Every agency uses AI now. Asking whether they do is no longer a useful question, and it produces a rehearsed answer.
These nine are the ones we would ask, in the order we would ask them. They are all about process rather than tooling, so you do not need to know anything technical to evaluate the answers. What you are listening for is specificity.
For each, the answer that should reassure you and the answer that should not.
1. Who checks the work, and is it the same person who produced it?
Good answer: a named separate step, with a person or role attached, that happens before anything reaches you.
Bad answer: "we all review each other's work" with no further detail. Self-review is not review, and a process everyone owns is one nobody does.
2. Show me the source for one number in this proposal.
Pick any figure — ideally the most impressive one. You are not testing whether the number is right; you are testing whether they can get to it.
Good answer: they name the account or property, the date window, and can pull it up. Bad answer: a category. "Industry benchmarks", "our client data", "recent studies". Those are not sources.
3. What happens when a claim can't be substantiated?
Good answer: it gets removed. Bad answer: it gets reworded. Rewording is how an unsupported claim survives — it sheds the precision that made it checkable and keeps the authority that made it persuasive.
We removed 22 claims from our own website in August 2026 under this rule, including five certification badges. Some are likely genuine. They come back when someone produces the certificate.
4. Will my results appear on your site, and do you need my permission?
Good answer: yes, we need it in writing, and here is what we would want to publish. Bad answer: any version of "we anonymise it".
Anonymising usually fails for local businesses. Trade plus city plus a specific figure is often one company to anyone in that market. If an agency thinks anonymising solves this, they have not thought about who reads their case studies.
5. What of my data goes into third-party AI services?
Good answer: a clear boundary, stated without defensiveness — what goes, what never goes, and what the retention terms are.
Bad answer: surprise at the question. It is a reasonable thing to know about your own analytics, customer records and ad accounts.
6. What will you do in month one, specifically?
Not a strategy. A list of actions with names and dates.
This question has always been worth asking, and it discriminates better than it used to: a generated plan is fluent about direction and vague about the first Tuesday.
7. Who owns the accounts?
Your Google Ads account, your Analytics property, your Search Console, your domain, your hosting. All of it should be owned by you, with the agency granted access.
Good answer: you own everything, we get invited. Bad answer: anything involving an agency-owned account you would have to leave behind. This is not an AI question at all, but it is the one that costs the most when it goes wrong.
8. What does your reporting say when a month goes badly?
Ask to see a real report from a bad month, with the client's name removed.
Good answer: they have one, and it names the cause. Bad answer: they cannot produce one, or every report is a success. Reporting that only works when results are good is not reporting.
9. Is my site readable to AI assistants, and was that a decision?
Good answer: they know, and they know why. Bad answer: a confident yes without checking, or a blank look.
Most sites are in whatever state their CDN's default put them in. That is forgivable. Not knowing which state, while selling digital marketing, is less so.
None of these ask about tools, models or prompts. That is deliberate. A tool tells you what an agency has. A process tells you what you will get.
What a good meeting sounds like
You are not looking for perfect answers. An agency that says "we do not currently do that, and here is what we do instead" is telling you something useful and is probably being straight with you.
What you are listening for is whether the answers are made of specifics or of categories. Specifics survive follow-up questions. Categories do not, and one more "such as?" is usually enough to find out which you are hearing.
Common questions
Is it rude to ask an agency to show me their sources?
No, and the reaction is informative. You are being asked to spend money on the basis of those numbers. An agency with a verification process finds the question easy and often welcomes it, because it is a question their competitors answer badly.
What if an agency says they don't use AI at all?
Treat it the same way: ask about process. Either it is true, in which case ask how they keep up on the volume work, or it is a positioning claim, in which case the same verification questions apply. The interesting question was never whether, but how it is checked.
Should the agency's own website be good?
It is fair evidence but weak evidence, and worth asking about rather than assuming. An agency's own site is the one project nobody is paying for, so it is often the most neglected thing they own. A poor site plus a candid explanation is a better sign than a polished site with claims nobody can source.
How do I compare two agencies that both answer well?
Ask each the month-one question and compare specificity, then ask who would actually do the work day to day and whether you would be speaking to them. Process quality and staffing continuity separate similar-looking agencies more reliably than proposals do.
What is the single biggest red flag?
A number that cannot be traced to a source, defended rather than removed. Everything else on this list is a question of degree. That one tells you what happens to the truth when it is inconvenient, which is the thing you are really trying to find out.
Want us to check your site?
We will look for the same faults on your website and tell you what we find, whether or not you work with us.
Book a free consultation