Semantic Token Optimisation

Search engines and AI models read a page as tokens: numbered pieces of words, each placed in a space where nearby points mean nearby things. A firm that calls one service by three names, or uses one name in three senses, spreads itself across that space, and no point in it belongs to the firm. Semantic token optimisation (in US spelling, optimization) is the opposite discipline: one term, one meaning, one page that defines it, the same words in every piece and the same names in the structured data. The word token carries three senses here, and the practice works on all three: the named decision of a design system, set once and used everywhere; the model’s numbered piece of a word; and the unit an engine spends when it reads, so the page that answers in the fewest and surest tokens is the cheapest to read and the easiest to quote.

Where the Term Comes From

Noo–Studio coined the term on 10 October 2026. That day no earlier use of the exact phrase, in either spelling, was found on the web, in Google’s search suggestions, or in the arXiv, Zenodo and OpenAlex indexes of research. It is the first of three terms the studio uses as one set: semantic token optimisation for the word, the semantic authority mesh for the structure the words build, and semantic entanglement optimisation for the field they work in.

The semantic gravity gradient, coined on 12 October 2026, is the slope all three measure, and this term measures it at the size of a term. The paper of the same name, in preparation, sets out the theory and tests it with an open model.

Why It Matters

Google wrote this layer down. Its internal documentation, published by mistake in March 2024 and confirmed as genuine by Google that May, keeps salient terms for each page: one normalised term, “the different ways we found this normalized term in the signals”, and a salience between 0 and 1, “the importance of the term as a descriptor” (the documentation). For each document it also keeps “the body words over tokens ratios”, an original content score and a keyword stuffing score (PerDocData).

The documents list what was kept in 2024, never what it weighs. Since then the engines have moved further toward meaning. Google’s AI Overviews and AI Mode may use a “query fan-out”, “issuing multiple related searches across subtopics and data sources” (Google Search Central), so one question becomes many, and each of them meets a site’s terms on its own.

A term used one way, on a page that defines it, is the one every one of those searches can place.

What It Is Not

Keyword density counts how often a word appears on a page. Semantic token optimisation counts how many meanings a word carries across a whole site, and keeps that number at one. The same documentation scores the first practice, at its extreme, as keyword stuffing.

Nor is it the token optimisation of AI engineering, which trims what a model spends on a prompt: this works on the site’s side, though a page that says a thing once, in its own defined words, also costs a model less to read.

How the Studio Uses It

Every term this site uses has one page at /term/ that defines it in its first sentence. The first time a paper uses a term, the word links to that page, every term page lists the papers that use it, and the structured data declares each term in the studio’s own set, so people and machines read the same definition. The studio’s satellites are written with the same defined words, so a term means one thing wherever the studio’s name appears.

How to Check It on Your Own Site

1. List the ten words your buyers search for. Count how many names your own pages use for each.
2. Paste a service page into a public tokenizer and see how your product and service names split. A name that breaks into many pieces is one a model has rarely seen, and one you have to define.
3. Ask an AI answer what your main term means. Does it use your words, and does it name you?