Entity coverage maps: a working template.

What an entity is

An entity is any distinct, nameable thing a model can hold a representation of: a company, a product, a person, a concept, a place. A B2B brand has multiple entities to manage: the organisation itself, its primary products, its key people, and the category concepts it wants to be associated with.

What the map captures

An entity coverage map is a structured inventory of what a model currently knows about each of those entities: how it describes them, how confidently, and where the description comes from. It is not a keyword gap analysis. It is not a backlink audit. It is specifically about the model's representation, which is built from training data and (for some engines) live retrieval - not from your sitemap or your link profile.

What it tells you

The map tells you three things: what the model gets right, what it gets wrong, and what it doesn't know at all. Each has a different fix. Confusing them leads to wrong interventions.

For a B2B company, I map across five dimensions. Each dimension gets a score of 0-2: 0 = absent or wrong, 1 = present but incomplete or inconsistent, 2 = accurate and consistent across sources.

Dimension 1: Category placement

Ask a model: "What category does [brand] operate in?" and "What problem does [brand] solve?" Compare the answers to your own positioning. If you build observability software but the model calls you a "DevOps tool," that's a category placement gap - you're being surfaced for a different set of buyer prompts than the ones that should trigger you.

Dimension 2: Customer description

Ask: "Who is [brand] designed for?" and "What size of company typically uses [brand]?" Mismatches here affect whether you appear in prompts from the right buyer segment. A model that describes your enterprise product as SMB-focused is sending the wrong buyers to your competition. This gap is almost always traceable to an inconsistency in how you're described across external sources.

Dimension 3: Differentiation accuracy

Ask: "How is [brand] different from [primary competitor]?" The model's answer should reflect your actual differentiation. If the model's differentiation narrative reflects how you were positioned two or three years ago, it's because that older positioning is more strongly represented in the training data: it appeared in more documents, more consistently, for longer.

Dimension 4: Key people

Does the model know who founded the company? Who leads it? Are those people described accurately? For B2B brands, executive credibility is part of the entity footprint. A model that doesn't know the CEO's background or confuses them with someone at a different company has a person-entity gap that affects the brand's overall source-worthiness perception.

Dimension 5: Factual accuracy

Founding date, headquarters, funding stage, key partnerships, major product releases. Errors here are usually fixed fastest via Wikidata and Wikipedia corrections, since those are the reference sources models weight most heavily for factual claims about organisations.

The mechanics

Run the five dimensions across three engines: ChatGPT, Perplexity, and Claude. That's fifteen question sets. Take notes verbatim: don't paraphrase what the model says, because the exact language it uses is the data. You're looking for the frequency of specific descriptions, not just whether the model gets it right or wrong.

Then compare what each model says against three sources: your own homepage copy, your Wikidata entry, and your most-cited Wikipedia text (if a page exists). The gaps between what the model says and what your authoritative sources say are your intervention targets.

The important nuance

Not all gaps are fixable by changing your content. If a model consistently describes you in a way that reflects how you were positioned two years ago, that description is likely dominant in the training data - the only way to shift it is to produce enough consistent, authoritative new signals that the newer description eventually becomes the more-cited one. That takes time. The short-term fix is Wikidata and Wikipedia; the medium-term fix is the editorial and reference source work described in building source-worthiness for AI search.

The output of an entity coverage map is a prioritised list of interventions, not a content plan. The interventions fall into three buckets by speed and mechanism:

Bucket 1: Reference source corrections (hours to fix, weeks to propagate)

Wrong or missing Wikidata entries. Outdated Wikipedia descriptions. Crunchbase entries with the wrong industry classification. These can be fixed in hours and affect the model's entity profile within weeks as training data refreshes and retrieval indexes update. Always fix these first - they're the highest leverage per hour of any GEO intervention.

Bucket 2: Description convergence work (medium-term, 3-6 months)

If the model's description of your category placement or customer type is inconsistent across sources, the fix is getting your current, accurate description into as many authoritative sources as possible: consistently, specifically, and repeatedly. This means editorial outreach that targets the right characterisation, conference talks where you describe yourself clearly, and named-author content on your own site that uses the exact language you want models to learn. This is the core of what entity coherence work looks like in practice.

Bucket 3: Factual gaps (variable, 6-18 months)

Missing information - the model doesn't know your founding story, your key partnerships, your flagship customer logos - is slower to fix than inaccurate information, because you're building net-new model knowledge rather than correcting existing knowledge. The mechanism is the same (reference sources, editorial, community) but the timeline is longer and the results are less predictable.

The map is a starting point, not a complete audit. For a full picture of AI search visibility (including citation share measurement, source-surface analysis, and a prioritised 90-day roadmap), that's the GEO audit service. But the five-dimension map will tell you in two to three hours whether you have an entity problem, and that's the question to answer before anything else.