Retail's Next Storefront May Have No Shop Window

Over the years I have often watched physical stores compensate for weaknesses in their own data. The same product carries different names in the POS, the ERP and the accounting system. Stock updates arrive late. The details of a promotion are known to the staff and never make it into the product record. None of it stops a customer, who walks in, sees the item on the shelf and asks an employee for whatever is missing.

Until recently the consequences of that disorder stayed inside the company. They distorted reports, complicated stock management and created extra work — the cost I have written about before. As more customers begin searching for products through AI, the same inconsistencies start to decide something else: whether a retailer's offer is seen at all.

AI is entering the buying journey before the customer opens a website or walks through a door. It helps people articulate what they need, compare alternatives and reduce a broad market to a few plausible choices. Agents are beginning to go further — checking price and availability, assembling baskets, reserving products and, within limits the user sets, completing payment.

Search is changing before purchasing does

The infrastructure is already taking shape. Google's Universal Commerce Protocol gives agents a way to retrieve live product variants, inventory and pricing from a retailer's catalogue (March 2026). Shopify switched its merchants into AI shopping channels by default the same month — ChatGPT, Copilot, Google's AI Mode, Gemini. Adyen connected product feeds, carts and payments into one integration in June, though so far only in limited availability for enterprise merchants in the US. And in July, Visa reported agents completing real purchases at participating merchants across Europe, rather than at test storefronts.

It is too early to speak of a mass shift to autonomous buying, and the survey evidence says so plainly. Gartner found willingness to let AI make the final purchase decision topped out at 11 per cent, and that was in low-stakes categories like personal care and household supplies. Willingness to let AI narrow the choices ran higher — 31 per cent for household supplies, 28 per cent for personal electronics. That was 322 US consumers surveyed in January 2026, which is a small sample in one market, and worth holding lightly.

For a retailer, though, the shortlist already matters more than the checkout. Adobe measured traffic from generative AI services to US retail sites up 393 per cent year on year in the first quarter of 2026, after a 693 per cent rise over the 2025 holiday season. The same report is honest about the shape of that curve: March on its own was up 269 per cent, so the growth rate is falling off a very low base even as the absolute numbers rise. NIQ found 42 per cent of US consumers had used at least one AI tool to shop in the past month, from a monthly sample of around 500.

The number I find harder to dismiss is not the traffic. In March 2026 that AI traffic converted 42 per cent better than everything else — a year earlier it converted 38 per cent worse.

Every consumer figure here is American, drawn from samples in the hundreds, and none of it should be mistaken for a comparable share of retail sales in Europe. The direction is still difficult to ignore. Some product discovery is moving into an environment where the customer receives a handful of recommendations instead of a page of links, and being absent from that handful may eventually cost as much as losing the sale at the till.

A digital storefront without e-commerce

This leads somewhere less obvious. A retailer with no intention of building a conventional online shop may still need a digital storefront.

It does not require customer accounts, online checkout, acquiring or home delivery. Its first purpose is more modest: to let a person, or an agent acting for that person, understand what the retailer sells, where a product is available, what it costs and what can happen next. Consider an ordinary request — find two cases of still water nearby that I can collect after 8 p.m. today, check whether my loyalty discount applies, and reserve them.

Knowing that a chain sells bottled water is not enough for that. The agent has to identify the relevant pack sizes, check availability at individual stores, understand their opening hours, apply the loyalty rules and work out whether a reservation is possible. A store may be around the corner, hold the product and offer the best price, and still fall outside consideration because none of that exists in a form a system can use. The transaction may still happen at the physical till. What has moved online is the choice of where to go.

Such a storefront has to be legible to both audiences at once. A customer sees ordinary product and location pages. An agent needs structured attributes, machine-readable prices and inventory, product feeds and clearly defined actions. A dedicated subdomain can host all of it, but publishing a few web pages does not by itself make a retailer visible to AI systems.

A consistent view of the business comes first

Creating product pages is technically much easier than making sure what they show is true.

The master catalogue lives in one system, current sales in another, purchase prices in accounting documents, promotions somewhere else entirely. Transfers between stores are recorded late. Returns and write-offs follow their own procedures. Total inventory can look perfectly plausible while its distribution across locations no longer reflects reality — the difference between a total and a breakdown that decides most investigations.

An experienced manager knows how to navigate those inconsistencies. They know which report needs a second look, whom to call, and which corrections the staff carry in their heads. An external agent has no access to that informal knowledge, and no way to ask.

This is why normalisation has to come before publication. Product identifiers, variants, units of measure, categories and locations have to be reconciled. The retailer needs to establish which system is authoritative for price and which for inventory, and how quickly a change reaches each channel. The next layer is the context the business operates in: a product connected to specific stores, their locations and opening hours, fulfilment options, additional services, exchange rules and loyalty programmes. That operational knowledge graph is what lets a system understand the conditions under which a product can actually be sold, rather than treating it as an isolated catalogue entry.

At marql we approach this from inside the operational environment. Data from POS, ERP, accounting and other systems is connected read-only and brought into one consistent model of products, stores, sales and inventory. Until now that foundation has served management analysis and operational recommendations. The same foundation can be used to make the business legible to customers and to external AI agents.

Two practical models

Neither of the two below is something marql ships today. What exists is the foundation under them — the normalised model of products, stores, sales and inventory described above, running on read-only connections to systems chains already have. These are what we intend to build on it.

Where a retail chain already has a website or an application, the first model will put a B2C agent inside that existing channel. The customer would talk to the retailer directly: asking questions, checking availability, comparing products, assembling a basket for a recipe or a celebration, reserving items at a particular store.

The same interface will be able to carry recurring purchasing instructions — a standing order of drinking water or a basic grocery basket, with the schedule, the budget, the permitted substitutions and the actions that still need confirmation all defined up front. Clear limits will matter there. The convenience of a recurring purchase should not become an unrestricted right to spend the customer's money, which is the same boundary question that decides where an automated system may act on its own inside the business.

For retailers with little or no online presence, the second model will be a hosted storefront on a dedicated subdomain, prepared for people and agents alike. The customer journey might end with a reservation, a phone call, directions to the nearest store or an in-person visit. Full e-commerce can follow later; it is no longer the only route to digital visibility.

Both would begin with the same work: operational data collected, normalised, checked and enriched with business context. Skip that stage and the new storefront simply exposes the company's internal inconsistencies to the outside world.

Even if agentic commerce develops more slowly than the platforms currently predict, this work is not wasted. Consistent product data improves reporting, assortment management, stock control and the sales channels that already exist. Readiness for AI agents becomes an additional return on infrastructure you were going to need anyway — which is the argument for treating it as something you keep running rather than a project you finish.

Physical stores will continue to matter, as will brand, merchandising and human advice. What is gradually changing is the order in which a customer meets them. Before a shopper sees the store, a system may already have decided which stores deserve to be considered. The next storefront may have neither a permanent window nor a sales floor. It may exist for a few seconds inside a conversation, assembled from whatever version of the product, price, availability and service terms the retailer can substantiate at that moment.