What should I ask an AI visibility agency before hiring them?

What does this actually cost in 2026?

Between roughly CAD-equivalent $1,500 and $25,000 per month, and the spread is not mostly about quality.

Published 2026 pricing surveys put ongoing AI visibility work for mid-market brands in a band from about $1,500 to $10,000 a month, with most mid-market service businesses landing between $2,000 and $8,000, and broader ranges reaching $3,000 to $25,000 depending on content library size and industry. One-off audits are reported between $1,500 and $7,500, and under about $3,000 an engagement is usually a monitoring dashboard with light reporting attached. Above roughly $20,000 it becomes an enterprise programme with dedicated strategists.

Two things follow from that spread. First, price alone tells you very little: the same number buys a senior strategist at one firm and a junior account manager plus a tool licence at another. Second, the cheap end of the market is frequently not a service at all. A dashboard that tells you what engines currently say about you is a useful instrument, and it is not the same thing as work that changes the answer.

So the questions below are mostly attempts to find out which of those two you are being sold.

What will I have after thirty days?

Ask it in exactly those words, and listen for whether the answer is a noun you can hold.

This is the single most efficient question in the list, because it collapses the entire proposal into one testable statement. "A baseline report and the first set of published corrections" is a noun. "We will still be in discovery" is a schedule. Buyer guidance published in 2026 makes the same point bluntly: if thirty days produces only a research document, the timeline does not match the pace of the thing being bought.

The follow-up matters more than the answer: and what will I have after ninety? A credible engagement can describe a first deliverable, a first measurable change, and a decision point where you can stop. A pilot structured that way, commonly quoted at $6,000 to $9,000 per month for ninety days, exists precisely so you can evaluate before committing to a year.

Compare that with the alternative on offer. A $5,000 monthly retainer on a twelve-month contract is a $60,000 commitment made before a single result is in.

Which promises should end the conversation?

Anything phrased as a guaranteed citation, mention, ranking, or placement inside an AI answer.

Nobody controls what a generative engine retrieves or quotes. Answer sources rotate frequently, engines cite only a fraction of what they retrieve, and different surfaces from the same vendor often share very few sources. A supplier promising a fixed outcome inside that system is either misunderstanding it or counting on you to.

Three more claims deserve the same treatment:

• "We have a relationship with the engines." There is nothing to have a relationship with. There is no submission queue and no inclusion desk.
• "We will get you into AI Overviews with our schema." Structured data is not a requirement for AI Overviews or AI Mode, and there is no markup that unlocks them.
• "Results in two weeks." Anything that moves in two weeks is a dashboard reading, not an established change in what engines say.

What an honest supplier can promise is work: pages written to be quotable, facts corrected at the source, third-party surfaces addressed, and measurement run on a fixed schedule against a real prompt set. That is a description of effort and evidence rather than a promise about someone else's system.

How will we know whether it worked?

By a prompt set agreed in advance, run on a schedule, recorded in writing, and shown to you whether it moved or not.

Ask which questions will be tested. They should be the questions your customers actually ask, not the ones your brand name already wins. Ask how often, ask which engines, and ask to see the raw record rather than a summarised score, because a proprietary "visibility index" that only the supplier can compute is a number you cannot audit.

Ask what they track beyond position. Buyer guidance in 2026 suggests asking whether the strategist tracks citation share rather than rankings, and whether they work on third-party surfaces at all, since a large amount of what engines repeat about a company comes from places the company does not own.

Then ask the question most proposals do not survive: what would count as this not working, and what happens then? A supplier with a real answer has thought about failure. A supplier without one has only planned for the invoice.

What happens when an engine states something false about me?

This is the question that separates marketing work from the thing most buyers actually need.

Wrong prices, discontinued products, a service you stopped offering, a location you left, an ownership claim that was never true. These are not visibility problems; they are correction problems, and they persist because the wrong fact is sitting in a source the engine trusts.

A credible answer describes a route: find the source that carries the error, correct it where it lives, publish an authoritative statement of the correct fact on a surface you own, and re-test on a schedule until the answer changes. An evasive answer talks about "reputation management" without naming a mechanism.

Ask specifically who does the correcting, whether they will contact third-party sites, and what they do when the source is one they cannot edit. The honest answer to that last one is that you publish a clear, well-structured, checkable version on your own property and keep it current, which is a slower remedy than anyone wants and is frequently the only one available.

What do I still own if we part company?

Ask before signing, not at the exit interview.

Three specifics are worth naming in the contract: the content produced during the engagement, the prompt set and measurement history, and any persona, assistant, or answering surface built on your behalf. Each of them can quietly become the supplier's asset rather than yours.

The third is the one that catches people. If a supplier builds you an answering presence that lives on their platform, trained on your material, and the relationship ends, the question of what happens to it should already have a written answer. Ask whether you can export your sources. Ask what happens to the thing itself. Ask what a wind-down looks like and how long you have.

The same applies to measurement. A year of recorded answers to your prompt set is genuinely valuable, and it is worthless if it lives inside a tool you lose access to on the day you stop paying.

What are our answers to these questions?

The list is only worth publishing if it can be applied to us, so here it is applied to us, at our published prices.

• What it costs. Paid work starts with a CAD $2,000 Persona Blueprint: audience and use-case definition, a source-readiness assessment, a rights and risk scan, a working prototype, and an evaluation and implementation plan. Larger engagements start from CAD $7,500 and are quoted against source volume, rights, review complexity and deployment. Payment is normally half to begin and half at approved launch.
• What you have after thirty days. The Blueprint itself, which is a deliverable in its own right and remains useful if the full project never proceeds. Blueprint fees may be credited toward a full project that begins within thirty days.
• What we promise. Not a citation, a mention, or a position, because nobody owns those. What we commit to is the work and the evidence for it.
• How it is measured. Against your prompt set, on a schedule, with the record shown to you.
• What you keep. Your sources are yours and exportable. The persona is yours.

The deliverable at the end of that is a public, source-grounded presence that answers questions about you from material you approved, and declines when it has nothing to answer from.