Learning

How to choose a GEO agency, and what an audit must include

Six checks: frozen bank, saved responses, rates with denominators, source list with owners, the right guarantee in the right place, inspectable raw output.

· 4 sections

“How do I know if a GEO agency is serious, or selling me repackaged SEO?” You do not need to take anyone’s method on trust. Use six checks to assess whether the provider can explain its approach and show evidence for its results.

The checklist

1. A frozen prompt bank, in layers. The questions are the instrument. Ask to see the bank: which questions, how many, and in which layers (unbranded, brand-named, competitive). A serious provider shows the bank before running anything, explains why each theme is there, and freezes it before the first prompt runs, so the numbers repeatable months later use the same questions.

2. The raw responses, saved and readable. Every answer from every model, kept. If the deliverable is a summary or a dashboard, ask where the responses live. You should be able to read exactly what an assistant told a buyer, not a provider’s interpretation of it.

3. Rates with denominators. Not scores, rates. “Mentioned 26% (103 of 396 responses)” is evidence. “Visibility score: 72” is a decoration; ask what population stands behind it and the conversation ends. Every percentage in the report should carry its count, and a percentage without one is a red flag on its own.

The rate quoted above is from our invented sample, printed the way every report prints its figures.

4. A source list with owners. The report should name the domains the assistants cite for your category, how often, and who controls each: the regulator page nobody can edit, the retailer page you can rewrite, the earned media you can pursue. An audit that ends at “you need more content” never measured anything.

5. The right guarantee, in the right place. If they promise positions in ChatGPT, walk out. Answers can vary by user, context and run. A provider should be able to commit to the instrument and the run: a prompt bank complete and useful for your industry, and a study that produces findings you can act on. That does not mean anyone can promise that later actions will move the numbers. Ask the provider to explain what its method can and cannot show.

6. Raw output you can inspect, on demand. Ask: “can I see the responses behind this number?” If a provider cannot explain how a metric is defined or show the evidence supporting its report, ask how you can audit it. For Google Search, Google’s guide says llms.txt, special markup and breaking content into tiny pieces are not requirements for its generative AI features. That guidance does not establish what other assistants use. Read Google’s guide.

The questions that expose weak agencies

Ask these in the first call and listen for specifics:

  • “Which questions would you ask about my category, and why those?” Vague answers mean the bank does not exist yet.
  • “If the same question returns different brands on different runs, how do you handle it?” The answer must involve repetitions and denominators, not “we take a snapshot”.
  • “Which sources does ChatGPT cite for my category, and how do you know?” The only honest answer starts with “we measure it”, not “we know from experience”.
  • “What does the report look like if nothing improved?” Weak agencies have no answer. The correct one: the report says nothing moved, with the same numbers behind it.
  • “What do you measure that a traditional SEO tool does not?” If they describe rankings or traffic, they have not understood the question.

What a real audit contains

For reference, the deliverable list we consider the floor: the frozen bank in its layers, every response saved and readable, the four outcome rates each printed with its denominator (mentioned, recommended, named as the best, with a purchase path), share of category conversation against each competitor, the ordered source list with owners, and a working session that leaves actions with names on them.

Notice what is absent: scores, positions, promises about outcomes, and “executive summary” slides that summarize data nobody can inspect. Notice what is present instead: a provider who commits to what can actually be committed — the bank and the run — and says plainly that whether your actions move the numbers later is something nobody can promise in advance.

How to use this list

Take it into every pitch, including ours. It is the evaluation we would apply to anyone selling AI visibility, and it is what our own study has to survive too: /agencia-geo-mexico/ shows the frozen bank, the saved responses and the denominators behind it. The AI visibility study explains the service and its deliverables.

Request the study