The retail-search procurement landscape in 2026 is dominated by the proposition that the search interface should be substantially upgraded with large-language-model integration; the upgraded search is sold as a means of permitting the visitor to query the catalogue in natural language and receive results that respect the visitor’s intent rather than the visitor’s specific keywords. The proposition is, as a description of the technology’s capabilities in laboratory conditions, broadly accurate; the proposition is, as a description of where the conversion uplift on a typical retail catalogue is to be found, decisively wrong.
The conversion uplift, on the data I have collected across multiple retail clients, is concentrated in a different layer of the search infrastructure entirely. The layer is the auto-complete and typo-tolerance layer, which the AI-search procurement narrative treats as a solved problem and which is, on the catalogues I have audited, almost universally under-implemented. The uplift available from completing the existing layer’s implementation exceeds, by a substantial margin, the uplift available from layering an AI-search capability on top of the existing search. This post is an account of why the procurement narrative has the priority order wrong, and where the actual uplift is to be found.
What the AI-search procurement materials are claiming
The AI-search procurement materials, in their dominant form, present a sequence of demonstrations: a visitor types “comfortable shoes for standing all day” and the AI-search returns the appropriate footwear category; a visitor types “duvet for hot sleepers” and the AI-search returns the appropriate bedding category; a visitor types “outdoor table that fits four people and survives winter” and the AI-search returns the appropriate garden-furniture category. The demonstrations are, in their isolated form, impressive, and the visitor’s experience of the search interface is, in the demonstrated cases, considerably better than the corresponding experience with the conventional keyword-matching search.
The procurement materials, having presented the demonstrations, project the conversion uplift the AI-search would produce against the retailer’s existing search interface. The projections are, in the materials I have audited, in the range of fifteen to thirty per cent uplift on the search-derived conversion contribution; the figures are presented with confidence and are typically supported by case studies of large platforms whose deployments produced figures broadly consistent with the projections.
What the procurement materials are not addressing
The conversion uplift projections rest on a comparison between the AI-search and the existing keyword-matching search, with the implicit assumption that the existing search is functioning at the level the keyword-matching architecture would, in principle, support. The assumption is, in essentially every retail catalogue I have audited, incorrect. The existing search is functioning substantially below the keyword-matching architecture’s capability, and the gap between actual and capable performance is, on the conversion data, the fraction of the search-derived conversion uplift the procurement materials are attributing to AI-search.
The under-performance of the existing search is concentrated in three specific failure modes that are, on inspection of the search infrastructure, almost universally present.
The first is the absence of competent typo tolerance. A visitor who searches for “shoes” with the typing accuracy that mobile devices and rapid input produce types, in a non-trivial fraction of cases, “shose”, “shoses”, “shooes”, or some other variant. The conventional search implementation, configured against an exact-match index, returns no results for these variants; the visitor’s experience is that the catalogue appears not to contain any matching products. The implementation work to introduce typo tolerance is straightforward — a single configuration change in the great majority of search implementations — and the implementation produces, on the data I have collected, conversion improvements that are, on the median catalogue I have audited, approximately twelve per cent of the total search-derived conversion contribution.
The second is the absence of competent stem-matching and synonym resolution. A visitor who searches for “running shoe” against an index that contains “running shoes” but not “running shoe” returns, on a great many search implementations, zero results despite the obvious match. The implementation of stemming and synonym lists is, again, straightforward in the standard search engines that retail platforms typically use; the implementation is nevertheless absent from the great majority of the catalogues I have audited.
The third is the absence of competent auto-complete. A visitor who has typed several characters of a query and is presented with no auto-complete suggestions is, in a non-trivial fraction of cases, more likely to abandon the search than a visitor who is presented with relevant suggestions. The auto-complete implementation is, in the dominant search engines, available out-of-the-box but requires explicit configuration of the suggestion source; the configuration is, in essentially every catalogue I have audited, either absent entirely or implemented against a poor source (such as the catalogue’s product titles only, with no behavioural input from the visitors’ actual queries).
The conversion uplift comparison
The deployments I have shipped that addressed the three failure modes against the existing keyword-matching search produced, on the conversion data, uplifts in the range of nine to seventeen per cent on the search-derived conversion contribution. The figures are broadly comparable to the AI-search procurement-material projections, achieved against the existing infrastructure with implementation costs that are, in the great majority of cases, less than ten per cent of the AI-search procurement cost.
The deployments where the AI-search was layered on top of a search infrastructure that had already been brought to keyword-matching competence produced additional uplifts in the range of three to six per cent. The additional uplift is real and positive; it is also substantially smaller than the procurement materials project, and is consistently smaller than the uplift from the underlying keyword-matching work that should have preceded it.
The implication is that the conversion uplift the AI-search procurement materials project is, in the great majority of catalogues, substantially attributable to the keyword-matching work that the materials’ demonstrations implicitly assume to be already in place. The work is, on the catalogues I have audited, almost never in place; the AI-search procurement is, in this sense, paying for the keyword-matching uplift via the AI-search invoice. The retailer who proceeds with the AI-search procurement before completing the keyword-matching work is paying for the keyword-matching uplift twice — once as part of the AI-search and once if the retailer subsequently completes the keyword-matching work — and the cost is, on the implementation costs of the two paths, an entirely avoidable expense.
An advisory close
The AI-search-for-retail procurement narrative is, on the data, addressing a real conversion opportunity but is doing so via a procurement path that is approximately ten times the cost of the alternative path that produces broadly comparable conversion outcomes. The alternative path — the completion of the keyword-matching, typo-tolerance, stemming and auto-complete work that the existing search infrastructure depends on — is operationally available, requires modest engineering investment, and produces uplift that the AI-search procurement is, in the procurement materials, attributing to itself.
It is recommended that retailers considering AI-search procurement audit the underlying keyword-matching search’s performance before committing to the procurement; the audit will, in the great majority of catalogues, surface the implementation gaps that the AI-search procurement materials assume to have been closed already, and the closure of the gaps is the considerably more cost-effective intervention than the AI-search layering above them. (For the related view on the broader procurement narrative around AI integration in retail commerce, see AI Interactive FAQ: building a customer-question widget that does not hallucinate.)