The Word Pick Was Enough To Publish An Award

Search engines and AI assistants do not read a product page the way you do. Alongside the page, every product publishes a block of structured data: a machine readable summary with named slots for price, rating, materials and so on. One of those slots is called award, and it means what it sounds like. This product won this.

On 78 of our 386 product pages we were filling that slot. We had never checked whether any of those products won anything.

What we were actually doing

Every product page carries a short trust line, written by hand, saying why we take the product seriously. Sometimes it names an award. More often it names a magazine that reviewed it, a shop that stocks it, a rating, or the year the workshop opened.

Our code read that line and looked for any of fourteen words: Award, Prize, Winner, Best of, Editor’s Choice, Pick, Selection, Honor, Medal, Recognition, Golden Pin, Red Dot, iF Design, Good Design. Where it found one, it copied the entire sentence into the award slot and published it.

So a machine asking what this product won got back a paragraph. On 72 of the 78 pages the answer we published ran past one sentence, with the award, where there was an award at all, sitting beside a star rating, a review count or a founding year.

On eight pages there was no award of any kind. They qualified on the word pick, or the word selection. An editor naming a product in a roundup is not an award, and we published it as one.

The same mistake, one slot over

A second slot, item condition, switched on wherever a product badge contained the word handmade. Condition is whether an item is new or used, and handmade says nothing about that either way. It was set on 45 records. The code also looked for the spelling “hand made”, which has never matched a single product we carry, and a branch that matches nothing looks exactly like a branch that works.

What we changed

Both slots are gone. Not narrowed, not given a stricter word list. Deleted.

A stricter list was the obvious repair and we refused it, because we tried that reasoning once already and it cost us. A pattern that reads a hand written sentence can only ever be tuned against the sentences that exist today, and those sentences get rewritten every time somebody improves one. An award is published from a field that holds an award, or it is not published. We do not have that field, so for now we make no claim.

Nothing on any product page looks different. The trust line sits where it always sat, saying what it always said, as a trust signal rather than a trophy.

How we found it

Not by reading the file. Two versions of this same mistake had already cost us: warranties of up to 320 years, parsed out of badges naming the age of a workshop, and a product page headed with a count of awards the product does not hold.

Three instances of one shape is not three problems. So rather than fix the third we built an instrument to find every one of them at once, and pointed it at every place our code inspects a written sentence and publishes a result from what it sees. It found 53 such places, and this was one nobody had ever counted.

The instrument got two of its own answers wrong on its first run, which is why every figure above was produced twice, the second time by a separate script that shares no code with the first.