This lesson covers the worst result the Knowledge Graph can give you: a confident match to the wrong thing. It explains why that is more damaging than no match at all, shows the demo site’s own brand caught in it twice, and gives a method for finding and triaging it across any corpus.
What this lesson covers
- What over-linking is and how it differs from being unlinked
- The mechanisms that produce it
- A real brand resolved to two wrong identifiers
- Detecting over-linking by clustering on MID
- Separating findings that matter from those that do not
- Rename, disambiguate, corroborate or leave alone
Unlinked means machines do not know you; linked means they do; over-linked means they think you are something else, and every later system builds on that mistake. It happens when a name overlaps a common noun or a famous entity, the surrounding text gives too little disambiguating signal, and the famous entity is a much stronger prior. On the demo site, the brand is attached to two identifiers: one for the generic concept of an academy — whose detailed description is an encyclopaedia entry about Plato’s school in ancient Athens — and one that another Google API cannot resolve at all. Neither is the company.
The detection method is mechanical: cluster mentions by MID, then read the wording, because one identifier covering several unrelated surface forms is the signal. Run across the corpus, the same check finds everyday words resolved to famous nodes — “models” to a 2008 comedy film, “teams” to Microsoft Teams, “reviews” to an academic publisher. The lesson is clear that most of these do not matter: over-linking matters on the entities you are trying to own. For those, the options run from renaming, the cheapest, through adding type words in context, to off-site corroboration, the slowest — and for generic nouns, doing nothing is usually right.
This lesson continues Module 3, Knowledge Graph Auditing & Entity Mapping, in The AI-powered Entity Audit, and leads into off-site corroboration. It suits SEOs auditing brands with generic or overlapping names, and anyone who needs to explain a wrong knowledge-graph match to a client calmly.