The fit score

Each product the assistant shows can carry a fit score from 0 to 100 and a label: Strong fit, Good fit, Partial fit or Weak fit. This page says what that number is, and what it is not.

What it is

The score says how well a product’s own facts match what the shopper asked about. It is worked out by a fixed calculation, the same way every time, from three things:

• The match to the question. Does the product contain the compound the shopper asked about, and is it a main part of the product?
• The certificate. Is there a lab certificate on file for the product, and how far has it been checked? A product with no current certificate scores lower.
• The evidence. For the compound and the outcome in the question, how strong is the published research the assistant can cite?

Each card also shows one short sentence naming what its score rests on. That sentence uses fixed wording. It is never written by the model.

If a fact is missing, the score goes down. A missing fact is never filled in with a guess.

What it is not

• It is not a rating of the product’s quality, and it is not a ranking of brands.
• It is not a promise that the product will work, and it is not advice for any one person’s health.
• It is not a recommendation to buy, and it does not take a person’s medications or conditions into account. Questions about those are handled by the assistant’s separate safety checks.
• It does not come from the model. The model never sees the number, is not asked to write it, and is told never to mention it.

Where it appears

The score appears on product cards when the assistant is matching products to a question. It does not appear on answers that only describe a product’s contents, on education answers, on help answers, or on refusals.

When several cards are shown, the strongest fit comes first. Cards with the same score keep the order the product search returned. A product that could not be scored is shown last, with no score.

What you can do to raise it

The inputs are your own catalog facts, so the useful steps are factual ones: keep the lab certificate current, fill in the product’s compound and amount fields, and keep product names and categories accurate. The score changes when the facts change, never because of wording on the page.

For developers

The fit object on each product in a chat response, and the debug.fit_scores list a test key receives, are described in the Widget endpoints reference.