Entity-Attribute-Value: The Data Model That Organizes Knowledge

Every service-business site eventually claims things no one verified: "same-day service," "competitive prices," a review that reads like it was written by the owner. Human writers invent because they have nothing real to say and a word count to hit; model writers invent faster. Prompts do not fix this — a model told to never fabricate will still fabricate the moment it has no source, because filling the gap is what text generators do. The fix has to be structural.

Entity-Attribute-Value is that structure. An entity is the thing a page is about — a service, a product, a city branch. Attributes are the questions you can ask about it: price band, duration, warranty, service area. Values are the literal answers, observed rather than invented. Logged this way, facts stop being testimonial prose in a CMS and become records a coverage matrix can grade: every entity × attribute cell is solid, thin, or missing. "Definitive pillar" stops being a prose aspiration and becomes a coverage property. Literal values matter because embeddings approximate meaning — they cannot reliably hold exact prices, model numbers or policy limits, which is precisely what engines need to extract verbatim.

Provenance is the quiet detail that keeps the stored facts honest. Facts sourced from private records stay visible to the operator but never reach published content. Facts cleared for publication enter the EAV matrix as observed values — observed price band, observed duration, verified review — scoped to their city, with the latest observation winning per entity × attribute and a quality floor for reviews. From that moment, the model no longer distinguishes between a fact someone asserted and a fact the business actually produced — both are rows with provenance.

Which is exactly what published content should draw from the model. Documents weave publishable EAV rows into their sections, so a page can state the observed price band for a job in that city, quote a verified review, and cite a real duration — sentences that are true because they are downstream of records, not because a model was asked to be careful. A fact budget per document keeps this disciplined: each page draws from a bounded, tracked set of facts instead of re-consuming the same rows until they blur.

And the pre-publish audit closes the loop. Fact-integrity checks verify that rendered claims trace back to rows; label-collision rules keep distinct facts from collapsing under one attribute label. If a cell has no proof, the pages for that cell simply cannot make the claim — the silence is enforced, not requested. Hallucination does not get punished by good process. It gets starved.