The five questions
Ask these in the first call. Any real practitioner answers all five without stalling.
One: what prompt set do you audit, and can I see it. A fixed question list is the instrument; no instrument means no measurement.
Two: show me a gap list from a real client --- the questions where the brand should be named and is not. If they cannot produce one, they do not run audits, they run reports.
Three: how do you score a citation. A mention is not a citation, and a 0-to-3 scale takes five seconds to explain. If the answer is "we track mentions," they track the wrong column.
Four: what do you publish. Agencies that publish their method, their probes, their zeros, are auditable. Agencies that keep the method inside a deck are selling mystery, and mystery is how you get billed for nothing.
Five: how do you separate classic rankings from AI citations in reporting. If the dashboard is one blended number, you cannot tell whether the work moved the page or the engine.
The red flags
Guarantees of being in ChatGPT or any engine --- nobody controls that, and anyone claiming to is lying. Rankings-only reporting for an AI engagement --- you hired them for the board they are not showing you. No prompt set, no schedule, no baseline --- an audit without a baseline is a testimonial with extra steps. And price before questions --- an agency that quotes before it knows your category is selling a package, not a measurement program.
What a real engagement looks like
Baseline first: a fixed prompt set run across two or three engines, scored, with the gap list as the output. Then content built one question at a time --- quotable blocks, entity naming, schema --- shipped in cycles with probes between them. Then the same prompt set re-run on schedule, deltas recorded, zeros published alongside wins. The full loop is what GEO search tracking describes, and it is the only structure in which you can tell whether the agency did anything.
Where to start without an agency
Run the baseline yourself. Ten questions, two engines, score every answer 0 to 3, and read the audit method --- the free diagnostic does the first pass against your host. You will either have the gap list an agency should be quoting, or you will have proof you do not need one yet. Either outcome is cheaper than a listicle.