Measuring AI Citation Rate: An Open Protocol

Why a protocol

AEO is full of citation-rate figures and short of comparable ones.

A vendor says a client reached 34% citation rate. An agency reports 12%. A tool shows 61%. None of these numbers can be compared, because none is accompanied by the four things that would make it meaningful: which prompts, how many samples, which engines, and over what window.

This document specifies a method. It is deliberately plain and deliberately public, so that results produced with it can be checked and reproduced. We will use it for any primary research we publish, and we would rather other people used it too than that we owned it.

Definitions

Fix these first, because most disagreement about AEO numbers is definitional rather than empirical.

Citation. An answer contains a link to, or an explicit named attribution of, the target domain. A brand name with no link is a mention, not a citation.

Mention. The brand name appears in the answer text, with or without a link.

Citation rate. Citations divided by total sampled answers, expressed as a percentage, for a given prompt set, engine and window.

Sample. One submission of one prompt to one engine, producing one answer.

Observation window. The calendar period over which samples were collected. Required, because engines change.

Protocol

1. Build the prompt set

Between 25 and 30 prompts. Fewer and the confidence intervals are too wide to act on; more and the sampling cost rises without much added signal at this stage.

Sourcing, in descending order of value:

• Sales discovery calls and pre-sales questions. The highest-fidelity source of how buyers actually ask.
• Support tickets with evaluative intent.
• Search Console queries with question shape.
• Competitor comparison queries: "X vs Y", "alternatives to X".
• Category questions: "best tool for job", "cheapest way to outcome".

Composition rules:

• No branded self-queries. Asking an engine about your own product name measures recall, not discoverability. Exclude them from the main set; track separately if you want.
• Cover both registers. Insider terminology and plain language, because query rewriting draws on both.
• One question per prompt. Compound prompts produce compound answers that cannot be scored cleanly.

Then freeze the set and version it. Changing prompts mid-study is the single most common way AEO measurement produces uncomparable data. If you must add prompts, start a new version and report both.

2. Choose engines and record versions

Report per engine. Never blend into a single figure — engines diverge systematically, and a blended number hides the common case of being strong in one and absent from another.

Minimum set: ChatGPT, Gemini, Perplexity. Add Claude, Copilot or others if your buyers use them.

Record the model or version where the engine exposes it, and note the date. An unexplained change in your numbers is frequently an engine update rather than anything you did, and without version notes you cannot tell the difference.

3. Set the sample count

At least 10 samples per prompt per engine. Thirty or more if you need to detect changes below roughly 10 percentage points.

This is the parameter people cut, and cutting it invalidates everything downstream. A single observation per prompt yields a citation rate with confidence intervals so wide that any period-on-period comparison is noise.

Control what you can: consistent region, no logged-in personalisation, samples distributed across times of day rather than fired in one burst.

4. Score each answer

For every sample, record:

• Cited: yes or no, with the URL.
• Mentioned: yes or no.
• Position, if the answer is a recommendation list.
• All competitor brands named.
• All domains cited.
• Any factually incorrect statement about the target brand, quoted verbatim.

Score against a written rubric, not judgement in the moment. Ambiguous cases — a brand named inside a quotation from another source, for instance — must be resolved by a rule fixed before scoring begins, or the rule becomes whatever produces the nicer number.

5. Compute

• Citation rate per engine, with a 95% confidence interval. Report the interval. A rate of 18% ± 7 is honest; 18% alone is not.
• Mention rate per engine.
• Share of voice: target mentions divided by all brand mentions across the sample, against a fixed competitor set.
• Mean position when the answer is a ranked recommendation.
• False-claim frequency: incorrect statements per 100 answers.

6. Report

State, always: prompt set version and size, engines and versions, samples per prompt, observation window, region, competitor set, and the scoring rubric.

A citation rate without these is an assertion. With them, it is a measurement someone else can check.

Known limitations

Stated plainly, because a protocol that hides its weaknesses is not much of a protocol.

Personalisation is not fully controllable. Engines vary answers by account history and geography in ways that cannot be entirely neutralised.

Engines change under you. A model update mid-window can shift results independently of anything you did. Short windows and version notes reduce but do not remove this.

Citation is not influence. Your content can shape an answer without attribution. This protocol measures attributed citation only, and therefore understates real influence — an important caveat when the number is used to judge content investment.

Scoring involves judgement. A written rubric narrows it. Two careful scorers will still disagree on some answers, so double-scoring a subset and reporting agreement is good practice.

Sampling cost is real. A full cycle at 30 prompts, 4 engines, 10 samples is 1,200 queries, repeated per period. This is the practical reason AEO platforms exist, and any honest protocol should say so rather than implying the work is free.

On tooling

Every step here can be done by hand once. None of it can be done by hand repeatedly, which is the only way it produces trend data.

If you are evaluating platforms against this protocol, the questions worth asking a vendor are: can I freeze and version a prompt set, how many samples per prompt do you take, do you report per engine, do you report confidence or variance, and do you surface false claims explicitly? Several tools we assessed report a single blended visibility score and cannot answer the second or fourth questions at all. Our comparison matrix records which do.

Contributing

If you run this protocol, we would like to see the results — including results that contradict ours. Send them through our submissions page. We will publish methodologically sound studies whether or not their conclusions suit us, which is the point of publishing the method first.