01 Run the test most buyers skip
Before any sales call, ask the assistants themselves. Put the agency's name into ChatGPT, Perplexity, and Gemini and ask what it does and whether it is credible. Ask each engine which sources it would cite on AI search optimization and see whether the agency's own content appears anywhere. An agency can be young and still legitimate, so a quiet result is not automatically disqualifying, but how they respond when you raise it is revealing. The good ones will show you their own measurement honestly. The bad ones will change the subject.
The same session teaches you something more valuable: what a real AI answer in a niche looks like, which is the thing you are about to pay someone to influence. Note which publications the assistants cite along the way, because coverage on sources of that caliber is what your future agency will need to earn on your behalf.
02 Questions that separate operators from sellers
Ten minutes of specific questions will sort any AI SEO agency shortlist. Ask these and listen for mechanics rather than magic:
- Which of our pages would you expect engines to cite, and why those?
- How exactly do you measure visibility, and can we see a sample report?
- How much of the work is press and mentions, and who does that outreach?
- What happens to our numbers when a model update lands?
- What will you refuse to promise us?
- Which publications ran your clients' stories last quarter, shown as a list, not logos?
Strong answers reference retrieval, crawlers, citation graphs, and above all the earned press coverage that feeds them, since backlinks from real publications are the trust input every engine absorbs (the mechanism is laid out in backlinks for AI trust). Weak answers reference proprietary secrets and relationships with AI companies that do not sell placement.
03 Red flags that should end the conversation
Some claims are not sales enthusiasm, they are physics violations. An AI SEO agency guaranteeing that a named assistant will recommend you cannot be honest, because answers vary by prompt, user, and model version, and no agency controls any of it. Claims of secret prompt tricks or insider access to ranking algorithms describe things that do not exist to be sold. Guaranteed placement in specific named publications, permanent visibility from a one-time fee, and dashboards with no explained methodology round out the list. Any one of these is a walk-away signal, however good the rest of the pitch sounds.
Watch for the softer versions too. A vendor who cannot explain where their links come from is describing a link farm with better fonts. A vendor whose case studies are screenshots with the prompts cropped out is showing you theater. And a vendor who talks only about content volume, with no press capability at all, is selling half the mechanism: coverage is what the engines absorb as trust, and content only gives that coverage somewhere to point.
04 Deliverables to request before you sign
Paper beats promises. Ask for a sample monthly report with real structure, even if anonymized. Ask to see how a prompt panel is designed: what kinds of questions, how many, how changes get logged. Ask for one example of a citation-worthy asset they have built, and for the categories of publication they typically place coverage in, stated as types rather than guaranteed names. Ask for the monthly placement minimum in writing too: an agency selling coverage as the lever should be willing to put a number on it. An agency that produces all of this quickly has an operating system. An agency that stalls has a pitch deck. None of these documents are proprietary, which is the point: a vendor should be comfortable showing you the machine before you pay for its output.
05 Interrogate the measurement method
Measurement is where weak agencies hide, because AI answers are noisy enough to excuse anything. The method you want to hear is boring: a fixed prompt panel held constant month to month, run across ChatGPT, Claude, Perplexity, Gemini, AI Overviews, and Microsoft Copilot, reporting mentions and recommendations separately, with competitor share of voice and an annotated log of model updates. If a vendor cannot explain how they separate real movement from noise, every number they ever report to you is negotiable. Ask one uncomfortable follow-up as well: what would make you tell us this is not working? An operator with a real method has an answer, because the method produces one. Our own approach is documented in measuring AI visibility, and you should compare anyone's method against that standard, including ours.
06 Pricing sanity checks
Real AI SEO work includes earning press coverage, building original assets, and running monthly measurement, and that has a cost floor. A very cheap retainer means one of those is missing, usually the coverage. Opaque pricing that requires three calls to learn is its own signal about how the relationship will run. Compare quotes against published numbers, ours are on the AI SEO pricing page, and ask every vendor the same closing question: what is excluded from this price? Insist on a stated monthly placement count too. A retainer without one leaves the core deliverable undefined, which converts every flat month into an argument about definitions. The scope we consider standard is on the AI SEO services page for reference.