Before you buy AI search, read what people couldn't find.

Start with the list nobody opens

Most stores have a report that nobody reads. It's the list of searches that came back empty. Someone typed "linen couch," hit enter, and saw "No results found." Maybe they tried once more. Maybe they left.

When a team decides its search "isn't good enough," the next step is usually a demo from an AI search vendor. The demo is impressive. It understands "something cozy for a small living room" and returns three lovely sofas. It's easy to sign.

Before you do, spend an afternoon with the failures you already have. Your store platform, your search provider, or your analytics tag can usually give you the searches that returned nothing, and often the ones that returned results nobody clicked. Export a month of each. A marketplace can do the same with buyer searches. This walk-through takes that export from a list of strange phrases to a short set of fixes, and then to a fair test for any tool you're thinking of buying.

You'll need the export, a spreadsheet, and someone who knows the catalog well. That's it.

Sort every failed search by why it failed

Read the list top to bottom, and give each search one of four labels. Don't fix anything yet. Just sort.

Their words, not ours. The customer said "couch" and your catalog says "sofa." They typed "sneakers" and you sell "trainers." They searched a color you call "oat" by the name everyone else uses. The product is there. The vocabulary isn't.

It's there, but the detail isn't. Someone searched "waterproof jacket" and you stock three, but the word lives only in a photo caption, not in any field search can read. A buyer on a marketplace looked for "same-day pickup" and a dozen suppliers offer it, but nobody captured it when they listed. This pile is about data, not search.

We don't sell it. Some searches are for things you've never carried. That isn't a search problem at all. It's a merchandising signal, and a useful one. If the same request shows up every week, someone should know.

Typos, codes and formats. Misspellings, part numbers with and without dashes, "10.5" against "10 1/2," a model number people copy from the box. Most search tools handle some of this already. The rest is a short list of rules.

Two things usually happen when a team does this. First, the piles aren't the size anyone expected. Second, the "their words" and "detail isn't there" piles are often bigger than the pile that needs anything clever.

Then open the second export: searches that found results but led nowhere. These tell you where search returned something, just not the right thing. Look at what came back for each. You'll often see the same pattern, such as accessories outranking the product, or out-of-stock items crowding the top.

Fix the cheap piles first

Each pile has an owner and a fix that doesn't require a new search engine.

Synonyms belong to merchandising. Keep a short, owned list that maps customer words to catalog words. Add to it every week from the failed searches. Most search tools accept a synonym list. The work is deciding, not configuring.

Missing detail belongs to the catalog. If people search for an attribute, it should be a field: material, fit, fabric weight, pickup option, whatever your customers keep asking about. On a marketplace, that means the supplier listing form asks for it, and listings without it are easy to spot. This is the same discipline as checking that a listing's claims are true, and it pays twice, because filters and product pages get better too.

"We don't sell it" gets a page and a report. A helpful empty result, with close alternatives or a way to ask, is better than a dead end. And a simple monthly note to the buying team about recurring requests costs almost nothing.

Formats get a few rules. Normalize part numbers. Accept common size notations. Test them against the actual searches in your export.

Do these, then rerun the export a few weeks later. Start with whichever pile holds the most searches, not the most interesting one. A dozen synonyms for your most-searched words will usually do more than a clever fix for a rare query. Whatever is still failing is the real case for something smarter.

Make any new tool search your failures

If you still want AI or semantic search, and for some catalogs it is genuinely worth it, now you have a test set nobody else has. Hand the vendor a sample of your remaining failed searches, plus a handful that work well today, and ask to see results on your catalog, not theirs.

Watch for these, query by query:

• Plausible but wrong. A search for "waterproof" that returns water-resistant jackets looks like success. It's a promise you can't keep. Count these separately, because they cost more than an empty page.
• The things that worked before. A new engine that handles fuzzy requests but buries exact part-number searches can make your best customers' day worse.
• Out of stock and discontinued. Does it know, and does it rank them down?
• Your controls. Can merchandising pin, boost, or exclude results? Can someone explain why a result appeared?
• What it costs as traffic grows. Ask how pricing works per search or per item, and what happens at your busiest week.

None of this is against AI search. It's about buying it for the searches you have, not the ones in the demo.

Keep the list alive

The export isn't a one-time project. Give it an owner and twenty minutes a week. Read the new failures, add synonyms, send missing attributes to whoever owns the catalog, and pass "we don't sell it" requests along.

Over time, the list gets shorter and stranger, which is how you know it's working. When something new and odd climbs to the top, it's often the first sign of a trend, a supplier problem, or a product page that broke. Seasonal words show up here first too, weeks before anyone updates the navigation.

This sits squarely inside the work on our ecommerce page: product discovery that understands its catalog. For a marketplace, the same exercise tells you what to ask suppliers for during onboarding. If you'd like a second reader for your export, send us a month of empty searches and we'll sort them with you.

Related work: Paddle8.