How to evaluate an AI Visibility vendor: 10 questions to ask before you sign

Buying AI Visibility is hard because almost everything promised is invisible to the buyer: nobody archives AI answers, so almost any claim of improvement is impossible to contest afterwards. The ten questions below fix that. They are useful even if you never hire ReBo, and three of them, on their own, eliminate most vendors.

First: why this market is hard to audit

In SEO, the buyer has counter-evidence. Search positions are public, Search Console belongs to the client, and any third party can verify. In AI Visibility, none of that holds: the answer ChatGPT gave your question on Tuesday is not recorded anywhere, it changes with the index and the model, and it cannot be reconstructed later.

That creates a perverse incentive: a vendor who never recorded the starting point can always say things improved. The questions below exist to remove that space, not to corner anyone, but because a serious vendor answers all of them without hesitating.

The ten questions

1. What is my baseline today, and was it frozen before you published anything?

This is the question that eliminates most vendors. Without a record of the prior state (which questions, which engines, which date, who was cited in your place), no number presented later is demonstrable. If the answer is "we measured after the work started", the rest of the conversation is about narrative, not results.

2. Which prompts will be measured, and are they category questions or only branded ones?

A set made only of questions containing the firm’s name produces a high citation rate and measures nothing: the AI cites you because you were named in the question. The set has to be mostly category questions, the ones an allocator asks without knowing your firm exists. Ask for the declared mix.

3. How do you prove citations: a number, a screenshot, a date and an engine, or a traffic report?

Organic traffic is not proof of AI citation. It is the most common substitution in this market, because traffic is easy to show and citations are work to measure. Require a per-answer record: question, engine, date, whether a citation occurred, and with which link.

4. What happens to my historical archive of PDFs?

Years of investor letters and reports are the most underused asset in an investment firm and, in their current format, effectively invisible. A vendor with no plan for the archive will produce new content from scratch, slower, more expensive and less trustworthy than what the firm already wrote.

5. Who signs the content: you or my firm?

Content signed by the agency does not build the institution’s authority. Models use named, verifiable authorship as a tiebreaker, and the byline has to belong to someone at the firm, with a public credential and proper authorship markup.

6. How does compliance approval work, and is it on record?

In financial markets, cadence without compliance is risk, and compliance without cadence is paralysis. Ask for the workflow: who approves, in which window, and where the approval record for each published piece lives.

7. Will the site depend on JavaScript?

AI crawlers do not execute JavaScript. If the page only exists after the browser runs a script (which is what most website builders do), the content does not exist for the model. The test is "view source": if the text is not there, it is wrong.

8. Do you verify the CDN and the WAF, or do you just publish robots.txt?

This is the technical question that most reliably separates people who know the field. Allowing crawlers in robots.txt is useless if the CDN or the web application firewall returns 403 to those user agents. It is the number-one cause of silent invisibility: the file is correct and the bot never read a thing.

9. Who owns the domain, the hosting and the content when the contract ends?

Authority has to accumulate on the institution’s domain, not the agency’s. Content published on a vendor subdomain, or hosting in an account you do not control, turns the result into a hostage of the commercial relationship.

10. Do you promise a guaranteed position or a guaranteed citation?

The right answer is no. Nobody controls what a model answers. A vendor guaranteeing position is guaranteeing something outside its control. And, worse, it will eventually have to choose between breaking the promise and dressing up the measurement.

How to use the list in practice

Send all ten in writing, before the meeting, and compare the answers side by side. The pattern shows up fast: vendors who measure answer with protocol, examples and caveats; vendors who do not answer with case studies, logos and adjectives.

Three answers are disqualifying on their own. A baseline measured after the work began (question 1), proof by organic traffic (question 3), and a promise of guaranteed position (question 10). Any one of them tells you the result will not be verifiable when the contract ends.

And ReBo’s own answers?

They are published, on purpose. The baseline is frozen in the first workstream, before anything goes live. The prompt set is declared, versioned and mostly category-level. The proof is a per-answer record, with date and engine, against the baseline, the full protocol is in how AI Visibility is measured.

The historical archive is released in the data-transformation workstream; content is signed by a partner at the firm; the site is static HTML with no JavaScript dependency; CDN and WAF verification is part of the delivery; domain, hosting and content belong to the institution. And we do not promise position, deadline or guaranteed citation. See the ReBo method and the scope of each workstream.