What AI visibility tracking actually measures.

What does AI visibility tracking actually measure?

The honest answer starts with a subtraction. AI visibility tracking does not measure the inside of a model, because nobody outside the company that trained it can. What it measures is the public evidence those systems retrieve and the outcomes that follow, and it names the source of every reading.

Six families cover almost everything worth watching. They are the same families the free AIOInsights evaluation scores at a point in time, which is why a baseline and a tracking history speak the same language.

The six, and what each one is read from

• Semantic clarity. Whether the most important pages still say what the business does, who it serves and where it operates. Read from the pages themselves, in the HTML a machine receives rather than the text a browser paints after scripts run.
• Entity consistency. Whether names, domains, profiles, categories and locations still describe one business. Read from the owned pages and the public profiles that reference them.
• Authority and corroboration. Whether independent sources still say the business exists and is competent. Read off-site, which is the point: an owned page asserting expertise is a claim, and a third party repeating it is evidence.
• Retrievability. Whether a system trying to answer a question right now can reach, parse and quote the page. Read from robots directives, response codes, canonicals and the raw markup.
• Reputation evidence. Whether reviews, responses and their recency still support a confident judgment. Read from the review platforms themselves, never inferred from a star average alone.
• Visibility outcomes. Search exposure and qualified clicks where a platform provides them, public citations and mentions where they can be observed, and dated samples where a person ran a query by hand.

Why the sixth family is worded so carefully

The first five describe conditions we can inspect directly. The sixth describes results, and results are where measurement most often turns into invention. A tool that reports a single number for "your ChatGPT ranking" has either measured something narrower than the label suggests or made it up.

So outcomes are tracked with their provenance attached: this figure came from Search Console for this date range; this citation was observed on this page on this date; this sample was one person asking one question once, and it is reported as a sample rather than a position.

Unmeasured and zero are different findings

Every scoring engine here treats them as different, and it costs tidiness to do so. If a source cannot be reached, the reading is not observable, not zero. Zero says the thing was checked and found absent. A network fault reported as absence turns a bad afternoon into a business finding, and a business that acts on it spends money fixing something that was never broken.

That single rule is what makes a tracking history worth keeping. Measurements are appended, never overwritten, and each carries the version of the engine that produced it, so a change in the number can be separated from a change in how it was measured.