How to read an AI visibility report
Guide · Continuous Intelligence · 4 min read · last verified 2026-07-25
The most common way to misread an AI visibility report is to open it, look at the one big number on the front, and stop. That number is the least informative thing in the document. The parts that change what you do next — the per-question breakdown, the absences, and the evidence trail — sit underneath it, and this is a guide to reading them in the right order.
What to look at first
Open an AI visibility report by going straight to the per-question presence, not the headline score. Find the highest-intent buyer questions and check three things: are you present, is the description accurate, and which sources the answer cited. That tells you where you are winning or losing on the queries that actually decide deals — which the aggregate score cannot.
Only after reading the per-question view should you glance at the aggregate, and then only as a rough thermometer. A report is a diagnostic tool; the front-page number is the temperature, and you do not treat a patient by staring at the thermometer.
What share of voice, mentions, and citations actually mean
These three metrics answer different questions, and conflating them is a classic misread:
| Metric | What it measures | What it does not tell you |
|---|---|---|
| Mentions | How often your brand is named in answers | Whether the mention was positive or sourced |
| Citations | How often you are named with a linked source | Whether buyers act on it |
| Share of voice | Your presence relative to competitors | On which specific questions you win or lose |
A mention is being named. A citation is being named with a live source link behind the claim, which carries far more weight — both to the model and to the buyer clicking through. Share of voice compares you to a competitor set, so it is only as meaningful as that set is honest. Read all three together; any one alone flatters or scares you for the wrong reasons.
Absence versus presence — read the zeros first
The most valuable cells in the report are the empty ones. Presence tells you where you already compete; absence tells you where you are eliminated before a human ever evaluates you. Sort the report by high-intent questions where you score zero and a competitor is recommended — that ranked list of absences is the working output, far more than any question you already win.
Absence also compounds. A question you are missing from today feeds fewer downstream citations tomorrow, because the sources that would have referenced you never form. Treat a zero on a high-intent question as urgent, not as a data point.
Per-question versus aggregate — and the averaging trap
An aggregate score is an average, and averages hide their own distribution. You can hold a healthy overall number while being absent from every one of your top five revenue questions, if you happen to dominate a pile of low-intent definitional queries. The report will look fine and your pipeline will not.
Always weight by intent when you interpret. A report that lets you filter to "comparison and selection questions only" is telling you something the blended number actively conceals. When a score moves, ask which questions moved — a gain concentrated in throwaway queries is not the same as a gain on the questions buyers ask before they buy.
Read the evidence trail, not just the verdict
A trustworthy report shows its work: for every mention or citation, the source the assistant leaned on and a link to check it. Read that trail. If the answers that name you cite only your own pages, your presence is fragile, because third-party sources tend to be cited more heavily than vendor-owned ones. If a rival wins a question on a single review-site page, you know exactly what to out-corroborate.
The evidence trail is also your defense against believing a number you cannot verify. If a report asserts a position but cannot show the source, treat the claim as unconfirmed rather than measured.
What not to over-read: the score itself
The headline score is the most gameable and most over-read figure in any AI visibility report. It moves when a model updates, when the question set changes, or when sampling happens to differ between runs — none of which mean your real position shifted. Adding prompts alone can move your score without moving your position at all.
So do not manage to the number. Manage to the trend on a locked methodology and to the queue of absences you are closing. If the underlying method changed between two reports, the delta is noise; only a fixed benchmark makes a score comparison honest.
How to sanity-check a report before you act
Before you act on any reading, confirm four things: the method was locked between periods, the run count is high enough that sampling variance is not driving the delta, the competitor set is the real one, and each claim links to a source. According to the Princeton GEO study (2024), citing sources and adding statistics are among the strongest drivers of whether content gets surfaced at all — so a report that cannot show its own sources is failing the same standard it is measuring you against.
Reading a Magrios report in practice
A Magrios report is built to be read in exactly this order: per-question presence first, absences ranked into a queue, every mention and citation carrying its source link, and the aggregate framed as a locked trend rather than a trophy. The design reflects the doctrine — the score is not the KPI; the blind-spot queue and the locked trend are, because those are the parts you can actually act on and re-measure.