How is AI Visibility measured?
AI Visibility is measured through six metrics: citation rate, AI share of voice, coverage by engine, accuracy and sentiment, AI referral traffic and crawler access, all compared against a baseline frozen before the first publication. The protocol is declared on this page: anonymous session, date and engine on record, a frozen and versioned target-question set. Publishing the protocol forces you to follow it, which is exactly why it is here.
In how many questions is the firm cited?
This is the citation rate: the share of a fixed target-question set in which the firm appears in the answer. It is the primary metric, because it matches the product: appearing inside the answer the allocator reads. We also record position (cited source with link, mention without link, absent) and how the citation is framed.
What it does not prove: citation rate measures neither purchase intent nor the quality of what was said. A firm can be cited in many answers and be described with outdated information, which is why factual accuracy is measured separately.
What slice does the firm hold against competitors?
This is AI share of voice: out of every brand cited across that question set, how much is the firm’s. We log every brand that appears, not only the client’s, because relative position usually says more about where to put effort than the absolute number does.
What it does not prove: share of voice is comparative presence inside a question set we defined. Change the set and the number changes. That is why the set is frozen, versioned, and any change is declared in the report.
On which engines do the citations happen?
This is coverage by engine. Each system retrieves from different indexes and weighs sources differently, so the same question produces different outcomes on ChatGPT, Perplexity, Gemini, Claude and Google’s AI surfaces: AI Overviews and AI Mode. We always report engine by engine as well as consolidated.
What it does not prove: coverage is not audience. Appearing on an engine with fewer users in a given market is worth less commercially than appearing on a larger one, and the report does not weight that automatically.
What do the models claim, and is it correct?
Two readings of the same check: sentiment (how the firm is framed) and factual accuracy (whether what the model states is correct and current). We check each claim against the firm’s facts, log the error and trace its likely source.
This is the most underrated metric in the set: being cited wrongly is worse than not being cited. A strategy the firm no longer runs, a partner who has left, a figure from two years ago: the reader receives the error with the same confidence they would receive the fact.
What it does not prove: correcting a wrong claim depends on the model retrieving the corrected source again, which is not under our control.
How much traffic arrives from an AI answer?
This is AI referral traffic, read in GA4 through assistant referrers and the sessions that arrive from those origins with no search term. It is a supporting metric, not the primary one: much of a citation’s value is consumed inside the answer itself, with no click.
What it does not prove: low AI traffic does not mean no citations, and high AI traffic does not mean citations: it can come from a shared link. Traffic never replaces measuring citations directly, which is the most common mistake in this market.
Are AI crawlers actually reading the site?
This is crawler access, read from server or CDN logs: which user agents visited, which URLs, how often and with which response code. It is the only metric that shows cause rather than effect. A site that never receives GPTBot, ClaudeBot or PerplexityBot does not have a content problem; it has an access problem.
This is where the number-one cause of silent invisibility shows up: robots.txt allows the crawler, but the CDN or WAF returns 403 to that user agent. The file is correct and the bot never read a thing.
What it does not prove: crawler access does not guarantee indexing, and indexing does not guarantee citation.
What is the measurement protocol?
The protocol is what makes the numbers comparable from one month to the next. It is declared below and does not change without a note in the report. Anyone can copy a list of metrics; publishing the protocol is what forces you to follow it.
| RULE | WHAT IT MEANS | WHY |
|---|---|---|
| Baseline before publishing | The first measurement happens before any new page goes live | The earlier state cannot be reconstructed later |
| Frozen questions | The target-prompt set is fixed and versioned at the diagnostic | Changing the question changes the result |
| Declared mix | Category questions and branded questions, in a declared proportion | A set made only of the firm’s name inflates the number |
| Anonymous session | No login, no history, no personalization | Account history contaminates the answer |
| Date and engine on record | Every run stores date, engine and set version | AI answers are not publicly archived |
| Repetition per question | Each question runs more than once and is reported as an average | Answers are not deterministic |
| Competitors logged | Every brand cited enters the record, not just the client’s | It is what makes share of voice and citation gap possible |
Source: ReBo monitoring protocol, 2026.
What ReBo does not promise
We do not promise a guaranteed position in an AI answer, a deadline for the first citation, or a guaranteed citation. Nobody controls what a model answers, and any vendor promising otherwise is selling what it cannot deliver.
What we control are the variables that move citations (crawler access, readability, structure, factual density, verifiable authorship and external signal), and what we guarantee is measuring their effect, on the same ruler, every month, against the baseline. That is what makes it possible to say, with a number, where it worked and where it did not.
Frequently asked questions about measurement
Can you measure whether ChatGPT cites my firm?
Yes. You define a fixed set of questions your clients would ask, run that set in an anonymous session on ChatGPT and the other engines, and record which answers name the firm, in which position, and who appears in its place. Repeated monthly on the same ruler, it becomes a time series.
What is citation rate?
It is the share of the target-question set in which the firm appears cited in the AI answer. It is the primary AI Visibility metric, measured per engine and tracked against the baseline.
Why does the baseline have to come before publishing?
Because AI answers are not publicly archived and change with the index, the model and the date. If the first measurement happens after the first publications, the starting point is lost and any later result becomes a claim with no counter-evidence.
How often are the numbers updated?
Monthly, on the same ruler and with the same question set frozen at the diagnostic. Crawler access is read continuously, because it reveals an access problem before the month closes.
Do you use an off-the-shelf monitoring tool?
We use tools for collection and verification, but the ruler is ours: the question set, the session protocol and the competitor log. A tool that changes the question set on its own breaks the comparability of the series, which is precisely the asset.