What “100% Brand Coverage” Can Hide in an AI Audit

A report says it found every market-leading brand. That sounds comprehensive, but it leaves an important question unanswered: how often did each brand appear for a relevant buying question?

A brand encountered once in thousands of runs can count toward complete coverage. And a brand encountered during retrieval may never appear in the answer the buyer reads. Before using the percentage to judge performance, establish what was counted.

A 37,000-run audit makes those distinctions useful. It studied 533 catalogued brands across 19 commercial sectors, separating retrieval, mention and recommendation. Its aggregate and per-query results illustrate why “coverage” needs a denominator and an observation layer.

The experiment was a bounded, single-day audit with selected providers, markets, prompts and an LLM-curated reference catalog. Its tiers describe that catalog. They are not permanent categories that determine how every business must behave in AI search.

“Appeared at least once” is not “appeared every time”

The study reports 100% aggregate surface coverage for its market-leader tier. Every catalogued leader surfaced at least once across the full set of model configurations and retrieval conditions. That does not mean every leader appeared on every query, or that a buyer saw it recommended.

The reported 77% per-query rate has a narrower experimental scope: Sonnet 4.6 with high reasoning effort and native web search. It measures surfacing in any retrieval layer on relevant, in-sector queries. It is not a final-answer recommendation rate.

The two percentages therefore differ in both denominator and configuration coverage. They are not a direct conversion from overall coverage into recommendation frequency under one setup. A report comparing them should preserve those differences.

For your own report, distinguish brands ever observed from how often each appears on relevant questions. Keep those measurements within a stated configuration and observation layer. Otherwise a large experiment can look comprehensive while providing little evidence about the answers a particular customer is likely to receive.

Specialist and regional brands had much lower aggregate coverage in the study. Around half never surfaced in the tested run set. That is a substantive observation about those conditions, not proof that no suitable prompt or retrieval system could ever find them.

Separate access from selection

The audit's staged view distinguishes material that was not retrieved, material retrieved without a brand mention, a mention without a recommendation, and an actual recommendation.

When you control the retrieval system, logs may let you inspect those stages directly. When you use a commercial assistant, you may only see the answer and its visible sources. An absent citation does not establish that a page was never retrieved internally. Mark that stage as unknown if you cannot observe it.

A retrieved page can fail to produce a recommendation for reasonable reasons. The buyer may need an integration the supplier lacks. A source may discuss the company without offering enough information to compare it. A named business may appear as a counterexample rather than a suggested choice.

Read the sentence around the brand. Counting all mentions as recommendations removes information you need to decide what to fix.

Retrieval changes helped some brands

The study compared retrieval conditions, including native search and external retrieval. Some lower-prominence brands surfaced only under external retrieval. That establishes that the route to the evidence can change the observed set.

Other brands remained absent across the tested routes. Their absence does not isolate the cause. The experiment cannot directly observe a model's training-time beliefs about which brands belong in a category. It also does not prove indexing or source availability is irrelevant.

Possible explanations include the selected questions, the catalog, retrieval coverage, available source material and the system's later selection. Declaring a training-prior barrier from the final absence skips those alternatives.

The same care applies to buyer personas. Different requirements changed recommendation sets, particularly for some middle-tier brands. A recommendation that changes with the buyer's needs can be appropriate. Consistency is useful only when the underlying need remains comparable.

Ask what the coverage figure lets you decide

Start with its definition: brands observed at least once, share of relevant runs, or recommendations shown to the customer. Ask which models and retrieval conditions contribute to the total. Then choose a concrete buying question and record the customer's constraints. Repeat it under a consistent product setup and keep the full answers. Identify whether your business was absent, merely mentioned or actually recommended, and inspect how it was described.

Where the retrieval path is available, check the documents that entered it. Where it is hidden, keep your conclusion at the answer level. Compare the visible evidence about your business with the requirements in the question before prescribing a rewrite or a technical fix.

If the business appears but a relevant capability is missing, verify whether that capability is stated clearly in the available source. If the answer accurately rules out the business for the requirement, that is not automatically a visibility defect.

A useful coverage report makes those distinctions inspectable. It tells you where the brand was observed, under which conditions and how often. A complete set of names is a starting point; deciding what to change requires the context behind them.