When an AI audit claims it found every market-leading brand, it sounds exhaustive. It also leaves the most important question unanswered: how often did each brand actually show up when a buyer asked a relevant question?
If a brand appears once across thousands of test runs, it can still count toward “100% coverage.” And if a brand only appears in the background retrieval layer, the buyer reading the final answer might never see it at all. Before you use a coverage percentage to judge your performance, you need to know exactly what was counted.
A recent 37,000-run audit shows why these distinctions matter. The researchers tracked 533 cataloged brands across 19 commercial sectors, separating the process into retrieval, mention, and actual recommendation. Comparing their aggregate numbers against their per-query results makes one thing clear: “coverage” is meaningless without a defined denominator and observation layer.
Keep in mind that this experiment was a bounded, single-day snapshot using specific providers, markets, prompts, and an LLM-curated brand catalog. The tiers in the study describe that specific catalog—they aren't permanent laws dictating how your business will always perform in AI search.
“Appeared at least once” is not “appeared every time”
The study reports 100% aggregate surface coverage for its market-leader tier. That just means every cataloged leader showed up at least once across the entire pool of model setups and retrieval conditions. It does not mean every leader appeared on every query, or that a buyer actually saw them recommended.
Meanwhile, the study’s 77% per-query rate comes from a much narrower setup: Claude Sonnet 4.6 running on high reasoning effort with native web search. Even then, that 77% measures whether a brand surfaced in any retrieval layer during relevant, in-sector queries. It is not the rate of final-answer recommendations.
Because those two percentages use different denominators and different model setups, you can't treat them as a simple funnel from overall coverage to single-setup recommendations. Any report that compares them needs to keep those boundaries intact.
When you build your own reports, separate the brands that were ever observed from how often they actually appear on relevant buyer questions. Lock those measurements to a specific model configuration and observation layer. Otherwise, a massive experiment can look reassuringly thorough while telling you almost nothing about what a real customer will actually see.
Specialist and regional brands fared much worse in the study's aggregate numbers—roughly half never surfaced in the tested runs at all. That is a real finding about those specific test conditions, though it doesn't prove that no prompt or retrieval system could ever find them.
Separate access from selection
The audit breaks visibility into four distinct stages: material that was never retrieved, material retrieved without a brand mention, a brand mention without a recommendation, and an actual recommendation.
If you control the retrieval pipeline, your system logs might let you inspect all four stages directly. If you are testing a commercial AI assistant, you usually only see the final answer and whatever sources it chooses to cite. A missing citation doesn't prove the model never pulled your page behind the scenes. If you can't see the retrieval layer, mark that stage as unknown rather than guessing.
Even when an AI system retrieves your page, it might skip recommending your brand for entirely fair reasons. Maybe the buyer asked for an integration you don't support. Maybe the source page mentions your company without giving enough detail to compare it against competitors. Or maybe your business was only named as a counterexample.
Read the actual sentences surrounding your brand. Lumping every mention into the “recommendation” column strips away the exact context you need to decide what to fix.
Changing the retrieval path helps some brands, but not all
The study also tested different retrieval methods, comparing native web search against external retrieval. Some lower-profile brands only surfaced when external retrieval was plugged in, proving that the route the system takes to find evidence directly changes which brands show up.
Other brands stayed invisible no matter which retrieval route was used. That absence doesn't automatically tell you why they were missing. An experiment like this can't peer inside a model's training weights to see what it inherently “believes” about a category, nor does it prove that indexing and source availability don't matter.
When a brand goes completely missing, the culprit could be the test prompts, the reference catalog, the retrieval coverage, thin source material, or how the model filters results at the very end. Blaming “training priors” just because a brand didn't show up skips over all those practical steps.
You need the same discipline when testing buyer personas. In the study, changing the buyer's requirements shifted the recommendation sets, especially for mid-tier brands. That isn't necessarily a flaw—an AI assistant should change its recommendations when the buyer's needs change. You should only expect consistent answers when the underlying buyer requirements stay the same.
Make sure your coverage metric helps you make decisions
Whenever you look at a coverage number, start by asking how it is defined. Are you looking at brands observed at least once, the share of relevant runs, or actual recommendations shown to the user? Ask which models and retrieval setups were blended together to get that total.
To make this practical for your own business:
- Pick a concrete buying question and spell out the customer's exact constraints.
- Run that question repeatedly under one consistent product setup and save the full responses.
- Check whether your business was completely absent, merely mentioned, or actively recommended—and read how it was described.
If you have access to the retrieval logs, inspect the documents that were pulled in. If the retrieval layer is a black box, keep your conclusions strictly at the answer level. Before you order a content rewrite or a technical overhaul, compare the visible evidence about your business against the specific requirements in the prompt.
If your brand shows up in the retrieval layer or gets mentioned, but the model notes that you lack a key capability, check your source pages to see if that capability is stated clearly. And if the model accurately rules your business out because you genuinely don't fit the buyer's constraints, that isn't an AI visibility bug—it's an accurate filter.
A good coverage report makes all of these layers easy to inspect. It tells you where your brand was seen, under what setup, and how often. Collecting a list of brand names is just the starting line; knowing what to change requires the context behind them.
