GEO Re-measurement: What to Put in the Contract

Re-run, do not redesign

Re-measurement means running the identical question set under identical conditions. Same questions, same punctuation, same number of runs per question, same treatment of web-enabled versus non-web-enabled.

Any change to the method breaks comparability with every previous round. If the method genuinely has to change, restart the series and say so — do not quietly splice the new numbers onto the old chart.

What to report

Two pages is enough:

Three metrics per platform: mention rate, share of voice, source coverage.
The change from last round for each.
The three questions that moved most, with the most likely reason for each.
Context breakdown — positive, neutral, negative — reported separately, never folded into the headline number.

Long reports are where weak methods hide. If a report cannot fit on two pages, ask what the extra pages are doing.

Reading a flat round

Some rounds do not move, and the useful response is diagnosis rather than reframing. Check in this order:

Did source coverage move? It usually moves before mention rate. If it moved and mention rate did not, the material is being seen but not preferred — a content quality issue, not a placement one.
Did one platform move while others did not? That points at which source ecosystem the round's work reached.
Did a competitor's share rise? Then the category moved and you held position, which is a different situation from standing still.
Did a model version change mid-round? Note it; do not attribute the flatness to your own work either way.

Attribution on the business side

GEO frequently produces no click. The realistic approach is:

Add "how did you hear about us" to your lead form, with an AI option.
Track brand-name search volume — it usually rises when AI mentions rise.
Ask on sales calls what the customer had already looked into. This also feeds the question set.

Do not construct a precise ROI figure. A defensible correlation, honestly described, survives scrutiny; an invented attribution model does not.

Three clauses that make an engagement verifiable

The question set is fixed for the service period and attached to the contract.
Each round provides raw records: question text, full answers per platform, cited source URLs.
Reported metrics are fixed as mention rate, share of voice and source coverage — no composite "AI ranking" score.

With those three, effect is something you can show a finance team. Without them, price is not worth negotiating, because there is nothing to accept against.

FAQ

Questions teams ask before they start

What if the provider wants to change the question set?

Ask why, and require that the old set keeps being run alongside the new one for at least two rounds. There are legitimate reasons to add questions — a new product line, a new market — but the original set has to keep running or the series is broken.

What should the contract say about results?

Name the indicators rather than a position: mention rate, share of voice and source coverage, measured against the frozen question set, with raw records supplied each round. A guaranteed ranking cannot be written because no such position exists.

How do we verify a provider's numbers?

Re-run a sample of the frozen set yourself under the same conditions and compare against the raw answers they supplied. Where the two diverge, the raw records on both sides show which step differs.

Should reporting be monthly or quarterly?

Monthly for most engagements. More often and the movement is lost inside the variation of generated answers; less often and there is no chance to correct course within the term.

What counts as a round that did not work?

Three indicators flat, source coverage included, with no version change recorded during the round. Where source coverage moved and mention rate did not, the round did work; the material is being read but not preferred.

Can results be tied to revenue?

Not through a precise attribution model. What holds up is a correlation described honestly: lead-form source fields, brand-term search volume, and what sales record about what a buyer already knew.

Re-measurement, acceptance and reporting