Testing a promise about AI answers

1. 01

   Which prompts, and who chose them?

   A prompt set written to be won reports an excellent share of an audience that does not exist. The prompts should be the questions your buyers actually ask, and you should recognise them.

   If insteadThe prompt set is proprietary, described as “thousands of queries”, or supplied only as a count.

2. ---

   02

   Is a citation being promised, or content that might earn one?

   No assistant operator sells placement in an answer. A promise of citation is therefore either a promise about somebody else's system that the supplier does not control, or a loose description of writing and markup work.

   If insteadThe distinction is treated as pedantry. It is the entire difference between a deliverable and a hope.

3. ---

   03

   Do your own logs show the assistants fetching the pages?

   Being read is upstream of being cited and is the part you can verify alone. Separate the three purposes: an index build makes citation possible later, a training crawl returns nothing ever, and a user-triggered fetch means somebody was waiting for an answer at that moment.

   If insteadNothing in the logs, months into the engagement. Whatever was produced has not been read, and citation was never in reach.

4. ---

   04

   Was the same measurement taken before the work started?

   Without a baseline over the same prompts, a share is a single reading of an unstable system. Assistants disagree with themselves between runs, so a first measurement taken after the fact cannot show a change.

   If insteadThe baseline is described rather than shown, or was taken on a different prompt set.

5. ---

   05

   What is being paid for that would be worth doing anyway?

   Most honest work here is not specific to assistants: state a claim, attribute it, publish the figure with its period and base, keep the entity records consistent across sources. It earns citations as a side effect and stands up whether or not the field changes shape again.

   If insteadThe proposal is priced around a file — an llms.txt, a schema block — as the deliverable rather than the by-product.

It is real

A named prompt set you recognise, a baseline on the same prompts, saved answers, and logs showing the pages being fetched. The number will still be noisy and the supplier will say so.

---

It is not

The prompt set is withheld, or citation itself is the promised deliverable. Neither can be verified, and the second cannot be delivered by anyone.

---

Cannot tell

The work is sound, the pages are being fetched, and the citation share moved by a few points across two months. That is inside the noise of the instrument. The honest report says the inputs are right and the outcome is not yet measurable.

This check, pointed at us.

Our own GEO practice

We sell answer-engine visibility and we have never sampled our own citation share. Not once, on no prompt set. What we measure is the layer beneath it — fetches by declared agents at our own edge, 99 of them from 10 agents on 18 August 2026, of which four were triggered by somebody's live question — because that is what a server can see. The glossary entry for citation share is the only quantity in the corpus with a formula and no worked example, and the reason is that we have no measurement of our own to put in it.

SinceNot yet, and the gap is named rather than closed quietly. A sampled prompt set with a published baseline is the next measurement this practice owes, and until it exists we describe what we count as fetches rather than as citations.

Citation

Being named or linked inside an answer a machine assembled — the unit of answer-engine visibility, and the one your own server cannot see.

---

Citation share

How often you appear among the sources for a set of questions — a number whose entire meaning lives in the set, which is the part suppliers omit.

---

Assistant fetch

One request for a page by a declared automated client of an AI system — the only unit of attention in answer engines that a site can count on its own server.

---

Retrieval

How an assistant chooses which pages to open before it writes — a selection problem, not a ranking one, and it rewards different things.

---

llms.txt

A file proposing which of your pages a machine should read first. It is a community convention rather than a standard, and nobody is obliged to fetch it.

---

GEO

Being citable by an assistant that is assembling an answer, rather than merely rankable by an engine that is returning links.

---

Entity

A thing a machine holds a record of — a company, a person, a product — as opposed to a set of pages that happen to mention a name.

Every checkThe whole glossary

If the answers did not arrive, that is the finding.

A supplier who cannot supply a base and a period has told you something, and it is not that they are disorganised. We do this for a living and the questions land differently when they come from outside.