Verumgraph is a method for publishing knowledge as an entity graph with persistent identifiers and source provenance — readable by people, verifiable by machines, citable by search engines and AI systems.
First deployment: Kanon Kosmetyków · ~700 addressable records · in production
People increasingly ask ChatGPT, Perplexity, Claude and AI Overviews — and read a single synthesized answer instead of ten blue links. In that world your website is no longer read by your customer. It is read by the machine that answers your customer.
Most content was never written for that reader. Marketing prose is ambiguous, undated and unverifiable — so AI systems either skip it or, worse, guess. Brands end up invisible in AI answers, or misrepresented by them.
Verumgraph takes the opposite approach: publish knowledge the way reference databases do — as records with identity, evidence and dates — so machines can trust it enough to cite it, while people still get pages built for humans.
Any single one of these is a known technique. Verumgraph is the discipline of applying all five together, from day one, to the same body of knowledge.
The unit of publication is a record, not a blog post: a product, a user need, a question, an ingredient, a brand. Each is a node in a graph with typed relations — question → need context → product → variant → ingredient — so knowledge composes instead of duplicating.
Every record carries a registry-style ID (PC-000021, QE-000027, CTX-000063, ING-000001) that never changes. Machines can track an entity across pages, exports and years — the way scientific databases and standards bodies do it.
Each statement is typed — verified fact, manufacturer declaration, or editorial interpretation — and carries its evidence (e.g. a label photo) and an as-of date. Uncertainty is data, not fine print. An empty field beats an unverified value.
One record, two renderings from the same source of truth: an HTML page for people, and machine layers for AI — schema.org JSON-LD wired into one graph via @id, a clean record.json per entity, and an llms.txt that tells AI systems what the base is and how to cite it.
Entities link out via sameAs to global identifiers — Wikidata, industry registries — so the graph is not an island. External anchors are added only on exact matches: fewer links, zero wrong ones.
This is a shortened, real machine record from the first deployment — the kind of object an AI system can fetch, parse and cite without scraping a single line of marketing copy.
// …/products/biolaven-eye-cream/record.json { "@context": "https://schema.org", "@type": "Dataset", "@id": "…/products/biolaven-eye-cream/#dataset", "name": "BIOLAVEN eye cream — reference record", "identifier": "PC-000021", "dateModified": "2026-08-07T09:14:00Z", "license": "https://creativecommons.org/licenses/by/4.0/", "distribution": [{ "@type": "DataDownload", "encodingFormat": "application/json" }], "record_fields": { "application_area": "eyes", "use_mode": "leave_on", "primary_function": "moisturize", "fragrance_status": "declared — source: label, as of 2026-06" } }
Shortened for display. The live record, and hundreds like it, are one click away below.
Kanon Kosmetyków is a Polish curated knowledge base about natural cosmetics — no ads, no rankings, no affiliate links. Every statement has a source and an as-of date. It is the first full deployment of the Verumgraph method and serves as its living proof.
The graph maps real purchase intent to verifiable product data: user questions resolve through need contexts to product records, market variants (label, GTIN, ingredients as printed) and individual ingredient entities anchored to Wikidata.
Everything below is public. Open the links, view source, fetch the records — the method is fully inspectable.
When a customer asks an AI about your category, your product should appear — described from your verified record, not from a model's guess.
Your clients are asking what you're doing about AI search. Verumgraph is a deployable answer: a record model, a context vocabulary, a provenance workflow and the technical layer to publish it all.
Tell us your category. We'll show you what your first ten records would look like — and what AI systems currently say about you.