Semantic Gravity Gradient

An engine that reads meaning places text as points in a space where near points mean near things. Where a firm’s terms are written one way, page after page, those points gather, and the place where they gather pulls: text that uses the same terms in the same sense lands near it. The semantic gravity gradient is the slope of that pull, how steeply it rises towards the centre and how far out it still reaches. The terms written one way set where the centre is. The terms around them, the shorter runs of a phrase, its neighbours in meaning and the vocabulary of its field, shape the slope and decide which shorter searches it runs towards. The studio reads the slope twice. In a model’s space it is a descent, a well that tokens fall towards. In a results page it is the mirror of that well, an ascent: the nearer a search sits to the centre, the higher the firm is expected to stand on it. The same picture holds at three sizes: the uses of one term, the pages of one site, and everything published under one name, by the firm and by others about it.

Where the Term Comes From

Noo–Studio coined the term on 12 October 2026. That day no earlier use of the exact phrase was found on the web, in Google’s search suggestions, or in the arXiv and OpenAlex indexes of research. Its parts are older, and so is the picture.

In physics, gravity gradiometry studies the Earth’s field through “the spatial gradient of gravitational acceleration” (Wikipedia). In 2016 Sándor Darányi and four co-authors used “gravitation as a metaphor to express changes in the semantic content of a vector field”, with a term’s PageRank as its mass (arXiv). In 2025 Maryna Vyshnyvetska described a query inside a language model as a mass “creating gravitational potential wells” that draw tokens during generation (arXiv).

The term takes that picture to what a firm publishes, and joins the well in a model’s space to its mirror in the results. It runs through the three terms the studio uses as one set, each of which measures it at its own size: semantic token optimisation for a term, the semantic authority mesh for a site, and semantic entanglement optimisation for everything under one name. The three papers of those names, in preparation, set out the theory and its tests.

Written short it is SGG, letters that computer vision already uses for scene graph generation, so the full name comes first.

Why It Matters

A pull says that a firm stands somewhere. Its slope says what follows: which searches the firm can expect to be shown for, and in what order it can win them. A page on a phrase of four words stands at the centre for those four words.

It does not thereby stand for the first of them, which lies far out, where many larger wells overlap. An old site with a deep well is hard to move from its centre, and a new one begins where no such well stands. Part of this is written down.

Google’s internal documentation of March 2024 stored, for each site, a siteFocusScore, “how much a site is focused on one topic”, and a siteRadius, “how far page_embeddings deviate from the site_embedding” (the documentation): a centre, and a distance from it. The documents list what was kept in 2024, never what it weighs, and nothing in them names a slope. That step is the studio’s theory: standing falls off with distance from a centre along a curve, and the curve can be read from positions.

Authority, in this picture, is where two slopes agree, the one a firm’s own text makes and the one made by what others write about it, and links are how an engine finds that the two belong together.

What It Is Not

It is not a ranking factor, and it is not a score. No engine publishes a number by this name and the studio offers none; a score that a tool computes from links estimates something else. It borrows a picture from physics and none of its mathematics.

It is not the semantic gravity of Legitimation Code Theory, a framework of education research in which the words mean “the degree to which meaning relates to context” (Journal of Chemical Education, 2020). Nor does it describe how a model is trained: nothing here is trained, and the slope describes where published text stands. And it is not answer gravity: answer gravity names the pull as it shows in AI answers, and the gradient is the slope of that pull, wherever it is read.

How the Studio Uses It

This site is the first case. Each of the studio’s terms is written one way and defined on one page at /term/, and the pieces around it are planned for the words they add near that term, not for a count of repetitions. A new satellite starts on the exact searches of its subject, where few older sites stand, and is widened from there.

The three papers measure the same slope at three sizes, each with its own deviation: a term’s forms and senses, a page’s distance from the centre of its site, and a property’s place among everything under one name.

How to Check It on Your Own Site

1. Take the phrase your firm most wants to be found for, three or four words long. Search it whole, then drop a word at a time down to the first. Note where you stand at each step.
2. Read the page that defines the phrase. Is the phrase written one way everywhere? Then list the other terms on the page: do they lead towards the shorter searches you noted, or somewhere else?
3. Search your firm’s name beside the phrase and read what others have written. Do they use your words for it, or their own?