Wednesday, 16 September 2026
Artificial intelligence is beginning to alter one of the foundations on which online retail has been built for more than two decades: the assumption that consumers will search for a product, click through to a retailer and make their purchasing decision inside the retailer’s digital environment. Increasingly, the first part of that journey can take place somewhere else. Consumers can describe what they need to an AI assistant, compare products and prices, refine their requirements conversationally and arrive at a retailer already carrying a much clearer purchasing intention. For retailers, that represents a potentially significant shift in who controls product discovery, customer acquisition and ultimately the economics of e-commerce.
At AI4 2026 in Las Vegas, representatives from Nordstrom, Staples, Mirakl and technology consultancy CI&T discussed how artificial intelligence is changing the journey from product discovery through conversion and post-purchase service. The discussion suggested that the immediate transformation is less about autonomous AI agents purchasing everything on behalf of consumers and more about a gradual restructuring of how shoppers find, evaluate and select products. One of the clearest changes is appearing in search behaviour. Retail search was historically dominated by relatively short queries: a customer might enter a product category, brand or basic description and then navigate filters to narrow the results. AI assistants are encouraging consumers to express requirements much more naturally and in considerably greater detail.
Nordstrom described searches evolving from short product requests towards longer, more contextual questions. Instead of looking simply for a dress or pair of shoes, customers can describe an entire occasion and ask for an outfit. They can subsequently reject one recommendation while retaining the rest, creating a conversation around the purchase rather than restarting a conventional search. The distinction matters because it changes what retailers need to know about their own products. Traditional e-commerce could rely heavily on categories, keywords and structured filters. Conversational shopping requires much deeper information about how a product will actually be used.
A customer may not search for a technical product attribute at all. They may ask for running shoes suitable for long distances and a particular type of foot, an outfit appropriate for a specific wedding, or office equipment capable of meeting a certain workload. An AI assistant must translate that intention into products whose characteristics satisfy those requirements. That makes product information one of the most strategically important assets in the emerging retail AI market.
Retailers have spent years improving titles, descriptions, images and structured attributes, primarily to support website search, merchandising and search-engine visibility. AI-driven discovery raises the standard considerably. Models need sufficient information to understand why one product is appropriate for a customer’s particular circumstances and why another is not. Staples provided an example involving printer cartridges. A model number may contain information indicating whether a cartridge is a high-yield version capable of substantially greater printing capacity. If that meaning is not captured correctly within the retailer’s product information, an AI system can struggle to answer a customer’s underlying requirement even though the correct product exists in the catalogue.
Reviews and customer questions can become important as well. A shopper asking for a highly rated product, or one that performs well under a particular set of conditions, is requesting information that may sit within thousands of customer reviews rather than conventional catalogue fields. Retailers therefore need ways of making that unstructured information accessible to recommendation and discovery systems. The potential consequence is a shift from traditional search-engine optimisation towards a broader effort to make products understandable to AI systems. The objective is no longer simply achieving a prominent position in search results. Retailers increasingly need their products to be selected when an external AI assistant reduces hundreds or thousands of possibilities to a handful of recommendations.
That reduction in choice makes assortment strategically important. In conventional online retail, consumers can browse pages of results. An AI assistant may present only several products. If the retailer does not have an item closely matching the consumer’s increasingly precise requirements, it may disappear from consideration before the customer ever reaches its website. Mirakl argued that large catalogues can therefore become more valuable rather than less valuable in an AI-led discovery environment, provided the underlying product information is sufficiently strong. A broad assortment gives the system a greater chance of finding an exact match for increasingly detailed customer requirements.
This also has implications for online marketplaces. Marketplaces historically benefited from aggregating supply and allowing consumers to compare products within one destination. AI assistants can perform part of that aggregation externally, potentially comparing products across several marketplaces, retailers and brands simultaneously. The retailer’s challenge consequently becomes twofold: it needs to be selected by external agents while also developing a sufficiently useful internal experience that customers have a reason to begin future searches directly with the retailer. That tension may become one of the defining competitive battles in e-commerce.
Retailers including Nordstrom are already seeing referral traffic originating from generative AI services, although conventional search and digital advertising remain much larger sources of traffic. The significance lies less in today’s absolute volume than in what happens if consumers become comfortable beginning commercial searches through AI. The economics of customer acquisition could change substantially. Search engines, social networks and marketplaces have built enormous advertising businesses around controlling the point where consumers discover products. If AI assistants capture a meaningful share of that activity, retail marketing budgets may eventually follow them.
