KPIs to Ask a GEO Agency For Before You Sign

No measurement can prove that a GEO agency caused a change in what ChatGPT or Gemini says about you, and that includes ours. The assistants swap models, refresh their indexes and reweight their sources on their own calendar, so a brand can appear more often in March for reasons that have nothing to do with the work done in February. What measurement can do is narrower and still worth a contract clause: it can make every claim the agency makes checkable by someone other than the agency. That is the job of the protocol below.

We should state our position before the steps. PSentry is a measurement tool, not an agency. We do not optimise anything, so we have no stake in whether you hire one. We do have a stake in the numbers being read honestly, and the steps below apply to any tool used for the measuring, ours included.

Step 1: freeze a baseline you own, before anything changes

A baseline is a dated record of how the assistants answer a fixed set of questions about your market, taken before the agency touches your site. Without it, every later number floats free.

Four things go into it, and you keep a copy of all four:

• The prompt set. The actual questions, written the way your buyers would ask them, in each language you sell in. Not your brand name typed into a chat box.
• The assistants. Which ones, named. A result on one assistant says little about another.
• The dates. At least two runs some days apart, because the same question asked twice can return different names. One run is an anecdote.
• The raw answers. The unedited text of every response, plus the sources each one cites. A score without the answers behind it cannot be audited later.

The agency can propose the prompt set. You should approve it, and it should not change during the contract without a written reason, or the before and after stop measuring the same thing.

Step 2: split the GEO agency KPIs into two lists

An honest proposal separates what the agency controls from what it can only influence. Most disputes begin when those two lists are merged into one dashboard.

Deliverables the agency controls. These can be checked in an afternoon, with a browser:

• pages rewritten or created, listed by URL;
• specifications and prices published as text rather than locked inside PDFs or images;
• pages produced in each export language, not only in the language of the head office;
• technical fixes, such as crawler access in robots.txt, each with a before and after.

Outcomes nobody controls. These are measured, never promised:

• Mention rate: out of the answers to your prompt set, how many name you, per assistant and per language.
• Competitor presence: which local competitors are named on the same questions, and whether that list moves.
• Cited sources: whether your own pages start appearing among the sources the assistants cite.
• Factual accuracy: how many answers state something wrong about you, counted against a short list of facts you supply.

Pay for the first list. Track the second. A contract that ties fees to the second list is a bet on machinery that neither side operates.

Step 3: add control prompts the agency does not work on

This is the step that separates a measurement from a story. Alongside the prompts the agency is working on, keep a second, smaller set about topics it will not touch: another product line, another market, a question about your sector in general.

The logic is the same as a control group. If the mention rate rises on both sets at once, the likeliest cause is a change inside the assistant, such as a new model or a new index. If it rises only on the prompts the agency worked on, the case for the agency is much stronger. It is still not proof, but it is evidence that survives a sceptical reading.

Agree before signing what counts as a change. Answers vary between runs, so a movement of one or two mentions out of twenty can be noise. Write down the threshold and how many consecutive runs must show it. Deciding that after the numbers arrive invites everyone to read them generously.

Step 4: write into the contract who does the measuring

An agency that does the work and also grades it is marking its own exam. That does not make it dishonest. It does mean that the person paying needs access to the same raw material.

Three clauses cover most of it. First, the raw answers, prompts and dates are delivered as an export at every review, not only as a slide with a score. Second, the method behind any composite score is written down well enough that you could rebuild the number yourself. Third, you may run your own measurement in parallel on the same prompt set, with whatever tool you choose.

For that parallel measurement, a tool such as PSentry queries ChatGPT, Claude, Gemini and Perplexity twice a month, keeps a score per language and market, lists the competitors named in your place and the sources cited, and exports the underlying answers to CSV. It does not cover the overview features inside search engines, and two scans a month is a series of snapshots, not a live feed. Those limits belong in the contract as well, whichever tool you use.

Step 5: read the proposal for the KPIs that should worry you

Some metrics in a proposal tell you more about the seller than about your brand. The broader warning signs of an offer with nothing behind it are covered in how to spot a GEO vendor selling nothing. These are the ones that turn up specifically in KPI tables:

• A ranking position in AI answers. Answers are generated text, not a list of ten results. A brand can be named first in one run and third in the next.
• A proprietary visibility index that you cannot recompute from the answers. If the number only exists inside the vendor's dashboard, it is not a KPI you can enforce.
• AI referral traffic as the headline result. It is worth tracking, but many answers name a supplier without producing a click, so traffic undercounts presence and can move for reasons unrelated to it.
• A promised improvement by a fixed date. The question to ask is what happens to the fee if an assistant changes model during that period. The answer tells you how the risk has been priced.
• Results reported only for the assistants where things went well. Agree the full list of assistants and languages at the start, and report all of them every time.

Step 6: set the review rhythm before the first invoice

A single run tells you very little, because answers vary from one run to the next for reasons explained in why one check proves nothing. What can be read is a trend across several runs, set against the baseline and the control prompts.

That makes the first review date a decision, not a formality. Too early, and you are reading noise. Pick the date when you sign, write it next to the threshold from step 3, and agree that the conclusion at that review will be drawn from the export, not from the summary.

None of these steps tells you whether to hire a GEO agency. They tell you what you will be able to say about the work when it is done, which is the question that tends to get asked a year later, by someone who was not on the call.