About 60% of shoppers do not know what they want when a high-consideration journey begins (Kleio data). Ecommerce product discovery serves exactly those shoppers: it gathers the means by which a merchant site helps a buyer find the right product in its catalog. It rests on navigation, site search, the search engine, semantic search, filters and product data. This blog article compares those levers, measures where each one fails by query type, then presents Kleio's agentic discovery strategy, a solution built for complex buying experiences.
What is ecommerce product discovery?
A definition of ecommerce product discovery
Ecommerce product discovery covers everything that helps a shopper find, compare and choose a product on a merchant site. It spans site search, navigation, visual search, filters, the product page and the data that describes it. Effective product discovery reduces zero-result pages, shortens the path to purchase and feeds loyalty. A clear discovery strategy gives every item a chance to be found, through a consistent experience from one channel to the next.
Search or discovery: what is the difference?
Search answers a need already put into words: the shopper types a product name and expects relevant results. Discovery helps build that need. The shopper explores the catalog, compares products and sharpens the criteria. A search engine stays indispensable for the exact query. It falls short when the user does not know how to name what they want. Search and discovery therefore complement each other: a site that performs treats search and discovery as one experience, from the first click to the add to cart.
Why does product discovery weigh on conversion?
A shopper who does not find the right product leaves the site. Improving product discovery acts directly on the conversion rate, on average order value and on customer loyalty. The context sharpens the stake for the whole customer experience. French ecommerce reached 196.4 billion euros in revenue in 2025 (+7%) across 3.2 billion transactions. Average order value fell 3% to 62 euros (Fevad, February 2026). In a market growing through volume, every failed search costs a sale.
Ecommerce product discovery or product management: do not confuse the two
Product teams use the words “product discovery” in another sense. For a product manager, discovery answers the question “what should we build?”. Marty Cagan structures it around four risks (value, usability, feasibility, viability) and Teresa Torres makes it continuous with weekly contact with users. That discipline draws on user research and hypothesis validation. Ecommerce product discovery asks a different question: how does a shopper find the right product inside an existing catalog? Both disciplines share one reflex, testing before deciding. They share neither the same actor nor the same tools.
How do you improve ecommerce product discovery?
Navigation and categories
Navigation is the first lever of product discovery. A readable tree groups products by the shopper's vocabulary rather than by the catalog's internal organisation. A few clear categories beat a deep menu. Breadcrumbs and category pages that announce their contents keep the customer from getting lost in the online store. On mobile, navigation has to stay usable with one hand, because a growing share of visits starts on a phone screen. Constant landmarks help the shopper explore the catalog through a customer experience that stays stable from page to page. Every category page acts as a guide. It presents the store, announces the available filters and points the shopper to the right product page.
Site search and semantic search
Site search carries real weight: the shopper who uses it often knows what they want and expects an immediate answer. A useful search bar offers autocomplete, tolerates typos and avoids the zero-result page. An analytics tool spots those empty pages and guides the work of fixing the engine's vocabulary. Semantic search goes further: it understands the intent behind natural language, where a lexical engine compares words. It also reads the product image and synonyms. The search bar stays visible on every page so the shopper keeps control of the journey. Our analysis of AI-powered search details this shift in ecommerce search.
Filters and facets
Filters and facets cut a catalog of thousands of items down to size. Every filter is useful on one condition: the attributes on offer must match the shopper's decisive criteria (size, budget, material, compatibility). A dynamic filter that adapts to the category shortens the time to the product through targeted attributes. A useless filter adds noise to the page. A well-tuned filtering tool keeps the shopper in the store through familiar criteria. A missing filter sends them off the site. The choice of facets is therefore decided category by category, from real queries and from user behavior. Their limit stays the same: the shopper has to know their criteria already.
Product recommendation
Product recommendation suggests items from browsing behavior and customer history. A recommendation engine feeds personalization on the home page, the product page and the cart through its recommendation algorithms. Shopper behavior (pages viewed, add to cart, past purchases) and browsing behavior act as signal. That engine calls for a dedicated team and a control tool to govern personalization across the site. Personalization applies to a marketplace as much as to a brand site. Personalization pushed too far ends up locking the shopper into the same products, which is why it pays to keep it measured. Back in 2021, McKinsey measured that 71% of consumers expect personalized interactions. This lever assumes a history though, which is exactly what a first-time visitor lacks. Our guide to product recommendation sets out its mechanics.
