It measures how close the meaning of your writing sits to a category you picked. If you review a draft, use that comparison to investigate whether the writing fits the intended subject or audience.
An embedding model turns your text into a position. Every category already has a stored position from the same model. The percentage is their cosine similarity multiplied by 100.
What it does not measure
- Not accuracy. A confident, well-written, factually wrong passage matches its subject perfectly. Similarity has no opinion about truth.
- Not quality. A short focused piece scores closer than a longer, better one that covers several subjects.
- Not a ranking or citation signal. No search engine and no assistant is involved in producing the number, so it says nothing about how either would treat the page.
- Not keyword coverage. A high score does not mean particular terms appear, and adding terms does not reliably move it.
How to read it
Do not chase a high number. Look at which categories came back and in what order.
The report gives you the nearest categories and tells you whether yours is among them. If the category you picked is missing from the top, your writing is closer to other things than to the one you were aiming at. Go and act on that.
Each taxonomy returns ten. Read the first few. After that the list is giving you the closest remaining categories, and close is not the same as relevant.
A worked reading
You write a guide for practice managers choosing a dental supplier and check it against an audience category for dental practice management. Back comes dental equipment manufacturing as the nearest match, with practice management well down the list.
The guide is written about suppliers and not for buyers. Go and look at the structure: the opening explains how the equipment gets made, and the pricing section explains how manufacturers set list prices.
No term count would have told you. Both audiences use the same words.
One operating limit
The comparison takes 8,192 bytes and refuses anything longer instead of trimming it. Scoring the first part of a long article and calling it the whole thing would describe a page that opens well and wanders as though nothing were wrong.
Check long pieces section by section. Read the results separately. Averaging them puts back the problem the refusal exists to prevent.
At editorial review, compare the finding with the brief and read the text it concerns. Decide whether that passage needs revision to explain the intended subject.
See how the comparison works, what similarity measures, or Content Optimizer.