What is a good citation rate?

It depends entirely on category competitiveness, so absolute benchmarks are misleading. What matters is your rate relative to named competitors on the same prompt set, and its direction over time. A rate of fifteen per cent that was eight per cent last quarter is a healthier signal than a static thirty per cent.

How often should I sample prompts?

Daily for a small set of high-value prompts, weekly for the broader set. Less often than weekly and you cannot separate a real change from ordinary variance; much more often and you are paying for noise.

Why do my numbers differ from my agency's?

Almost always prompt set or sampling differences. Answers vary by region, account history, engine version and time of day. Two honest measurements of the same brand can differ substantially, which is why the prompt set and sampling method should be documented before anyone argues about the numbers.

Is share of voice better than citation rate?

They answer different questions. Citation rate asks whether you are a source. Share of voice asks how much of the category conversation you occupy relative to rivals. Report both; they move independently and a gap between them is diagnostic.

The AEO Metrics That Matter

The sampling problem comes first

Before discussing which metrics to track, one methodological point determines whether any of them mean anything.

AI answers are non-deterministic. Ask an engine the same question twice and you can get different sources, different framing, and a different verdict on your product. Vary the region, the account, the time of day or the model version and the spread widens further.

So a single observation is not a measurement. If you asked ChatGPT once whether it recommends your product and it said yes, you have learned that a favourable answer is possible. You have not learned your citation rate.

A usable AEO metric requires:

The same prompt set, held stable over time so comparisons are valid.
Repeat sampling — the same prompt asked many times, not once.
Multiple engines, reported separately, because they diverge systematically.
A fixed cadence, so trend lines mean something.

This is precisely the work that does not survive being done manually, and it is the main reason AEO tooling exists at all. Everything below assumes you have solved sampling. Without it, these metrics are decorative.

The six that matter

1. Citation rate

The proportion of sampled answers that link to or explicitly attribute your domain.

This is the closest thing AEO has to a north-star metric, because it measures the outcome you actually want: being the source. Track it per engine and per topic, never as a single blended number — a blended figure hides the common situation where you are strong in one engine and absent from another.

Watch for: a citation rate that moves while nothing else does. That usually indicates an engine changed its retrieval or citation behaviour rather than anything you did.

2. Mention rate

The proportion of answers where your brand is named, whether or not it is linked.

Mention rate is almost always higher than citation rate, and the gap between them is informative. A wide gap means the model knows you and is drawing on you without attribution — an authority and corroboration situation rather than a retrieval one. A narrow gap with both low means you are simply not in the picture.

Mention rate also matters commercially in its own right. In a recommendation answer, being named is most of the value; the link is a bonus.

3. Share of voice

Your mentions as a proportion of all brand mentions across your prompt set.

This is the competitive metric, and the one executives understand fastest. It normalises for category noise: if everyone's mention rate rose because the category got more attention, share of voice reveals whether you actually gained ground.

It requires a defined competitor set. Choose it deliberately and keep it stable, because quietly adding or dropping a competitor changes the number without any underlying reality changing.

4. Average position within the answer

Where you appear when you appear: first recommendation, third, or a passing mention in a closing caveat.

Being named first in a recommendation is worth substantially more than being named fifth, and a metric that treats those identically is throwing away the most commercially relevant distinction available. Any tool reporting mentions without position is under-reporting.

5. Sentiment

How the model characterises you: positive, neutral, negative.

Two reasons this matters more in AEO than in conventional search. First, the model is not showing your words; it is describing you, so the framing is out of your hands. Second, negative characterisations frequently rest on stale or wrong information — a price that changed, a limitation that was fixed, a feature you do have described as absent.

The actionable sub-metric is false-claim frequency: how often the model states something about you that is factually incorrect. That is a correctable problem, and correcting it usually produces the fastest measurable improvement available in AEO.

6. AI crawler serve rate

The proportion of AI crawler requests to your site that returned a successful response with actual content.

This is the only purely technical metric in the list, and it belongs here because it gates all the others. If GPTBot receives a 403 on your pricing page, no amount of content work will get that page cited.

Read it from server logs or a CDN, not from a content tool. Track it per crawler and per section: partial blocks affecting one directory are common and easy to miss.

What to leave off the dashboard

Prompt volume. How many prompts you track is an input cost, not a result.

Total mentions, unnormalised. Rises when you add prompts to the set, which makes it trivially gameable and analytically useless.

Sentiment scores to two decimal places. Sentiment classification of AI prose is noisy. Directional bands are honest; 7.34 out of 10 is false precision.

A single composite "AI visibility score". Vendors like these because they are easy to put on a chart and they always show improvement. They obscure which of the six underlying metrics moved, which is the only thing you can act on. Use them as a headline for executives if you must, but never diagnose from them.

Estimated AI traffic value. Built on assumptions about click-through from answers that vary wildly and are not independently verifiable.

Assembling the report

A monthly report that actually drives decisions:

Citation rate and mention rate, per engine, trended over at least three months.
Share of voice against a fixed competitor set.
Average position when mentioned.
Sentiment distribution plus a count of detected false claims, with the specific claims listed.
Crawler serve rate, flagged red on any drop.
A short list of shipped changes since the last report, so movement can be attributed to work rather than to weather.

Item 6 is the one teams skip and the one that makes the report useful. Without a record of what you changed, a trend line is an observation rather than a finding.

On tooling

Most platforms in the category cover metrics 1 through 5 adequately. The differentiators are elsewhere: whether crawler serve rate is included, whether position is reported at all, whether false claims are surfaced explicitly, and whether the tool attributes a change to a cause.

Attribution is the capability worth paying for. Knowing your citation rate fell is a fact. Knowing it fell because a competitor published a comparison page that now out-ranks you on eleven of your rewritten queries is a task. CiteCue's thirteen-factor analysis and its grounded assistant are built around producing the second kind of output; most tools stop at the first. See the full comparison matrix for how the field divides on this.