How we measure

SEOAIO measures your visibility across traditional search and the generative engines, the discipline now called GEO (Generative Engine Optimization) or AEO (Answer Engine Optimization). We run on one doctrine: measure, never manufacture. Every number in the product is something we actually observed, and every number tells you how it was observed. This page explains the method.

The doctrine

A visibility score is only useful if it reflects reality. So we do not estimate where we can measure, we do not extrapolate a single engine into a market-wide claim, and we never present a modeled figure as an observed one. If a pillar has not been measured, the dashboard shows it as not present. It is never filled in with a plausible-looking placeholder.

Sampling AI answer engines

For each site we run a fixed set of prompts against each AI answer engine we support, several samples per prompt, because these engines are not deterministic. We count a citation only when the engine actually returns your domain as a source in its answer. A passing mention of your brand name is not a citation, and we do not count it as one. Results are reported per engine: a Perplexity number is labeled as a Perplexity number, never generalized to engines we have not sampled.

Confidence bands and sample sizes

Small samples produce uncertain percentages, and we say so. Every sampled rate ships with a Wilson score confidence interval and the exact sample size behind it. When you see "33%, CI 12-65%, 3 prompts / 9 samples," that width is honest: it tells you the measurement is early, and it narrows as more samples accumulate. We would rather show you a wide interval than a precise-looking fiction.

Provenance on every number

Every metric carries one of three labels. Measured: we observed it directly, and you can see when and how. Derived: computed arithmetically from measured values, with the inputs disclosed. Modeled: an estimate, clearly marked as one, and never blended silently into measured figures. If a collection run fails, it writes no aggregate at all; a crashed measurement is a gap, not a zero and not a guess.

How often we measure

We re-check your pages every 24 hours and re-sample the AI engines every 7 days for each site you monitor. Checking your pages is cheap, so we do it daily. Asking the AI engines costs real money on every question, so that runs weekly and the sample size is set by your plan.

The cadence matters more than it sounds. Anything we tell you about movement is only as good as the number of times we looked, which is why we publish the interval and the confidence range beside the figure rather than just the figure.

When we do the work ourselves

We will help you fix what we find. That creates an obvious problem: a company that does the work and then grades it is marking its own homework, which is exactly what most tools in this category do. So the separation is a rule, not a preference, and these are its four clauses.

• We never sit in the path your site serves from. No proxy, no DNS takeover, no runtime injection. Your server sends your bytes.
• The test that decides whether something moved does not know who moved it. The same statistical gate is applied whether you made the change, your agency did, another tool did, or we did. We do not get an easier test for our own work. This is the clause that does the real work.
• Every applied change is logged with who applied it. Including when the answer is us. Your change log tells your work apart from your agency's and from ours.
• Nothing is ever applied without you approving it. We do not run an agent that edits your site while nobody is watching, and we are not planning to.

What we do NOT do

• No manufactured citations. We never seed, plant, or fake mentions of your brand anywhere.
• No auto-published spam backlinks. Link schemes carry penalty risk, and we will not take it with your domain.
• No black-box single number. Every score decomposes into the measurements that produced it, and you can inspect each one.
• No silent extrapolation. One engine's result is never presented as an all-engine result.
• Every question in its own language. A tracked question is asked, and its answers filed, in the language you set for it (English by default, or one of 7 others). Google AI Mode is measured in English only, so a question in another language reads as not measured there, never as an English answer.
• No unattended edits to your site. Other tools run an agent that rewrites your pages on a schedule. We do not, and the approval step is not a setting you can switch off.

What this instrument cannot tell you

Every measurement has limits. Ours are these, and we would rather you read them here than discover them in a number that surprised you.

• The engines change underneath us. A model updates and yesterday's answer is not today's, with nothing on your site having changed. We cannot prevent that. What we can do is record when OUR ruler changed, which is why the instrument changelog exists and says in writing which dates a trend must not be read across.
• The same question can get different answers. These systems are not deterministic, and two identical prompts minutes apart can disagree. This is why a single answer is never a measurement here, why we sample repeatedly, and why every share carries a confidence band instead of a clean number. A vendor quoting you a precise figure from one run is quoting noise.
• We measure the questions in your set, not every question. Sampling one prompt a hundred times makes the result precise. It does not make the prompt representative. If your tracked questions are not how buyers actually ask, a tight confidence band can still sit around an unrepresentative number, and no amount of repetition fixes that. Choosing the set well matters more than anyone in this category admits.
• We cannot see inside an answer we did not sample. Results are personalised, regional and account-dependent. We report what our sampling observed, under the conditions it observed it, and we do not model what some other user might have seen.
• Only one engine tells us what it read. Claude exposes the pages it retrieved before answering, so "read but not cited" is measurable there and nowhere else. For the other engines that reads as not measured, because it is, and we will not infer it.
• Correlation is not attribution. We can show that a change happened and when. We cannot prove your edit caused it, because nobody can run the counterfactual, and any vendor showing you a clean before-and-after without a control is showing you a story.