Kesko's search case: product discovery before agentic checkout

Research synthesis of external sources, not a Sander Gansen client implementation. Published 3 November 2026; case and observation dates are distinguished below.

A manufacturer can offer a suitable drill and still lose the buyer before price becomes relevant. A search for the main product may return batteries, spare parts and loosely related accessories first. An assistant working from those results inherits the same confusion. The practical priority is therefore not permission to check out autonomously. It is helping the buyer identify the right product and understand what belongs with it.

What changed across Kesko's six domains

LupaSearch describes an implementation for Kesko Senukai Digital covering 1a.lt, 1a.lv, 1a.ee, senukai.lt, ksenukai.lv and k-rauta.ee. Its case identifies approximately 500,000 active SKUs, migration from Nosto and five months of A/B testing. This is a Baltic, multi-domain retail case, not evidence of an Estonia-only experiment. The public page does not give calendar dates for the test. 1

The intervention combined language handling, ranking and interface work. Estonian compounds and singular/plural forms were addressed alongside AI-assisted synonyms and longer-query interpretation. High-traffic keywords were tested manually. Sales data informed ranking, broad brand searches received category suggestions, and principal products were separated from accessories. Faster loading and seasonal relevance were also part of the work. These details matter because an improvement cannot credibly be assigned to one language model feature when several parts of the buying journey changed together. 1

LupaSearch reports search-conversion growth of 22.9–69.3% and revenue-per-session growth of 33.5%. Those are provider-published outcomes. The page does not disclose per-domain baselines, experimental assignment, traffic denominators, confidence intervals or contribution margin. Nor does it define the precise event counted as “search conversion.” The upper end of the range must not become a promise that every Estonian manufacturer will increase sales by 69.3%. 1

The retained research includes a narrower observation from 27 October 2026: a query for “akutrell” returned categories and product cards, and a product URL included src=lupasearch. That supports the existence of an accessible search interface on that date. It neither reproduces the A/B result nor demonstrates an autonomous purchase. No new checkout test was performed for this brief.

Product identity comes before recommendation

Discovery is finding a potentially relevant offer. Recommendation adds a judgement about suitability. Checkout adds commercial and transactional decisions. A result can pass the first step and still fail the next two. A battery appearing for a drill query is not necessarily irrelevant, but it should not be presented as the drill itself. A plausible accessory is not evidence of approved compatibility.

For a manufacturer selling through dealers, this makes product relationships a commercial responsibility. The manufacturer part number, dealer identifier and local product name need to refer to the same variant. Voltage, size, material and pack quantity must remain distinguishable. Main products, consumables and spares need explicit relationships, including exclusions where a combination is not approved. Otherwise better query interpretation merely retrieves ambiguous records more efficiently.

Google's currently accessible guidance supports existing SEO foundations for its generative Search features. It says there is no special AI schema or llms.txt requirement and warns against producing scaled content primarily to manipulate visibility. This guidance concerns Google Search, not every shopping assistant. Clear technical information and usable feeds can support discovery without guaranteeing inclusion, recommendation or a sale. Current page wording should not be treated as a frozen October observation. 2

An illustrative manufacturer test

Consider an electrical-component producer exporting through several distributor catalogues. This is an illustrative application, not a reported deployment. Select one product family with frequent buyer confusion. Preserve its canonical identifier, translated names, controlled synonyms, variants, accessories and approved compatibility statements. Ask the technical owner to confirm which alternatives are genuinely acceptable before marketing adds broader matching terms.

Build a versioned query set from actual buyer language: local compounds, plurals, part numbers, shorthand, application descriptions and ambiguous brand queries. For each query, record the expected family, acceptable alternatives and prohibited alternatives. Save the observed result, channel, date, source record and correction owner. Include queries that should lead to clarification rather than an automatic recommendation.

Measure zero-result rate separately from wrong-item rate. A search engine can reduce empty results by returning more unsuitable items. Measure correctly defined search conversion separately from clicks, and connect qualified enquiries to quotations and orders where the channel permits. Price, stock and delivery must be confirmed at the merchant-owned commercial step, not inferred from an old product description.

Questions before expanding the experiment

Who owns a wrong variant: the manufacturer, distributor or search provider? How quickly does a corrected fact reach every channel? Can the team distinguish an accessory click from a principal-product purchase? What result should trigger abstention or a question to the buyer? Would the test still pass if a popular but incompatible product ranked first?

Start with one governed family and a repeatable correction loop. That produces a useful discovery asset before any agent receives buying authority. Explore the broader OfferSignal research context, the Business & AI Workshop, and the related brief on governed product information.

References