Teach the model your facts.

How a model decides what to say about you

Every LLM answers brand questions through the same chain: disambiguate the entity, resolve it to a known identity, retrieve facts, synthesize a response. A brand fragmented across domains, sister companies, and half-maintained profiles breaks the chain at resolution. The model hedges, or quietly substitutes a competitor it holds with more confidence.

LLM Optimization works on the memory layer above any single answer: what models already hold about your brand in weights and knowledge graphs. There are two ways into a model, training data and retrieval, and they follow different playbooks. LLMO covers both, so that any model, in any session, returns facts that are accurate, specific, and sourced from surfaces you control.

One canonical entity file, one sameAs chain

The foundation is a canonical entity file at /.well-known/entity.json: Organization JSON-LD declaring the legal entity, parent and sub-organizations, and a sameAs chain that connects every owned domain, LinkedIn, Crunchbase, and register listing into one identity. The identical block publishes on the about page and in the site footer. A short file, and it resolves a company-sized confusion.

Groups with sibling brands get disambiguation hardening: a brand-family page stating each sibling's name, legal entity, product, and canonical URL, so models stop crossing wires between companies that share a parent. And when encyclopedias are out of scope, no Wikipedia and no Wikidata, the grounding load ships entirely through first-party surfaces: entity.json, layered schema, llms.txt, structured about pages, confirmed by press and registers.

Triples, co-occurrence, and vector neighbourhoods

Models hold facts as triples: subject, predicate, object. Brand isRegulatedBy authority; brand headquarteredIn city; brand offers product. We write out the exact triples your brand needs models to hold, then place each one in at least three independent confirming sources. A fact asserted once is a claim; confirmed three times, it becomes knowledge a model will repeat.

Embeddings are co-occurrence machines, so co-occurrence gets engineered rather than left to chance: at least three sentences per article placing the brand name in the same sentence as the target attribute, accumulated across the content base until the association is dense. Pronouns are the silent failure; 'we' and 'our platform' are stripped during embedding and the signal is lost. The brand name does the work.

Retrieval is similarity search, which makes vocabulary a ranking factor. Pages written in legacy vocabulary embed into the wrong neighbourhood and surface for the wrong category. We rewrite money pages in category-native vocabulary so the brand embeds beside its true competitors. That is vector-neighbourhood control, applied page by page.

The acceptance test: fresh-session interrogation

Entity work is verified, never assumed. We interrogate each model in a fresh session, no history and no context, and check that it returns the correct identity, the correct corporate chain, and citations from owned or approved sources. The pass condition is specific: right facts, right structure, owned-source citations, zero hedge language.

A monthly hedge scan reads answers for 'appears to be', 'reportedly', and 'according to its website', then traces each hedge to the under-evidenced attribute behind it; three independent sources later, the retest comes back clean. And because models keep separate per-language entity memories, every entity foundation republishes per language path. An English-only truth source leaves every other language guessing.