A correction to the slogan

Things, and strings.

“Things, not strings” was accurate marketing for the index shift, but as a description of the mechanics it overstates the break. The Knowledge Graph did not retire strings; it gave them a destination. Entity resolution is the string layer: a business name, an address, a phone number, a domain, a profile identifier are all strings, and the machine’s job is to decide which thing they point at. Every identifier is simply a string that has been promoted to a name.

This matters because strings are the only evidence a machine ever receives about a thing. When one surface says “Anytown Ford” and another says “Ford of Anytown,” the underlying thing is identical — but the strings disagree, and string disagreement is what the resolver actually measures. Confidence in the entity is, operationally, a function of agreement among the strings that reference it.

The language-model era made this more true, not less. A large language model is the ultimate string machine: token statistics with no native store of things at all. Retrieval grounding is the industry’s acknowledgment that string machines must be handed well-formed things at answer time. The history is a pendulum — strings only (keyword search), then “things, not strings” (the graph), then things for string machines (retrieval-grounded answers) — and the stable position was always both at once.

The entity is the destination; the strings are the roads. If the roads disagree, nobody arrives.

Everything that follows on this page is engineering for that pair: model the thing once, then make every string that references it agree.

The mechanics

Why consistency is a measurable signal.

Entity resolution — the process of deciding whether two observations describe the same real-world thing — is a well-studied problem in information systems. Its core mechanic is corroboration: when independent sources report the same value for the same attribute, confidence in both the value and the entity rises. When they conflict, confidence falls, and systems either pick a source they trust more or decline to state the fact at all.

This is why fragmentation has a real cost. Conflicting hours between a website and a business profile, a department page that describes a different legal entity than the profile it is linked from, or markup that disagrees with the visible page are not cosmetic issues — they are contradictory claims about the same node, and they lower the probability that any system will state your data as fact.

The engineering consequence is a design rule we apply everywhere: establish the truth first, then render it. Every attribute must have exactly one canonical source, and every surface that publishes the attribute — page, markup, profile — must derive from that source rather than maintaining its own copy. Copies drift; derivations don’t.