E-commerce Search & Discovery

The storefront search worked the way most storefront search works. It took what the shopper typed, looked for those words in the product data, and ranked whatever contained them.

That is fine when someone types a product code. It falls apart the moment they describe what they want in their own words, which is what most people actually do.

Merchandising could see the problem in the numbers and could not do anything about it. Every ranking change was an engineering ticket, so the queue grew and seasonal adjustments landed weeks after the season had started.

Search is easy to under-rate because it looks like a solved problem. It is a box, it returns products, and it has been on every storefront for twenty years.

But a shopper using search has told you exactly what they want. They are further along than someone browsing a category, and they are closer to buying. Getting it wrong does not just lose that sale. It teaches the shopper that your search does not work, and they stop using it.

An online retailer's storefront search matched on keywords only. Shoppers typed full sentences describing what they wanted and got back results ranked by whichever product happened to contain one matching word. Search abandonment was high, and merchandising had no way to fix it without an engineering ticket.

The obvious fix is to match on meaning instead of words, and that alone does not get you there.

Search has to respect things that have nothing to do with language. A perfect match you cannot ship is a bad result. A product that is about to go out of stock should not sit at the top of the page. Margin, availability and what is actually selling all belong in the ranking, and none of that is in the query.

There is also the catalogue itself. Search can only work with the attributes that exist. When products come from different suppliers with inconsistent naming and missing fields, no amount of clever ranking on top will find something that was never described properly in the first place.

And search is not a page anyone will tolerate being slow. Whatever you add has to answer in the time a shopper will wait, which is not long.

We put a layer in front of the existing search rather than replacing the platform. That is faster to ship and much easier to undo, and replacing a working storefront is a decision that needs a better reason than this one.

The ranking was built against the real catalogue, real inventory and what shoppers actually did after searching, not against a clean sample. That distinction is most of the difference between a demo that impresses and a system that survives.

The other half of the work was giving merchandising the controls. If ranking rules and campaign boosts need an engineer, they will always be late. Handing that to the team who owns the outcome is what makes the improvement keep going after we leave.

Working against the real catalogue meant meeting it as it actually was, gaps included. Attributes missing on some products, the same thing named three ways by three suppliers, fields left empty years ago.

That is normally the least interesting part of a project like this and often the part that moves the number most. Ranking cannot surface something that was never properly described.

An AI-powered search and recommendation layer that understands shopper intent, not just tokens, ranking results against the actual catalogue, inventory and behavioural signals. Merchandising could tune ranking rules directly, without waiting on engineering.

The number was conversion from search, agreed before the build.

Conversion on its own can mislead, so it was not watched alone. Search abandonment says whether people are giving up. How far down the results page people click says whether the ranking is improving, and it usually moves first.

That early signal matters. It tells you the ranking is getting better before the revenue confirms it, which means you find out you are on the wrong track in days rather than at the end of a quarter.

Conversion from search rose 27%, search abandonment dropped sharply, and the merchandising team started shipping ranking changes themselves the same week they were requested.

Search understands what shoppers mean rather than which words they used.

The change that outlasts the number is who owns it. Merchandising ships ranking changes in the same week they are asked for, instead of filing a ticket and waiting. The system improves at the pace of the people who understand the catalogue, not the pace of the engineering backlog.

If storefront search is losing you sales, resist starting with the algorithm.

Look at your catalogue data first. In a lot of cases the reason search cannot find a product is that nothing in the record describes it the way a shopper would say it. No ranking model fixes a missing attribute, and enrichment is usually cheaper and faster than a search rebuild.

Second, check who is allowed to change ranking. If the answer is engineering, that is a constraint on how fast search can ever improve, no matter what sits underneath. The teams that get the most out of search are the ones where merchandising can adjust it directly and see what happened.

Third, be careful which number you chase. Conversion is the one that pays, and it moves late. Watch search abandonment and click depth as well, because they tell you whether you are heading in the right direction while there is still time to change course.

And you almost certainly do not need to replace your search platform. Adding a layer in front is faster, reversible, and enough for most of what is wrong.