Product data and attributes
No lever makes up for poor product data . Search, filters and recommendation read the same attributes: title, description, images, variants, stock, price. Every attribute (colour, material, compatibility) and every feed powers the filter, the search and the recommendation. A missing attribute makes the product invisible to an engine that only reads what is declared. A clean product feed, with one image per variant, is built upstream with the catalog team and the ecommerce team. This catalog work is first of all a question of technical quality. A control tool, run by the catalog team, keeps every product page complete through automatic rules. For a marketplace as much as for a brand store, enriching product data is often the most profitable project to run before buying any tool.
E-merchandising and discovery pages
L’E-merchandising drives the display order of products against business goals: margin, stock, seasonality, new arrivals. Sorting rules, boosting, visual merchandising and a seasonality attribute turn a result list into a storefront. Discovery pages (collections, themed selections) guide shoppers who arrive without a precise query. A merchandising strategy is judged on the engagement it creates, not on the number of rules configured. The goal is to deepen shopper engagement on every page of the store. The ecommerce team tunes the site through those rules, page by page, and keeps a best-practice guide up to date.
Why classic levers reach their limit
Where each lever fails, query type by query type
The Baymard benchmark Ecommerce Search UX 2026 rests on more than 170 sites and 10,000 ratings. It concludes that 56% of sites fail to adequately support their users' search needs, a figure detailed in our article on AI-powered search. The failure rate varies by query type: 12% on an exact query and 66% on a non-product query. No lever fails everywhere. Each one fails on a precise query type, which changes how you choose a discovery lever.
Query type and matching lever: the diagnostic table
The table below ties each query type measured by Baymard to an example. The examples come from travel, real estate, automotive and retail. The last column names the lever that answers best.
The reading is clear: use case and non-product queries show the highest failure rates. A symptom query, like “sore throat” or “oil leak” in Baymard's own examples, surfaces unrelated results. Semantic search and filters fix part of the problem. Recommendation then helps the shopper, provided recommendation has a history to work from. For a need that is still vague, conversational discovery takes over with the shopper, step by step.
What do you do when the shopper does not know what they are looking for?
About 60% of shoppers do not know what they want at the start of a high-consideration journey (Kleio data). A traveller looks for “a family trip somewhere sunny”. A household looks for “a flat near a school”. Each one states a use case rather than an exact query. The need sharpens by discovering offers along the customer journey. Filters and the search engine assume a need already put into words, which explains the 43% failure rate on use case queries.
Which indicators measure product discovery?
Measuring product discovery starts with four indicators. Track the zero-result query rate, the exit rate from the results page and the conversion rate of search. Then compare average order value across sessions with and without search or recommendation. Cart abandonment completes the picture, because it reveals the doubts that survive discovery. No blog post replaces your own data: setting these indicators before any optimisation is what makes the real effect of each lever readable. Every indicator should tie a lever to a figure, step by step. A reporting tool guides the team on which lever to activate across the store and on each key page.
The pitfalls of ecommerce product discovery
Four pitfalls come back in most projects:
- stacking tools without cleaning product data
- optimising only the exact query
- measuring the click rather than the preference built
- neglecting mobile, where typing is painful
Avoiding them takes more discipline than technology. A team that first fixes its data and its zero-result queries moves faster than a team that changes tool every year. User feedback stays the best compass for improving the strategy over time.
Diagnostic: which lever should your catalog activate first?
Does your catalog need better search, better data or guided discovery? Answer four questions to get the priority lever, two concrete actions and the indicator to track.
This diagnostic sometimes recommends fixing site search first. It is a prioritisation guide for the team: it complements a full audit rather than replacing one. It puts words on the gap between what your store offers and what your shoppers are able to express.
Agentic discovery: the AI layer that takes over
How does AI-driven discovery commerce work?
AI-driven discovery commerce understands the request in natural language and reasons over catalog attributes. It talks with the shopper to narrow the selection. It differs from keyword search, which compares terms. The stake is growing. According to Adobe Analytics, traffic to US retail sites from generative AI assistants rose 693% year over year during the 2025 holidays. Those visits convert 31% better than other sources. This conversational buying experience joins Google and Amazon among the shopper's entry points.
