Marqo and the Search Bar That Forgot to Sell

Imagine asking a shop for a “quiet luxury wedding guest dress.” The shop has dresses. It may even have the right dress. But its catalog says “midi,” “satin” and “sage”; the customer says “quiet luxury.” A conventional search box can return a poor guess, or nothing at all. The shopper leaves. Somewhere in a spreadsheet, the retailer records a visit that failed to become a sale. The culprit looks like a small rectangle at the top of the page.

Marqo’s bet is that this gap between a person’s language and a merchant’s data is commercially interesting. The San Francisco company trains AI models around each retailer’s own catalog and shopper behavior. Its system can search words and pictures, rank category pages, suggest related products and supply a conversational shopping assistant. It is enterprise software, sold to the people responsible for conversion, merchandising and the systems underneath them.

01 / The first customer was a developer

Tom Hamer and Jesse Clark, both former Amazon engineers, founded Marqo in 2022. The early pitch was broad: make multimodal vector search easier to deploy. A developer could index unstructured material - text, images and other data - through an open-source engine and use Marqo Cloud to avoid running the infrastructure. In 2023 the company raised a $5.2 million seed round led by Blackbird Ventures. Lightspeed led a $12.5 million Series A in February 2024, bringing announced funding to $17.8 million.

The public pitch has since narrowed. Marqo’s current site leads with retail revenue, product discovery and store-specific models. That change tells a familiar business story: a general tool finds a sharper use case. Vector search is a capability; helping a shopper find a product is an outcome a retailer can test. Marqo still has its open-source repository, APIs and developer documentation. But the buyer it now courts is a retail organization with a catalog, traffic and a conversion target.

02 / The vocabulary problem

Keyword search is excellent when the shopper knows the merchant’s exact words. It becomes awkward when the words are descriptive, visual or incomplete. A sneaker fan might search by product code and colorway; a jewelry customer by metal, stone and occasion; a fashion shopper by a photograph. Marqo uses multimodal embeddings to put text and images into a common search space, then combines that product understanding with observed behavior. The store’s clicks, add-to-cart events and purchases teach the model which matches matter in practice.

This also addresses a dull but expensive failure: the new product with no behavioral history. A click-driven ranking system knows the bestseller and knows very little about an item added this morning. If the model can read its image and description, the new item has a chance to appear before it has earned a click. For retailers with frequent launches or immense catalogs, that is more than a technical nicety. It is a way to avoid making the freshest inventory invisible.

Marqo does not ask merchandisers to surrender the shop floor. Its tools allow products to be pinned, boosted, buried or excluded, and business rules can sit over learned relevance. That matters because “relevant” is not always “the item this business should show first.” A retailer may have stock to clear, a brand promise to honor or a margin goal to meet. The company’s differentiation is the attempt to train product understanding at the model layer, then give merchants a steering wheel.

“Search the way you think.”Lightspeed Venture Partners on its Marqo investment

03 / The figures have names

AI search vendors all promise relevance. Marqo’s more persuasive material is its named retail testing. Mejuri gave Marqo half of its on-site test traffic and reported a 14.72% increase in purchase conversion and a 19.84% increase in search revenue per user. At KICKS CREW, whose sneaker inventory runs to hundreds of thousands of styles, the reported gain was 17.7% in search and discovery conversion. SwimOutlet ran a controlled test and reported a 10.6% increase in search add-to-cart rate; it later expanded Marqo from search and browse into recommendations.

Customer-reported results use different measures and periods; they should not be treated as one pooled benchmark.

These examples also show why a single “better search” metric is too blunt. Mejuri needs to read layered requests about metal, style and gifts. KICKS CREW must distinguish a particular model, code, size and colorway while stock shifts across regions. SwimOutlet has performance and leisure products whose shoppers do not speak one vocabulary. Marqo’s case is that each store deserves its own model because each store’s product language is peculiar.

The largest figure in Marqo’s published material is Kogan’s $10.1 million in incremental revenue attributed to improved search. Kogan’s enormous multi-category catalog is exactly the kind of place where a long tail of rare queries can add up. Such numbers are claims about particular deployments, not promises for the next shop. Their useful lesson is methodological: put a proposed search system against the current one on live traffic, choose a commercial measure in advance, and watch the whole journey rather than the prettiness of the first result.

04 / A search engine wants the whole storefront

Marqo now presents its product as one intelligence layer across search, collection pages, recommendations and conversational discovery. A pixel gathers behavior. MarqTune fine-tunes models on a retailer’s data. Merchandising controls preserve human direction. Sibbi, introduced in 2026, adds a conversational interface that can take a shopper from a vague request to products grounded in the retailer’s inventory. Integrations include Shopify, Adobe Commerce and Salesforce Commerce Cloud, alongside an API.

This puts Marqo against specialized search vendors such as Algolia, Constructor, Coveo and Bloomreach, and against retailers’ own mixtures of Solr or Elasticsearch, manual synonyms, recommendation widgets and chat software. The comparison is not settled by the word “AI.” A retailer has to ask whether a dedicated model improves actual sales enough to justify integration, training and ongoing spend. Marqo’s managed, enterprise-led sales approach makes sense for merchants with enough traffic to run credible tests and enough catalog complexity for relevance to hurt.

A smaller shop with a short, tidy catalog may get little return from custom model training. A retailer with thin traffic cannot quickly prove a conversion lift, however good the results look. Stale inventory or missing product images can weaken the very signals Marqo depends on. The practical move is to start with one search or browse surface, record a baseline, split real traffic and agree on the measure before anyone sees the result. The machinery is elaborate; the test should be plain.

There is a small irony here. The company began by making a complicated search technology easier for developers. Its newer work makes a far more ordinary human action - asking a shop for something - easier for shoppers. The search bar was never merely a way to retrieve records. It was the moment a customer tried to describe a desire in the store’s language. Marqo is betting the store can finally learn to answer in the customer’s.