From Search to Intent: The next generation of classifieds

Every classifieds platform in the world runs the same loop. A seller lists. A buyer searches. The platform ranks. Somebody contacts somebody.

Enormous engineering effort goes into the ranking step, and the returns on it flattened years ago. The gains available now are not in ranking the results of a query better. They are in the fact that the query was a lossy compression of the intent in the first place.

The query is not the requirement

Someone types honda city 2018.

What is the actual requirement? Almost certainly not “any 2018 Honda City”. It might be: a reliable family car, under a certain budget, from a seller who will not waste my Saturday, available within driving distance, ideally one owner and full service history. The user typed a model and a year because that is what the search box accepts.

The platform then treats the typed string as the requirement, ranks against it, and both sides lose information. The buyer sees listings that match the string but fail the requirement. The seller with a 2019 Honda City in immaculate condition at a better price never gets shown at all.

Every classifieds platform is sitting on this gap. It is not a ranking problem. Ranking is doing exactly what it was asked to do.

Where intent actually reveals itself

The signal is already in your logs, and most platforms do not use it.

Query reformulation is the most direct expression of intent available. When a user searches honda city 2018, then honda city 2019, then sedan under 8 lakh, they have just told you their requirement is a budget and a body type, and the model was a proxy. That third query is the real one. Most systems treat these as three unrelated sessions.

Rejection is more informative than engagement. A user who views forty listings and contacts none has communicated something precise. Platforms instrument clicks obsessively and near-misses almost never, which means the richest preference signal in the product is being thrown away.

Dwell patterns separate browsing from buying. Someone who returns to the same three listings across four days is close to a decision. Someone sweeping two hundred listings in ten minutes is calibrating on price. These deserve entirely different treatment and usually get the same feed.

Cross-category movement reveals the real job. A user moving between apartments-to-rent and flatshares is not confused. They have a budget and a location constraint and are exploring the solution space. The platform sees two unrelated category sessions.

The three shifts that follow

From matching strings to modelling requirements

The unit of the system stops being the query and becomes the requirement — a structured, persistent, evolving model of what this user is trying to accomplish, assembled from everything they have done rather than from what they last typed.

This is a genuine architectural change. It means holding state across sessions, updating on rejection as well as engagement, and being willing to show a result the literal query did not ask for. That last part is where most attempts fail, because the relevance metrics in place will initially score it as a regression.

From reactive search to standing intent

For high-consideration categories — property, vehicles, jobs — the purchase cycle runs for weeks. The platform is only present for the minutes the user is on it. A held requirement inverts that: the platform works the whole cycle and re-engages when something genuinely matches.

This changes the economics of the relationship. Return visits stop depending on the user remembering to come back.

From listing fields to listing understanding

Sellers write terrible structured data and reasonable prose. The condition, the history, the reason for selling, the flexibility on price — it is all in the description, and none of it is in the filters. Extracting it makes both sides of the marketplace legible in a way that fifteen more form fields never will, because sellers will not fill in fifteen more form fields.

The uncomfortable part

Intent-based discovery reduces measured search volume. Searches per session falls. Time-to-contact falls. If your reporting treats engagement as the goal rather than as a proxy for it, this will look like the product getting worse in exactly the quarter it got better.

Marketplaces that navigate this successfully change the metric before they change the product: successful transactions per user, and time-to-outcome. Marketplaces that do not will kill the work at the first review.

There is a second tension worth naming. Better intent matching means showing fewer, more relevant listings — which is directly at odds with a monetisation model built on paid prominence within a long result set. That is a real strategic conflict, not a technical one, and it gets resolved at board level or not at all.

Where to start

You do not need a platform rewrite. You need the signal.

Instrument rejection. Persist requirements across sessions. Extract attributes from listing text. Measure outcomes rather than activity. Then earn the right to change the interaction model.

The platforms that win the next decade of classifieds will not have a better search box. They will have stopped needing one.