Clients now ask whether they appear when someone asks an AI assistant about their category, and the honest first answer is that measuring this is harder than measuring search rankings ever was. GEO aio aeo reporting covers three overlapping practices: generative engine optimisation, which is about being represented in generated answers; AI optimisation, the broader work of making content usable by these systems; and answer engine optimisation, which concerns being the source a direct question resolves to. The reporting problem is common to all three. The systems do not publish rankings, answers vary between users and sessions, and the referral data that reaches your analytics is incomplete by design.
We build these dashboards anyway, because the alternative is having no view at all of a surface that is clearly taking share from conventional search. But we build them with the limitations written on the page, next to the numbers, because the failure mode here is not missing data. It is confident data that is wrong, presented monthly until someone makes a budget decision on it.
What can be measured with reasonable confidence
Three things hold up. First, referral traffic from assistant domains, where the assistant passes a normal referrer and your analytics records it. That is real traffic, real sessions and real conversions, and it can be segmented like any other source. It undercounts badly, because many assistant interactions never produce a click at all, but what it does record is trustworthy.
Second, branded search volume and direct traffic movement. If people are encountering your brand inside answers without clicking, one of the visible effects is more people searching for you by name later. This is indirect and slow, and it is confounded by everything else you do, but a sustained shift is meaningful. Third, your own content’s technical eligibility: whether pages are accessible to the relevant crawlers, whether structured data is valid, whether key facts appear in extractable form. That is a property of your site, so it can be checked precisely.
Citation share and how to treat it carefully
The metric most vendors sell is citation share: run a set of prompts, record which brands and sources get named, express yours as a percentage. It is useful as a directional measure and it is not a ranking. The output depends on the prompt wording, the model version, the user’s location and history, and on randomness in generation itself. Run the same prompt twice and you can get different sources.
If you use it, and we do, then run it properly. Fix the prompt set and keep it stable over time, because changing the prompts changes the number more than any optimisation will. Run each prompt multiple times and report a distribution rather than a single result. Record the model and date on every measurement. And treat the absolute percentage as close to meaningless while treating the direction over several months as informative.
The numbers we tell clients not to trust:
- Any single-run citation percentage presented without a sample size or a date
- Share of voice figures that aggregate across different assistants, since they cannot be compared like for like
- Estimated impressions inside AI answers, which no one outside the platforms can observe
- Traffic forecasts built from citation share, which assume a click-through relationship nobody has established
- Week-on-week changes, which are almost always variance rather than movement
Structuring GEO aio aeo reporting so it stays honest
Our dashboards separate measured, inferred and estimated into three visually distinct sections, in that order. Measured holds referral sessions and conversions from assistant sources, plus the technical eligibility checks. Inferred holds branded search and direct traffic trends with a note that attribution is not possible. Estimated holds the prompt testing results, with the sample size, the date and the model stated on the panel itself rather than in a footnote.
The order matters because dashboards are read top to bottom and the first number is the one that gets quoted. Putting the weakest data at the top, which is what most GEO aio aeo reporting does because it is the most impressive looking, guarantees it will be misused.
What the reporting should actually change
A dashboard earns its place if it changes a decision. In this area the decisions are mostly about content. If prompt testing shows competitors named consistently on a topic where you are absent, that is a content gap worth investigating. If your pages are being cited but not clicked, the question becomes whether the page gives a reason to visit beyond what the answer already contained. If assistant referral traffic converts well but arrives in small volume, that argues for expanding coverage rather than optimising conversion.
What it should not change is budget allocation based on precise comparisons with paid or organic channels. The measurement is not mature enough to support that, and presenting it as if it were will eventually be found out. We say this to clients directly at the start of engagements, because the credibility cost of correcting an over-claimed number later is much higher than the cost of qualifying it now.
This will change. The platforms have commercial reasons to publish more data, and some already surface limited information to site owners. Until then, the useful version of GEO aio aeo reporting is a small dashboard with a short list of reliable numbers, an explicit statement of what is not measurable, and a prompt testing panel that is read as a trend and never as a score.
Keep reading: GEO – AIO – AEO · GEO