We Were Filling Three Slots When We Should Have Been Answering a Question

At the foot of every lab note there is a short row of other products we have verified. It exists to answer one question: if you came here for this object, what else should you look at?

The rule that picked those products was not answering that question. It was answering a different one, and it was answering it correctly, which is why nobody noticed for a long time.

Here is how it worked. We look for products that share three things with the object on the page: its category, its feel, and its price range. If three such products exist, those are the three you see. They rarely do. So the rule widens: drop the price range and match on category and feel, or drop the feel and match on category and price range. Whichever of those two returns more products wins.

Price ranges are wide and feels are narrow, so the price version almost always returned more. A price band inside a category holds a lot of objects. A particular feel inside the same category holds a handful. The rule compared the two counts, saw a bigger number, and took it.

The effect on the page was that we would find two products that genuinely belonged beside the one you were reading about, discard both, and show you three that happened to cost about the same. Nothing looked wrong. Three cards rendered, the heading above them was accurate, and no reader could tell that a better answer had been found and thrown away.

The rule was optimising for filling three slots. It should have been optimising for relevance. That is the decision, and it is the whole of it.

From today, when matching on category and feel returns two or more products, those are the products you see, and the wider price match does not overrule them. The row is then two cards instead of three, and it is centred rather than leaving a gap where a third would have been.

Where the narrower match returns only one product, or none, the wider match still stands, exactly as before. That guard is deliberate. A single lonely card at the foot of a page is worse than three loosely related ones, and this change never produces one.

We measured what it moves before we made it, over every passed lab note with a linked, categorised product. The full three-way match reaches 54 of them. On 56 pages the old rule was discarding a two-product feel match to fill the third slot, and those 56 pages now show the two. The remaining pages are unchanged, including the ones where the narrower match found nothing at all.

One thing worth saying plainly, because it would be easy to present this as a bug fix and it is not. Both versions of the rule did what they were written to do. The comparison that picks the wider match is correct when the question is how many products can we show. It is wrong when the question is which products belong here. Nothing was broken; something was aimed at the wrong target, and we had written down the wrong target ourselves.