Does describing products by feeling change what you find?

We describe every product two ways: how it feels, and the moments it suits. That costs money on every product, and it keeps costing as the catalog grows. So the question worth answering early is whether it finds anything a normal search would miss.

The test

It takes 645 items in our catalog and describes them twice, from different inputs.

One description is the merchant's own title and copy, embedded directly, with no tags in it. The other is only the tags our pipeline assigned after looking at the product photograph and that copy.

Getting that independence right is the whole design. An earlier version of this test compared two things that were both built from our own tags, so the result was arithmetic rather than a finding, and we threw it out. Even here the two sides are not fully separate, since the pipeline reads the same merchant copy.

Across the catalog they correlate at 0.41, which is weak. They are mostly looking at different things.

That shows up hardest in what each one returns. Ask both for the ten nearest products and they agree on 2.25 of the 10. For 101 of the 645, the two lists have nothing in common.

The nearest product by feeling also sits in a different product category 65% of the time, against 23% by text. Text search stays near the object.

Two cases make it concrete.

Take a ceramic French press. By text the nearest things are a ceramic coffee dripper, a ceramic mug, a ceramic coffee jar. By feeling they are a wooden cheese board, a wooden salad serving set, and a meditation singing bowl. Text returns the same object. Feeling returns different objects for the same kind of morning.

Now the other direction. A bamboo bath towel, sold as two variants under one title and one description. To text they are identical, at 1.00. By feeling they sit at 0.80, which makes them less alike than two products pulled at random from this catalog, where the average is 0.84.

The photographs say why. One is a charcoal herringbone with a white fringe. The other is sand-coloured plush terry. The tags follow: cool tones, monochrome, geometric and woven for the first; warm neutrals, soft, cosy and comfortable for the second. They agree on bamboo, breathable and quick-drying, which is the part the words already carried. What separates them lives in the picture.

Brands make that choice deliberately. A red chair photographed in an ornate room and the same chair in a bare white one are two different offers, and the brand decides which one to make. Reading the photograph reads that decision.

What it does not show

It does not show anyone prefers these results. It shows the two descriptions differ, and differ in the direction we expected. Whether that helps a person find something is a separate test we have not run.

Both descriptions also run through the same embedding model, so some of the agreement could come from the model rather than from the products. It is 645 items, one catalog, one version of our vocabulary.

What we changed because of it

Two things.

The feeling description belongs inside ranking, not bolted on beside it as a filter. Most of what it surfaces is missing from text matching's top results, so running it as an optional extra throws away the part doing the work.

Then we went looking for the obvious failure. If results cross categories that often, does a plain request stop working? Ask for a coffee item and the real coffee items still come back inside the top hundred, while the closest few cross into other objects. Close enough to keep the literal answer, far enough to add something.

The preview at okidokes.com runs on this. It is small, nothing on it is for sale yet, and links out to brands start at beta. What we still cannot tell you is whether any of it makes finding something easier. That is the next thing we are trying to measure.