Before a model can recommend you, it has to hold a stable concept of you: one name, one domain, one category, and a set of facts that agree with each other everywhere machines look. Engineers call that concept an entity. Google rebuilt search around entities in 2012 with the Knowledge Graph, framed at launch as "things, not strings".Google's 2012 announcement, Introducing the Knowledge Graph, marks the shift from matching query strings to resolving real-world things. LLMs arrive at the same place by a different route: training compresses everything written about you into a cluster of associations between your name, your category, your products, and your website.
How models resolve brands, products, and people
A model has no lookup table of companies. When a prompt touches your category, the model draws on associations learned from every page that mentioned your name during training: what you sell, who you serve, what site you live at. The more independent sources state the same facts about you, the more stable that cluster is. Embeddings: How Machines Represent Meaning covers the representation; what matters here is the failure mode. When sources disagree, or your name collides with someone else's, the model's concept of you smears across two things, and answers about you go wrong in ways you will not notice until you audit them.
The entity checklist
Entity establishment is consistency work. One canonical form of your name, one canonical domain, one short description of what you do, one category label, used verbatim everywhere machines read:
- Your own site, especially the about page and footer.
- Your Google Business Profile and knowledge panel, claimed and corrected.Google's knowledge panel documentation covers how panels are generated and how an entity's representative claims one and suggests changes.
- LinkedIn, Crunchbase-class company databases, and the industry directories machines crawl.
- Wikidata, where your identifiers and facts live in structured, machine-readable form.Wikidata's introduction: a free, collaborative knowledge base of structured data, organized so computers can process and reuse it.
- Organization markup on your site with sameAs links tying your domain to those profiles; Structured Data and Machine-Readable Facts has the implementation.
None of this is glamorous, and that is the point. You are removing every opportunity for a machine to learn a conflicting fact about you.
Wikipedia and Wikidata: a threshold, not a program
Two findings from OppAlerts' correlation research should cap how much you invest here. First, entity presence works as a threshold: brands present in Wikipedia and Wikidata are far more visible in AI answers, but past that line, added depth predicts little. Crossing the line mattered roughly [number: verify] times more than polishing after crossing it. Second, entity presence is largely redundant with strong backlinks: control for link strength and the independent contribution of Wikipedia and Wikidata presence drops to near zero.The threshold and redundancy findings come from the Wikipedia, Wikidata, and signal-importance analyses summarized in What Actually Correlates with AI Search Visibility, with method in The Research Behind This Guide. Treat entity presence as a checkbox to clear, and stop funding it once it is cleared.
Clearing the checkbox has rules. Wikipedia presumes a subject suitable for an article only when it has significant coverage in reliable sources independent of the subject.Wikipedia's notability guideline, the standard editors apply when deciding whether a company merits an article. If that coverage does not exist yet, a page is not achievable, and manufacturing one backfires: conflict-of-interest editing is strongly discouraged, and paid editing must be disclosed.Wikipedia's conflict of interest guideline. Brands that edit their own articles get reverted, tagged, and sometimes publicly covered for it. Wikidata is more open, since its claims need verifiable references rather than notability, and for most companies it is the realistic first entry point. The independent coverage that eventually earns a Wikipedia article gets built through Digital PR, News, and the Gatekeeper Publishers.
Disambiguation: shared names and wrong domains
The expensive entity failure is the wrong-domain failure: a model recommends your brand and attributes it to a website that is not yours. It happens when your name is shared with another company, when your name is a common word, or when a stronger domain sits near your name in training data. The recommendation you worked for then sends buyers somewhere else. OppAlerts' testing found a real case where most of a model's memory-only mentions of a brand pointed at an unrelated company's domain, and turning search on corrected it.Documented in the domain-attribution analysis within the AI Search Visibility research; the share of wrong-domain mentions for the affected brand: [number: verify].
Audit it directly. With search off, ask each model you care about: what is brand, what does it sell, who owns it, and what is its official website. Run each question several times, because single runs mislead; Measuring AI Search Visibility covers the sampling. Log every wrong fact and wrong domain. The fix is the checklist above applied against the specific confusion: your domain stated on every profile, sameAs markup in place, and coverage that mentions your brand and your domain together, so the next training run and every live retrieval sees the association you need it to learn.