Entity Coverage: How Many of Your SKUs Can AI Recognize?

Tuesday’s quarterly review. A German garden-tool brand selling into the US has a slide showing its name in a healthy share of tracked AI answers, and the room moves on. Then support pastes a screenshot: an assistant recommends the brand’s cordless mower, names last year’s model, quotes a 38 cm cutting width for a deck that now measures 42 cm, and files it under string trimmers. The brand was recognized; the SKU was not. Nobody knows how many of the hundred storefront products the assistant gets right.

Being mentioned and being recognized are different capabilities; entity coverage measures the second — whether AI can name a specific SKU, state its key attributes correctly, and compare it.

Three numbers, one metric

Answering a comparison question, an assistant resolves products into entities — a graph, not a row — as why AI sees a product graph rather than a SKU argues. It shortlists you only when your facts match the ones it holds. A model number at the end of a title, variants behind a dropdown, one product under three names: each lets it merge SKUs into a fuzzy entity or answer at category level.

“Not mentioned” is the shallow failure. The expensive one is mentioned wrongly: last year’s spec, a digit off in the model number, a use case the product contradicts. Buyers compare on bad numbers and discover it at the returns desk.

Metric	Definition	Question it answers
Recognition rate	Share of sampled SKUs whose answer names a specific model	Does the assistant know this SKU exists?
Accuracy rate	Share of recognized SKUs whose key attributes are all correct	Is it talking about this model?
Comparability rate	Share entering the candidate set in a comparison or scenario, with a reason	Will it spend a sentence comparing you?

These are diagnostics, not a scorecard: they tell you which layer to fix and sit with the rest of the GEO measurement and growth work. Prompt wording is part of the definition: “have you heard of X” measures brand memory; “how do the A and B differ” forces attribute commitment, and the rate collapses. Change the prompt and last month’s number is gone.

Sampling: 100 SKUs, four strata

Sampling by sales volume builds a useless number: the sample is mostly bestsellers, which AI gets right anyway. Draw 20 hero SKUs, 20 long-tail, 20 launched this season, 20 variants and bundles. Variants need their own stratum because shared copy is where models swap attributes, and hero SKUs hide the damage. Freeze the list first — SKU ID, canonical URL, aliases, three key attributes — because a sample edited mid-quarter is a new ruler. Then confirm retrieval crawlers can read those pages, or a low score is crawl coverage in disguise.

Ask questions the attribute must answer

Prompt shape	Example	What it tests
Named	“What is the cutting width of the X 420 and how many square metres does it handle?”	Recognition and attribute accuracy
Comparison	“How do the 420 and the 450 differ, and which suits a sloped yard?”	Whether you enter the comparison
Scenario	“Best robotic mower under $1,500 for a quarter-acre yard with dogs”	Whether you make the candidate set

Cover every sample with one shape; hero SKUs get all three. Run each three times, a day apart, on a fixed engine, region, and language. Log eight fields: date, engine, locale, prompt, answer, wrong attributes, cited URLs, candidate yes or no. A ninth field and nobody runs it.

Read accuracy before recognition. Low recognition with high accuracy is an off-site corpus gap. High accuracy with low comparability means the facts are right and the reason to choose you is missing: fix answer units and corroboration, not more copy.

A five-step playbook

Step	Action	Owner	Passes when
1 Freeze	Lock 100 samples and three prompt shapes in one sheet	Growth	Sample reproducible, prompts unedited
2 Sample	Run the prompts, store raw answers and citations	Ops	Every row has an answer, not an impression
3 Compute	Recognition, accuracy, comparability, split by stratum	Growth	The weakest stratum is visible
4 Classify	Bucket errors: model mismatch, wrong attribute, variant confusion, no corroboration	Content and engineering	Three or more instances per bucket
5 Fix and retest	Change the largest blocking bucket, rerun the same sheet in 30 days	Owner	A before-and-after you can explain

Order fixes by whether the gap blocks comparison, not by volume. A model or GTIN mismatch is structural: you cannot be placed beside a competitor in one table. Variant confusion yields wrong answers; attribute inconsistency breaks the comparison; missing corroboration only suppresses recommendations. Fix the first three before buying PR. The root cause is usually master data, not copywriting — see why product master data comes before GEO — and a corrected attribute still needs turning into a reason a buyer can use.

Example: a 100-SKU audit on a US storefront

Example — a hypothetical DTC seller of portable power stations shipping to the US and Germany; no real brand, and the figures are illustrative, not a benchmark.

First pass: 100 SKUs, English, two assistants. Recognition 62%, accuracy 54%, comparability 19%; hero SKUs sit near 80% recognition, variants below 40%. Two errors dominate: capacity quoted from the previous generation, and “can it fly” answered as a product attribute when it depends on airline battery limits.

The fix did not start with articles: twelve variants got canonicals and model-bearing titles, category wording was unified, and capacity moved into each page’s first sentence. Thirty days later recognition and accuracy rose; comparability barely moved. That flat line is the useful result: the next constraint is off-site, in reviews and community threads.

Where this breaks

Under 30 SKUs. Measure the whole catalog; report which models fail, not percentages.

Revenue closes in a marketplace. Platform fields define the entity, so test listing attribute fill and model mentions in reviews.

Custom or made-to-order. No stable model exists; measure category recognition and delivery boundaries.

A new brand with no off-site corpus. Low recognition is normal; settle whether the brand entity exists before making it a content KPI.

Month-over-month comparison. Model updates, re-indexing, and regional differences move all three numbers; hold engine, region, and language fixed; treat one swing as noise.

For this month, choose one action, freeze the 100-SKU list, add model, GTIN, and three key attributes per row, run one named and one comparison prompt three times each, and compute recognition and accuracy only. Sort errors into model mismatch, wrong attribute, and variant confusion, fix the largest bucket, and rerun the sheet in 30 days.