Constructor
built the search bar.

It is a Tuesday afternoon at a Sephora distribution center, and a customer in Lyon is typing the word "matte" into a search bar. By the time her finger leaves the "e", the page already knows she has bought two foundations in the last six months, scrolled past three primers, and never once clicked on a setting spray. The results she sees feel personal. They are not. They are the output of a model trained on roughly 322 billion shopping interactions a year, running out of a building on Bush Street in San Francisco.

The company behind that search bar is Constructor. It does not have a logo most shoppers would recognize. Its founders do not give TED talks. Its product is, on the surface, the least glamorous thing in ecommerce: site search. And it is currently one of the more interesting companies in retail technology, mostly because it figured out, before anyone else cared to admit it, that the search box was where retail was going to be won or lost.

A bar, a query, and a lot of money on the table.

For most of the last twenty years, ecommerce search was treated as plumbing. Type in a word, get a list of products that contain that word. If the catalog was big enough and the query specific enough, this worked fine. If not, the shopper bounced - and the retailer blamed marketing. Algolia made the plumbing faster. Elasticsearch made it cheaper. Nobody made it smarter in a way that actually moved revenue.

Eli Finkelshteyn, who had spent years building data infrastructure at Tumblr and Shutterstock, noticed something strange in the clickstream data. The shoppers who searched were the ones most likely to buy. They were also the ones most often handed irrelevant results. The gap between "the shopper knows what they want" and "the site shows them what they want" was, by his rough estimate, costing enterprise retail tens of billions a year. So in 2015 he and Dan McCormick, the former CTO of Shutterstock, started Constructor.

Price the result, not the query.

The bet was small in concept and enormous in implementation. Most search vendors charged per query. Constructor proposed to charge for outcomes. If the search results did not move revenue per visitor, the customer did not pay for it. This was, depending on your view, either a generous business model or a contractual statement that the founders believed their math.

Their math, it turned out, was good. Constructor's models did not just rank products by string match. They trained on clickstream - what a shopper actually clicked, lingered on, added to cart, and ultimately bought. Then the system ranked future results against the KPIs the merchandiser cared about: conversion rate, average order value, revenue per visitor. The team called this approach "AI glassbox." It was a deliberate poke at the black-box vendors then dominating the conversation. Every ranking decision had to be explainable to a merchant. Otherwise, the merchant would not trust it, and the model would lose.

Nine pieces that all look like one thing.

From the outside, Constructor is a search box. From the inside, it is nine products that pretend to be one. There is Search and Autosuggest, the parts shoppers actually touch. There is Browse and Collections, which builds personalized category pages in real time. There are Recommendations, which run across email and product pages and homepage carousels. There are Quizzes, which translate "what kind of moisturizer" into "this moisturizer, for you, today." There is Attribute Enrichment, which uses AI to tag the messy catalog data enterprise retailers usually pretend they have already cleaned up.

Then there are the agents - the part Constructor has been quietly shipping while everyone else announced theirs. The AI Shopping Agent went live in the AWS Marketplace AI Agents and Tools category in mid-2025. The AI Product Insights Agent, known internally as PIA, lives on product detail pages and answers the kind of questions that used to require a chat window and a patient human. The Merchant Intelligence Agent, MIA, launched in March 2026 and turns merchandiser questions like "why did dresses drop 12% last week" into actual answers, in actual sentences, in roughly the time it takes to make coffee.

Numbers, because numbers are the point.

If you charge for outcomes, your case studies had better look like case studies. Constructor's do. The platform has generated documented revenue lifts north of $35 million for individual retailers. It has been credited with 47% increases in revenue per visitor and ROI multiples north of 21X. In its most recent fiscal year, customer count grew 82% and gross revenue retention sat at 96% - which is the polite way of saying that almost nobody who signs up ever leaves.

Make the internet feel like a great shop.

Finkelshteyn talks about a specific kind of retail experience - the one where you walk into a store, the clerk recognizes you, and within ninety seconds you have the thing you wanted but could not name. That is, in his telling, what ecommerce search should feel like. The fact that it usually does not is a failure of imagination on the industry's part, and a business opportunity on Constructor's.

There is something quietly serious about the company's approach to AI. While other vendors race to bolt generative chatbots onto everything that holds still, Constructor has been mostly focused on the part that pays: ranking, retrieval, and the unglamorous work of explaining why a specific product showed up first. The phrase "AI glassbox" appears more often in their marketing than the word "magic." This is, in 2026, an unfashionable position. It also happens to be the one enterprise buyers find easier to sign.

Shoppers will stop typing. Then what?

The interesting question is not whether Constructor will keep growing - they will, for as long as enterprise retail keeps caring about conversion rates. The interesting question is what happens when shoppers stop typing into search bars at all. Voice, agents, multimodal queries, image-based search - the inputs are about to fragment. The companies that win the next decade in retail are the ones whose ranking systems do not care what input they get, only what outcome they need to produce.

Constructor has been preparing for this for years. The same models that rank a typed query rank an agent-generated one. The same explainability that comforts a merchandiser today will comfort a regulator tomorrow. The same KPI-tied pricing that won them their first dozen enterprise contracts is going to look like a very good deal when the rest of the AI economy starts being asked to prove its return on investment.

Back in Lyon, the shopper has added two products to her cart and clicked through to checkout. She will never know what ranked the results. She will probably tell a friend that Sephora's site "just gets her." Somewhere on Bush Street, a model registers the click, updates the weights by a fraction, and gets very slightly better at its job. The search bar moves another small amount of money. Constructor takes its cut. The shopper, the merchant, and the model all walk away happier than they were a minute ago. None of them are typing the word "matte" anymore.