What AI product search has to get right.

Most on-site search is still lexical. It matches the words a shopper typed against the words in your product titles, and ranks by how often they overlap. That works when the shopper already knows your vocabulary. It fails the moment they describe a problem instead of a product — which is what people do when they are not experts in your category.

The interesting question is not whether a language model can hold a friendlier conversation than a search box. It plainly can. The question is whether it returns things you actually sell, in stock, at the price you are charging today. Everything below is about that.

Retrieval is the product, not generation

The useful mental model is that the model never invents an answer about your catalogue. It retrieves, then explains what it retrieved.

Semantic search works by turning each product into a vector — a list of numbers positioned so that things with similar meaning sit near each other. “Tap washer assembly” and “stop a dripping tap” land close together even though they share no words. At query time the shopper's sentence is embedded the same way, and the system returns the nearest products. Only then does the model write prose, and it writes it strictly about the items it was handed.

That ordering matters enormously. A system that generates first and checks later will eventually describe a product that does not exist, in fluent and completely convincing language. A system that retrieves first can only ever talk about real rows in your catalogue. It can still be unhelpful — but it cannot be fictional.

Three input modes, one retrieval problem

Text, voice and image look like three separate features in a product brochure. They are the same pipeline with different front doors.

Photo search is the mode that most often surprises retailers. A shopper holding a broken component rarely knows its name, and would otherwise have abandoned the session or phoned during business hours.

Because all three converge on the same retrieval step, catalogue quality — not model choice — sets the ceiling. Thin product titles, missing attributes and inconsistent category names hurt a semantic system just as much as a lexical one. They simply fail more politely.

A correct answer about yesterday's catalogue is a wrong answer

This is the operational trap. An assistant that recommends confidently from a stale index will cheerfully sell an out-of-stock item at last month's price. The shopper does not experience that as a synchronisation bug; they experience it as being lied to, and the cost lands in your support queue.

In practice that means the assistant should treat its own index as a way to find candidates, then verify each candidate against live product data before it speaks. It is slower by a few hundred milliseconds. It is the difference between a tool that deflects support tickets and one that creates them.

Scale changes which problems you have

At a few hundred products, almost any approach works and merchandising judgement matters more than retrieval quality. At tens of thousands, the failure modes invert: near-duplicate variants crowd out genuine alternatives, and the assistant needs to understand that six listings differing only by cable length are one decision, not six.

PandoraBot's own deployment with ergo-tel.gr, a Greek electronics retailer, runs across a catalogue the company reports at over 50,000 SKUs, handling a reported 1,600+ conversations a month at a typical two-second reply. The retailer reports a 31% increase in sales following deployment. Those are customer-reported figures from a single retailer rather than a controlled trial, and we present them as such — but the catalogue size is the part worth noting, because it is where naive approaches stop working.

Most of the value arrives outside business hours

The clearest argument for an always-on assistant is not that it is cheaper than a person. It is that the demand it serves was previously going unanswered entirely.

Across smartbnb.gr, a Greek property management operator running SalesBot as a 24/7 guest concierge over 500+ properties, the company reports that 73% of chats arrive outside staffed hours, alongside a reported 44% increase in conversions. Again: one operator's numbers, self-reported. But the shape of the finding is consistent with what any business with international or leisure customers will recognise — the enquiry arrives when the shopper is free, not when you are open.

Knowing when to stop

A good assistant has an explicit boundary. Complaints, warranty disputes, anything touching a refund, and any case where retrieval confidence is low should hand off to a person with the full conversation attached, rather than improvising.

The handoff is worth designing deliberately, because a bad one is worse than no assistant at all: it makes the customer repeat themselves, which is precisely the experience they were promised they would avoid.

If you cannot attribute it, you cannot justify it

The last piece is measurement, and it is the one most often skipped. A conversational assistant that cannot be tied to revenue will be cut in the first budget review, regardless of how well it performs.

The minimum viable instrumentation is: tag the conversation, carry that tag onto the checkout link, and land it in analytics alongside your other channels. That lets you ask the only question that matters — what did assisted sessions convert at, against unassisted ones? Everything else is a vanity metric.