What Gartner says: explore yes, delegate no
Only 11% of consumers would let an AI decide a purchase (Gartner, May 2026). On the other hand 31% accept that it narrows the choice in household products and 28% in personal electronics. Among shoppers who used generative AI, 54% had to double-check everything and 62% consider they lost time. Those US surveys date from November and December 2025 and from January 2026. The lesson is clear: the shopper wants help exploring, what the trade calls guided selling, not a delegated decision.
The propose, react, revise loop
L’Agentic discovery helps a shopper build a preference by reacting to real products. It is about guiding the shopper, not about exposing a catalog to external agents. The diagram compares the funnel journey, where every step loses shoppers, with the agent's loop: it proposes, the shopper reacts, the agent revises. Our page on the agentic discovery concept sets out that mechanism.
A refusal becomes a signal. A refusal on price revises the estimated budget. A refusal on distance revises the area. The loop converges toward a preference clear enough to move to a booking, a quote or a meeting with an advisor.
Agentic discovery or product recommendation?
A recommendation engine ranks products from behavioral history. It therefore assumes a history, which is missing for a shopper discovering a rare category: recommendation stays silent without a past signal. Agentic discovery differs because it works without history: it starts from the conversation and from live reactions. The two approaches complement each other. Recommendation serves the last step of the journey. Discovery serves the first ones, through a conversation that needs no history at all.
Hospitality, real estate, automotive: where discovery is decided
Buyer-side discovery weighs most on high-consideration purchases, with a long decision, a large catalog and weak intent at the start. In hospitality and travel, the traveller discovers the destination at the same time as the offer. In real estate, the buyer revises the project with every viewing. In automotive retail, they move from a budget to a use case then to a model. Travel is not Automotive and Real Estate is not Wholesale: each industry has its own exploration logic.
How do you choose a product discovery platform?
Choosing a platform starts with one question: where in the journey do you lose the shopper? Tools fall into three families: search and merchandising, recommendation, conversational discovery. Strong recommendation demands clean data where conversational discovery starts with no history at all. Compare each discovery platform, the product discovery tools and the product discovery software on four criteria: natural language understanding, integration with your product data, impact measurement and answer governance. Every application has to fit your technical context without rebuilding the site. An internal selection guide, kept by the ecommerce team, avoids comparing tools on vague criteria. A vendor blog comparison stays tilted toward its own offer: ask for a demonstration on your own data. One obligation is added: since 2 August 2026, article 50 of the EU AI Act requires telling users they are interacting with an AI.
Kleio: agentic discovery for high-consideration purchases
The Knowledge Engine: a live catalog
Kleio is neither a site search engine nor a merchandising tool. We step in when the shopper's need is vague, on a complex catalog. The Knowledge Engine distributes the catalog across three stores: products, documents and memory. It feeds every answer from approved data. It reaches 99% precision with answers in under 3 seconds and zero hallucination. The agent answers only from approved data. Our blog article on the Knowledge Engine sets out that architecture.
The Triple Business Ontology: 7 industries
The Triple Business Ontology layers three pre-built levels: function, industry and customer. It exists for 7 industries: Travel, Automotive, Real Estate, Wholesale, Manufacturing, Insurance, Energy & Utilities. The Knowledge Engine and the ontology work together, so the agent knows what counts as a decisive criterion for a traveller or for a property buyer. We extend that ontology to each client's data model in a matter of weeks.
Orpi, Selectour, Havas Voyages, Altarea Cogedim: journeys in production
At Orpi, the Kleio platform runs on Orpi.com and across its 1,250 agencies, serving 8,000 advisors. The deployment took three months (see the Orpi case study). Selectour equips its travel advisors and its customers. Havas Voyages is growing its lead generation through a conversational experience fitted to each traveller. Altarea Cogedim absorbs the complexity of its offers and financing schemes, with a customer experience fitted to each project. We deploy Kleio in 8 to 12 weeks with a dedicated project team, within a SOC 2 framework with tenant isolation. The Kleio blog carries other field reports on recommendation and discovery.