For the moment, some AI-generated referrals can reach retailers without the same established advertising economics associated with traditional search marketing. It would be risky, however, for retailers to assume that this will remain permanently free. As AI platforms become more important to commercial discovery, new advertising, referral, commission or transaction models are likely to develop. This makes direct customer relationships increasingly valuable. A retailer that receives a customer from an external AI platform has an opportunity to persuade that shopper to use its own search, recommendation or shopping assistant in the future rather than repeatedly relying on an intermediary.
Once consumers arrive on the retailer’s website, AI is also changing how products are recommended. Traditional recommendation engines frequently relied on historical behaviour processed through scheduled batch systems. More advanced models can interpret sequences of behaviour including products viewed, items placed into baskets, previous purchases and the timing of those actions. That can provide a better indication of purchasing intent. A consumer who has researched one category repeatedly over several weeks may require a different recommendation strategy from someone making an impulse purchase. AI can potentially distinguish between those contexts and adjust recommendations accordingly.
Conversational commerce is also returning in a more capable form. Retailers experimented with chatbots and virtual shopping assistants long before the current generative AI cycle, but the systems were often constrained by predetermined questions and limited product knowledge. Generative AI allows customers to communicate in much more natural language. The opportunity is particularly significant for retailers with large assortments. Instead of navigating extensive menus and filters, consumers can describe what they need. For the retailer, those conversations provide another valuable source of information because they reveal how customers naturally think about products.
A conventional filter might tell a retailer that a customer selected a colour, price range and size. A conversational request can reveal that the customer is attending a birthday party, starting a new job, renovating a kitchen or buying a gift. That context can improve merchandising and product development as well as recommendations. Yet the panel repeatedly returned to a less glamorous requirement underlying these experiences: data.
Retailers have accumulated information across decades of store systems, e-commerce platforms, loyalty programmes, supply chains, merchandising tools and point-of-sale systems. Much of it remains fragmented, duplicated or difficult to reconcile. Generative AI cannot automatically transform inconsistent information into reliable commercial decisions. According to CI&T, data modernisation is consequently becoming one of the most important areas of AI-related investment among large retailers and brands. Companies are consolidating information, improving context and connecting previously isolated datasets before attempting more sophisticated customer-facing applications.
The reason becomes obvious when AI begins making recommendations. A conversational shopping system may appear impressive while still failing basic retail tests. Mirakl described testing a retailer’s AI shopping experience and finding that products presented by the system frequently did not correspond with the retailer’s established best sellers. In another test, a customer request specifying a maximum budget produced recommendations whose average price was substantially above the stated limit. These failures demonstrate that sophisticated language capability does not replace retail fundamentals. The right product still needs to be offered at the right price, with accurate availability and realistic delivery information.
This becomes even more important when the AI system rather than the consumer is evaluating the alternatives. If an external assistant receives an incorrect price, outdated stock position or incomplete product specification, it can remove the retailer from consideration before the customer sees the offer. Inventory accuracy and substitution therefore become part of AI readiness. If a recommended product is unavailable, the retailer needs sufficiently rich product information to provide an appropriate alternative. Sizing, compatibility, delivery dates and location-specific availability all need to remain synchronised across channels.
Trust consequently moves earlier in the purchasing journey. Historically, retailers earned much of their trust after the customer had chosen a product: the item arrived as described, delivery occurred on time and problems were resolved effectively. AI-assisted commerce requires some of that trust before the purchase because consumers increasingly rely on automated recommendations to decide what they should buy. The panel argued that transparency will be important. Consumers should know when they are communicating with an AI system rather than discovering this after believing they were speaking with a person. Accuracy is equally important. A retailer cannot easily blame an external AI assistant when the price displayed to a customer differs from the price available at checkout. The customer ultimately associates the failed experience with the brand selling the product.
Post-purchase service represents another major application. AI agents can increasingly handle routine questions involving delivery status, returns, product information and compatibility without requiring customers to navigate telephone menus or wait for an employee. The potential extends beyond simple order tracking. If the retailer has a sufficiently detailed knowledge structure around its products, an AI service agent can potentially answer technical questions about whether a purchased component works with another product or explain how something should be used.
