How Bestpicky chooses

What the model weighs

Each dimension is scored from evidence records, then combined at the weights below. Weights are configuration, not code, so a category can be re-weighted without rewriting the engine — and doing so creates a new model version rather than silently changing old results.

Missing evidence lowers confidence, not the score

When a dimension has no usable source, it is removed from the calculation and the remaining weights are renormalised so they still sum to one. The confidence score falls in proportion to the weight that was dropped.

This matters because the alternative — treating an unmeasured dimension as a zero — would punish a product for our own gaps in coverage. A chair we know little about is reported as a chair we know little about.

A product with too little evidence to place against others is not ranked at all. It has not scored badly; we cannot yet judge it.

Ratings are adjusted for sample size

Where a permitted source supplies both an average rating and the number of ratings behind it, the raw average is never used directly. It is pulled towards the category mean in proportion to how thin the sample is:

adjusted = (v / (v + m)) × R  +  (m / (v + m)) × C

R = the product's average rating
v = how many ratings produced it
C = 4.1   (the category mean)
m = 250     (the confidence threshold)

So 4.9 stars from 21 ratings lands below 4.7 stars from 15,000. A small sample is evidence of very little.

Bestpicky currently operates without a permitted live rating source, so customer satisfaction is dropped from every score and its weight redistributed. You will see that reflected in the confidence figure on each product.

What cannot influence a ranking

Bestpicky earns a commission on some purchases made through links on this site. That relationship cannot reach the scoring code: affiliate, sponsorship and commission data are not inputs to the model, not tiebreaks, and not accessible from it. The boundary is enforced by the build — a change that imported commercial state into the scoring module would fail continuous integration rather than ship.

An editor may override an algorithmic placement, but not silently: an override cannot be saved without a recorded reason.

What we do not claim

No chair here will treat, cure or prevent a condition, and nothing on this site should be read as a medical claim. Where a chair provides stronger adjustable lumbar support than another, we say that and show the source.