On this page
Here’s an illustrative version of a pattern we see often: a brand tracks 30 prompts and comes back with a healthy, respectable visibility score, trending up, nothing that would make anyone nervous in a monthly review. Then someone actually reads the answer to “best [category] for [the one use case that drives half the pipeline]” and the brand is not in it, on any engine, on any of the last six runs. The blended score was never wrong. It just never promised that the query that pays the bills was covered, and nobody checked until now.
This is a different failure than the usual ones. It is not a crawling problem or a missing schema field. It is what happens when a single number has to stand in for thirty separate answers, and thirty separate answers do not all matter the same amount.
Why an average can’t tell you this
A blended score is a mean, and a mean treats every prompt as interchangeable. Rankscale’s own breakdown of what a composite score hides points out that the same headline percentage can come from genuinely different situations: being mentioned often but never cited, being cited but always buried past position six, or simply not coming up half the time, and calls these three “a Trust Gap, a Position Gap and a Detection Gap” that each need a different fix (Rankscale). None of those three show up differently in the one number at the top of the dashboard. They only separate once you look at the prompts individually.
The same masking happens across query types, not just across causes. Metricus, in its explainer on why composite AI visibility scores mislead, gives the example directly: a brand can score 90% on broad category questions and 0% on the specific, industry-level ones, with the broad questions outnumbering the narrow ones in the set and dragging the blended number up regardless (Metricus). Gracker’s guide to calculating a visibility score makes the same point from the engine side: “a brand can be dominant on Gemini and invisible on Perplexity, and a single average hides that entirely,” and notes that a composite number is fine for a headline but “hides where the problem actually lives” (Gracker). Swap “engine” for “query” and the same arithmetic applies: a handful of strong prompts can carry a dozen weak ones without the average ever moving enough to look wrong.
It is worse the more concentrated your revenue is
A 30-prompt tracking set usually isn’t 30 equally important questions. It’s a handful of best-of and comparison prompts that map to real buying moments, and a longer tail of definitional and category questions that build background trust but close nothing by themselves. Choosing the prompts worth tracking covers why those higher-intent prompts deserve the heaviest weight in the first place, precisely because a weak score on one of them costs more than a weak score on five low-intent ones combined. A plain average doesn’t know this tradeoff exists, so it reports the same healthy number whether the weak prompts in the mix are the unimportant ones or the one that actually matters.
This is also why the dead zone survives for months instead of getting caught. Nobody opens a dashboard to go looking for a problem when the top-line number is fine and trending the right direction. The gap only gets found when someone happens to run the exact query by hand, usually because a prospect mentioned it, and that’s a bad way to find out your best-converting question has a hole in it.
How to check whether this is happening to you
Pull your tracked prompts and rank them, not by score, but by how directly each one maps to a real purchase decision: best-of and head-to-head comparisons first, problem-led questions next, pure definitional questions last. Then look at the scores for the top tier on their own, separate from the blended number. If the top tier is weaker than the overall average, the average has been quietly flattering you, and the gap is real rather than noise. Measuring share of voice walks through the per-prompt math behind the rollup, which is the same breakdown you need here.
Check consistency too, not just the current reading. A prompt that returns you once in six runs and a prompt that returns you zero times in six runs can both show up as “weak” in a single snapshot, but they are different problems: one is noise around a real but fragile presence, the other is a genuine absence. Tracking competitors in AI answers is the same discipline applied to who else shows up in that dead-zone query. If a specific competitor owns it consistently across engines, you are not missing data, you are losing a decision, and that changes what the fix looks like.
Finally, check it per engine as well as per prompt, since the two gaps compound. Why different AI visibility tools report different scores explains how prompt set, engine coverage and weighting can each move a blended number independently; a query that is merely weak on one engine can look like total absence once it’s averaged in with five engines where you were never going to be the answer anyway.
Treat the top-tier prompts as their own scoreboard
The fix isn’t a better formula. It’s refusing to let one number speak for queries that don’t deserve equal weight. Keep the blended score for the monthly summary if that’s useful to the rest of the business, but track your highest-intent prompts as a separate, smaller scoreboard that has to be read on its own, because that is the one a hidden zero actually costs you money on.
Run a free Visibility Check on the specific query you suspect is a dead zone rather than your category in general, and if it comes back empty, Monitor tracking will tell you whether that’s a one-off miss or a standing pattern across every run for the last several weeks. A query that matters and returns nothing, six times in a row, is not noise. It’s the one gap in the plan worth fixing before the next one you add.