works on us: the formula is public, the study data is downloadable, and our audit hands over all twelve AI answers in full rather than a summary of them.

Two things should be said plainly. Answerable sells exactly the service this page is about, so read it as an interested party's version of a fair test rather than a neutral one. And the strongest number in our own dataset argues against overbuying from anybody: across the 300 businesses in that scan, site readiness averaged 74.9 out of 100, yet across the 137 with complete records from both engines it correlated with per-engine visibility at r = 0.116. If a supplier's entire pitch is a package of on-page fixes, that figure is the thing to put in front of them.

Get a baseline before you brief anyone.

Walking into a supplier conversation with your own number and your own answer text changes it completely. The free scan gives you both, and it commits you to nothing.