The LLM brand entity collision test
Test whether AI separates your brand, legal entity, products, parent, old names, and namesakes under deliberate ambiguity.
Recognition and identity resolution are separate tests
An entity collision happens when an LLM merges facts from a brand, product, parent company, former name, regional business, or unrelated namesake. Paired ambiguous and disambiguated prompts expose the problem. Score identity, relationships, attributes, abstention, and citations separately.
A correct brand mention is weak evidence of understanding. The answer may name the right company while assigning it another business’s location, treating a product as the legal entity, or carrying an old parent’s ownership into the present. Recommendation and sentiment can then be wrong even though basic recognition looks successful.
This entity work audits outputs rather than assuming that the model maintains one visible knowledge graph. In natural language processing, named entity recognition identifies mentions of people, organizations, locations, and other types. Entity linking or disambiguation connects a mention to the intended real-world entity. IBM’s overview of named entity recognition explains the distinction.
Schema.org provides vocabulary for entities and relationships such as Organization, Product, brand, and parentOrganization (Schema.org Organization). Markup can make visible page meaning more explicit. It does not guarantee that an LLM retrieves the page, resolves the identity, mentions the brand, or recommends it.
Create an entity disambiguation map
Map the identity before writing prompts. Include:
| Node | Evidence to record |
|---|---|
| Public brand | Official name, URL, logo, category, headquarters |
| Legal entities | Registered names and jurisdictions |
| Products | Product names, URLs, owner, current status |
| Parent or subsidiary | Relationship and effective dates |
| Former names | Rebrand dates and replacement name |
| Regional variants | Country domains, availability, local legal entity |
| Namesakes | Unrelated companies, products, people, or places |
Attach a current primary source to each relationship. Company documentation may establish product ownership. A regulator or registry may be stronger for legal status. Keep effective dates because acquisitions and rebrands create valid historical statements that are false in a current-tense answer.
Do not publish unnecessary legal details solely for an audit. Use information already public and appropriate for buyers.
Construct deliberate ambiguity tests
Run prompt pairs. The first uses the name alone: “What does Meridian offer?” The second adds a stable modifier: “What does Meridian at meridian.example offer to UK logistics teams?” Further prompts test relationships:
- Is Product X a company or a product?
- Who owns Product X now?
- Is Brand X the same organization as similarly named Brand Y?
- What was Brand X called before its rebrand?
- Which regions does Brand X serve?
- Compare Brand X’s product with its parent’s separate offering.
Use the same wording across products and save full responses. Run neutral identity questions before feeding corrective context. Otherwise the test measures the model’s ability to follow your supplied facts rather than reconstruct the public entity.
A useful panel varies one modifier at a time: domain, category, geography, parent, product, or former name. This shows which clue resolves the collision. The related persona-dependent recommendation test changes buyer constraints, while this test changes identity clues. Do not combine both variables in the first pass.
Score collisions and abstention
For each response, rate:
- Identity: Did it select the intended entity?
- Relationship: Did it correctly separate brand, product, and owner?
- Attributes: Are category, location, capability, and status assigned to the right node?
- Temporal scope: Does the answer distinguish current and historical facts?
- Abstention: When the prompt is ambiguous, does it ask for clarification or state uncertainty?
- Citation support: Do visible sources support the exact relationships?
Abstention can be the best result. A confident answer about the wrong Meridian is worse than “Which Meridian do you mean?” Set pass criteria according to risk. A wrong product owner may be severe in procurement, while omission of a former tagline is trivial.
Repeat priority prompts because generated answers can change between runs, modes, accounts, and regions. Report collision frequency inside the declared sample, with its size and conditions, rather than presenting a small panel as a universal accuracy statistic.
Diagnose before changing content
A collision may reflect ambiguous public naming, stale external profiles, contradictory owned pages, retrieval of the wrong result, or unsupported generation. Inspect citations and search results for the exact prompt. Compare page titles, organization descriptions, sameAs links, product ownership language, and dates.
The AI source-gap provenance map helps distinguish first-party and external evidence. If the answer repeats a false fact across several copied pages, inspect citation laundering.
Avoid claiming access to internal model state. You can say a response cited the namesake’s website while discussing your product. You cannot conclude that a particular hidden embedding or graph edge caused it.
Correct canonical identity relationships
Give each entity a stable canonical page. State the relationship in plain language near the relevant name: “Product X is a service from Company Y” or “Company Y acquired Brand X in May 2025. Brand X remains the customer-facing product name.” Add geography and effective dates where needed.
Use consistent names in title tags, headings, organization pages, product pages, documentation, and legitimate external profiles. Link parent, product, and regional pages with descriptive anchors. Retire or redirect obsolete pages when appropriate, while preserving needed historical context.
Structured data should match visible content. Use the most specific supported type and validate syntax, but do not add unsupported identifiers or unrelated sameAs profiles. Google’s structured data guidance says markup should represent the page content and warns against misleading markup (Google Search Central).
Leaf’s existing schema markup guide for AI search explains how to keep markup aligned with visible, verifiable entity facts.
Request corrections from high-impact third parties when they merge entities. Supply precise primary evidence. Then rerun the unchanged panel across relevant products and regions. Treat improvement as an observed association. Establishing that one edit caused the change, or that the change will persist, requires stronger evidence.
A collision audit should lead to content and technical tasks that also help human buyers. For a broader review of access, structured data, canonical sources, and AI-facing claims, request a SEO and AEO audit.
Frequently asked questions
What is entity disambiguation?
Entity disambiguation identifies which real-world entity a name refers to. “Leaf” could mean a company, a product, or a plant. A domain, category, and location can resolve the intended brand. This goes beyond detecting that the word looks like a name.
What is entity recognition in LLMs?
Entity recognition is the identification of named items and their types in text, such as a company or product. In an output audit, test whether the model detects the name and then separately whether it links facts to the correct entity.
How does entity optimization relate to knowledge graphs?
Clear, consistent relationships can help search and data systems reconcile entities, and structured data can express some relationships. Verify by checking rendered content, valid markup, search representations, and repeated answer tests. Do not infer a private knowledge-graph update from one correct response.
Why does an LLM confuse a brand with its parent company or product?
Possible causes include overlapping names, vague relationship language, old sources, copied profiles, or retrieval of the wrong page. Compare citations and canonical facts to diagnose the output. The public response rarely proves the internal cause.
Does schema markup help resolve brand entity collisions?
It can clarify machine-readable relationships when it accurately matches visible content, but it is neither sufficient nor guaranteed. Test before and after with fixed prompts, validate the markup, and inspect whether citations and identity accuracy change across repeated runs.
How do I know whether entity disambiguation is working?
Use ambiguous and disambiguated prompt pairs across relevant models and regions. Pass a high-risk identity check only when the intended entity, ownership, scope, and current facts are correct and supported. Track abstention and variability rather than relying on one favorable answer.
Maintain the identity record
Revisit the map after acquisitions, rebrands, product launches, and regional changes. Keep historical relationships dated instead of deleting all context. The goal is a public record that lets buyers and machines distinguish what is current from what was once true.