The next stage is agentic commerce, where AI moves from advising consumers towards taking actions for them. Predictions about how quickly fully autonomous shopping will develop varied considerably across the panel. Some participants expect agent-to-agent transactions to emerge relatively quickly in repetitive categories, while others believe consumer commerce could remain primarily focused on AI-assisted discovery for years. The distinction between categories is important. Consumers may readily allow an AI agent to reorder milk, household supplies or other frequently purchased goods within predetermined spending limits. Allowing the same system to autonomously select a wedding outfit, luxury product or other emotionally significant purchase requires a very different level of trust.
This suggests that agentic commerce will not arrive uniformly across retail. Business-to-business commerce may provide an earlier proving ground. AI agents can already assist with demand forecasting and routine ordering between companies where purchasing rules, product specifications and approved suppliers are clearly defined. Subscription-style consumer purchases could follow a similar model. Even in a world of autonomous agents, loyalty is unlikely to disappear. An AI assistant acting for a consumer can incorporate that consumer’s existing memberships, rewards, preferences and retailer relationships into its decisions. A customer who repeatedly shops with one retailer may instruct an agent to favour that company even when another seller offers a marginally lower price.
That means brands and retailers still need differentiation beyond price. If autonomous systems can compare thousands of offers almost instantly, competing purely on product availability could become increasingly difficult. Loyalty programmes, service quality, trust and exclusive assortment may become more important rather than less. Payments and fraud prevention will also need to evolve. Retailers must be able to determine whether an automated agent has legitimate authority to make a purchase on behalf of a customer, while payment networks need mechanisms for authenticating those transactions. Work on agent-authorised payments is already developing across the financial and technology industries, but mainstream adoption will require confidence that automated transactions can be controlled, disputed and audited.
Retailers simultaneously face the operational problem of scaling AI. A demonstration involving several products can work convincingly while an enterprise deployment covering millions of customers, locations and product combinations fails because of latency, cost or inconsistent data. The panel cautioned against assuming that the largest or most sophisticated model will necessarily produce the best commercial result. Search and recommendation systems operate in an environment where response speed directly affects conversion. A highly accurate model that takes too long to produce results may generate worse commercial performance than a simpler model capable of responding almost immediately.
The strategic question is therefore moving away from who has the most advanced AI model. Foundation models are becoming increasingly accessible. Competitive advantage is more likely to emerge from proprietary product information, customer understanding, operational data and the infrastructure connecting those assets. For large retailers, that changes the investment priority. The visible AI shopping assistant may attract customer attention, but the more valuable work can be happening underneath it: rebuilding product information, connecting inventory, modernising data infrastructure, establishing governance and creating systems that can deliver consistent information to websites, stores, external AI assistants and future autonomous agents.
The transition also has implications for physical retail. AI-led commerce does not necessarily mean the disappearance of stores. In categories where customers want to touch, try or experience products, physical locations can remain central to the relationship even if AI performs much of the initial research. Instead, AI could make the boundaries between digital and physical retail less important. A customer might begin with an AI conversation, visit a store to evaluate a product, complete the purchase digitally and subsequently use another automated agent for service.
The larger transformation is therefore not the replacement of the e-commerce funnel but its fragmentation across more interfaces. For the past two decades, retailers invested heavily in attracting consumers to websites and applications where they controlled search, merchandising, recommendations and checkout. AI assistants threaten to move some of those functions outside the retailer’s own environment. That creates a new competitive question. Retailers must decide whether they primarily want to supply products to other companies’ AI agents or become trusted AI destinations themselves.
The companies best positioned for either outcome may not be those with the most impressive chatbot. They are likely to be those with the cleanest product information, most accurate inventory, strongest customer relationships and infrastructure capable of exposing that information consistently wherever the purchasing decision takes place. AI may eventually allow software agents to negotiate and purchase products without consumers visiting a conventional online store at all. But the more immediate transformation is already simpler and commercially significant: customers are becoming accustomed to describing what they want rather than searching for it.
That changes the route to conversion. And as control over product discovery begins to move away from search engines and retailer websites, the battle over who owns the customer relationship is entering a new phase.
Source: CIJ.World Research & Analysis Team
