Your product on SleekNova? 99% don't make it. Try anyway

August 21, 2026

We Published Warranties Nobody Ever Promised

Five product pages published a warranty promise of up to 320 years in their structured data, taken from badges that name the age of the workshop rather than any warranty.

Some of our product pages carry short trust badges under the title. They are written by hand, in whatever words fit the product: where it was made, whether it is handmade, how old the workshop is. A reader sees exactly the words we typed.

Machines saw something else. Behind those badges, our code read each one and turned some of them into structured data, the machine readable summary that search engines and AI assistants take from a page. One of the things it produced was a warranty.

What we told machines

Five products were published with a warranty promise we had never been given, and one of them was a warranty of 320 years.

The others said 172 years, 145 years, 116 years and 60 years. Every one of those numbers was lifted out of a badge naming the age of the workshop that makes the product. Shoyeido has been blending incense in Kyoto for 320 years. That is a fact about the company, and we turned it into a promise about the box.

The pages themselves never said it. The badge on the Shoyeido assortment reads exactly what it always read, and the same is true of the Oigen incense holder, the Iittala plant pot, the Syouryu suzugami and the Tsuchiya Kaban randoseru. You could not have caught this by reading our site. It was only ever visible to something reading the page the way a machine does.

Why the code thought it was a warranty

It asked the wrong question. It looked for a number sitting next to the word year, and treated every hit as a term of cover. It never asked whether the badge said warranty or guarantee at all.

That question returns an answer on any badge with a date in it. An anniversary, a founding year, a workshop age: each one produces a confident, well formed, completely invented promise, and nothing downstream could tell the difference.

What we changed

We deleted the guess instead of improving it. A warranty is now published only where the badge itself promises one, in the word warranty or the word guarantee. Five products publish a warranty today, and every one of them says so on the page.

We considered the obvious repair, which would have been to demand the word warranty somewhere near the number. We turned it down. A cleverer reading of our own prose is exactly what produced this, and the fix would have been the same bet placed again. If a fixed term warranty is ever worth publishing, it gets a field of its own, filled in on purpose by a person.

The same block was guessing about origin too

The badge that says where a product is made was matched against a list of thirteen country names. Anything the list did not hold produced nothing at all, silently. Sixteen products said plainly where they were made and we published none of it.

Seven names have been added, taken from what our badges actually say rather than from what we assumed they said: Denmark, Austria, Israel, Sweden, Finland, Taiwan and Thailand. Nine more products now publish their origin.

Six still publish nothing, and that is deliberate. Their badges name a city, not a country. Paris is in France almost always, and almost always is not a standard we are willing to meet on a claim a machine will repeat without checking.

One product taught us something we had not asked. The Traveler’s Notebook is made in Thailand and Japan, and its badge says so. We were publishing Japan on its own, not because we had decided Japan was the answer but because Japan came first in our list. It now publishes neither. A field that holds one country cannot answer a question with two true answers, and picking one would have been our choice quietly wearing the maker’s voice.

The rule we are keeping

This is the second time this month we have found a published claim that was pattern matched out of our own writing rather than read from something we stored. The first was the page that counts our rejections, which had been re-reading its own sentences to work out why we had turned each product down. Both were fluent, both were confident, and both were wrong in a way that no page, no test and no reader could see.

A claim about a product is read from a field somebody filled in deliberately, or it is not published. A sentence that has to be interpreted is not a source.

Nobody knew the third instance was already live while this correction was being written. An award property on 78 product pages was switched on by any of fourteen words in the same kind of hand written line, and it is deleted rather than narrowed, for the reason above.

This correction was approved byMark van OosterwijckFounder and editor, SleekNova Labs