Where would you start if you were handed a retailer's search and told to improve it?

Start where you know you are wrong

Relevance work is unusual in having no ground truth. For most queries there is no correct ordering to compare against, only signals in tension — textual match, margin, availability, popularity — and a merchandiser with an opinion.

Zero-result queries are the exception. A customer typed exactly what they wanted, arrived with intent, and got nothing. There is no ambiguity about whether that is a failure.

So the first artefact is a table: distinct zero-result queries over the last month, ordered by volume, with the search volume next to each. The top hundred rows are usually a few hours of work and the highest-return hours available, because you are fixing known failures rather than adjudicating between plausible orderings.

The causes are mundane

Sorted roughly by how often they turn out to be the answer.

A synonym you do not have. The customer said "trainers", the catalogue says "running shoes". Or "hoover" against "vacuum cleaner", "capsicum" against "pepper". Regional vocabulary, brand-as-generic, and the retailer's internal terminology leaking into public-facing titles.

A spelling you do not correct. Long tails of one-character errors, and phonetic guesses at brand names people have heard and never read.

A filter still applied. The customer navigated into a category, then searched for something outside it, and the category filter silently constrained the results to nothing. This one is a user-interface bug that presents as a search failure, and it is common.

Attributes that are not indexed. Someone searched a capacity, a size or a model number that exists in the product data but was never added to the searchable fields.

You genuinely do not stock it. The residual, and the interesting one.

Only the last is a real absence. The first four are self-inflicted, which is why the exercise pays.

Separate "no results" from "no stock"

An important distinction that changes both the fix and the metric: a query matching a product that is out of stock is not a zero-result query, and treating it as one sends you to the wrong team.

If your index excludes out-of-stock items, a popular product going out of stock silently becomes a search failure. Usually the better behaviour is to keep it in the index, rank it lower, show it as unavailable and offer alternatives or a back-in-stock notification — because a customer who searched for it by name is already committed, and an empty page sends them elsewhere.

Reporting the two categories separately is what makes the distinction visible. "Queries with no match" is a search problem; "queries whose only matches are unavailable" is a merchandising and stock problem, and conflating them means neither team acts.

Do not show an empty page

Whatever the cause, the response matters. An empty results page is the end of a session.

The escalating fallbacks, in order of preference: correct the spelling and say so ("showing results for…", with a way back). Relax the query — drop the least important term and show what matches the rest, labelled honestly. Fall back to the closest category. Offer a way to be told when it arrives.

And record what the customer did next. A zero-result query followed by a successful reformulation tells you the synonym to add; one followed by a session ending tells you the volume you are losing.

Relevance beyond zero results

Once the known failures are fixed, you are into the part with no correct answer, and the honest position is that experimentation is the only arbiter.

The signals pull against each other. Pure textual match ignores whether anything sells. Popularity buries new products and entrenches whatever was already ranking. Margin optimisation quietly degrades the experience until it does not. Availability weighting helps conversion and hides products that are about to be restocked.

Two structural things are worth saying. Merchandiser control needs a home in the design rather than being fought: pinning a product to a term, boosting a brand during a campaign, burying a discontinued line. If the system has no sanctioned mechanism, someone will achieve it by editing product titles, which is far worse. Make it a rules layer applied on top of the ranking, visible and dated, so it can be reviewed and expired.

And conversion is the metric, not click-through. A ranking change that moves clicks to the top result while sales fall has made things worse; the two diverge routinely, and a team measuring the easier number will optimise the wrong thing.

What good looks like after a month

A concrete answer to "where would you start" ends with what you would have. A report of zero-result volume trending down; a synonym and spelling dictionary under version control with an owner; the category-filter bug fixed; out-of-stock handled deliberately rather than by exclusion; and an experiment framework so the next argument about ranking is settled by a test rather than by whoever is more senior.

None of that requires a new search engine, which is the answer candidates most often jump to and the one least likely to help.

Zero-result queries are the only part of search where you know the current behaviour is wrong. Fix the known failures before arguing about ranking, where nobody can prove anything.