What to Check in a GEO Agency Proposal

The first thing to check in a GEO agency's proposal is three details: how many times they measured, which AI engines they tested, and when they measured. If a proposal doesn't spell these out, the citation rates and ranking numbers that follow are hard to verify.

Most proposals on the market lead with summary figures: citation rates up by some percentage, conversion rates several times higher. What matters more than the number itself is asking how it was produced, and that alone tells you whether the proposal holds up.

There are five ways to check this, and none require asking the agency for a full evidence packet. Rereading the proposal itself, asking a couple of direct questions, or typing a query into a search box yourself is enough.

Check How Many Times They Asked

Generative AI doesn't give the same answer to the same question every time. Concluding "our brand is recommended first" from a single answer is no different from flipping a coin once and declaring the result a rule.

Check whether the proposal states a repeat count. How many times they asked, and whether each round used a new chat window or continued the same one, are different conditions. In the same chat window, an earlier question and answer can influence the next one.

There's no fixed rule for how many repeats are enough. But one or two rounds make it hard to separate luck from a real signal, and a result that holds up after close to ten rounds is far less likely to be a fluke.

A proposal built on only one or two rounds may just be a lucky single answer. Data that states "appeared in X out of Y attempts" as a fraction is more trustworthy than a single captured screenshot of one good result.

Check Which Engines, and When

A company that shows up in ChatGPT often doesn't show up in Gemini or Perplexity. If a proposal only says "cited in AI search" without naming the engine, ask again.

A proposal covering only one engine isn't automatically a problem. If it states which engine it checked and limits its claims to that scope, you can simply read the rest as unchecked.

Timing matters just as much. AI models keep updating, and the same question can produce a different answer weeks later. Without a measurement date, there's no way to know if a result still holds.

Some proposals take a result from one of four engines and generalize it to "cited across AI search." It's more accurate to ask, engine by engine, how many times out of how many attempts a citation appeared.

Reproduce the Same Question Yourself

Take the exact question and repeat count from the proposal and run it yourself, spread across several days rather than all at once. Asking everything in one day leaves you at the mercy of that day's one unusual answer, while spreading it out shows whether the result actually holds steady.

Don't stop at typing the proposal's sentence word for word. Real customers don't type an agency's exact phrasing into a search box; they ask in their own words, with slight variation. Asking the same underlying question three or four different ways shows whether a good answer was a fluke tied to one phrasing, or something that holds up across phrasing.

Save each answer, either as a screenshot or copied text, along with the date you asked. When you compare against the proposal's numbers weeks later, that record is what lets you pin down what actually changed.

If the proposal presented results across several engines, reproduce it on each of those engines separately. Checking one engine and taking the rest of the proposal's claims on faith undercuts the point of verifying anything.

An agency reluctant to disclose its exact question wording may simply not want its results reproduced and checked. An agency that discloses both the wording and the repeat count up front is one that's comfortable with you verifying it yourself.

If what you get running it yourself diverges sharply from the agency's reported results, the first things to ask about are differences in measurement timing or chat-window conditions.

Check When the Next Measurement Happens

As covered above, AI models keep updating and the same question's answer can change over time. That means it's worth confirming whether re-measurement and reporting happen during the contract term, not only at the end.

Most proposals include a measurement taken at the start of the contract, or a final result at the end, but many stop short of naming a recurring schedule for re-measuring and reporting in between. Ask whether a monthly or quarterly re-measurement cadence is written into the proposal or contract.

Without a re-measurement cadence, the entire contract term rests on a single number from the outset. If the model changes months later and the results shift, there's no way to know it happened.

Separate the Agency's Own Numbers From What You Verify Yourself

Citation-rate growth, conversion rates, and client results published on a GEO agency's website or in its proposal are, for the most part, figures the agency compiled and released itself. That doesn't mean the numbers are wrong, but whether a third party gets the same result measuring the same way is a separate question.

When you receive a proposal, ask who compiled a given number, how, and across how many cases. A figure that highlights a growth rate without disclosing the case count or measurement period is worth asking about again rather than accepting outright.

The same applies to client case studies. Look for a start date, an end date, and the comparison period alongside any growth percentage. A case that states a growth rate with no start or end date leaves you unable to tell what period was actually compared.

For the same reason, absolute language like "100% cited in AI answers" or "guaranteed top ranking" isn't evidence on its own. As shown earlier, since the same question produces different answers on repeat, promising a fixed outcome amounts to guaranteeing something the agency doesn't actually control.

Ask about repeat counts, engines tested, and measurement dates, reproduce the results yourself, and separate the agency's own reported numbers from what you've verified independently. Do all of that, and a single-page proposal starts to reveal what it didn't show at first glance.