Is the website becoming less relevant?

The answer isn’t that the website is about to disappear. Billions of consumers already understand how websites work, and much of the world’s commerce infrastructure is built around them.

But the website is likely to lose its position as the unquestioned starting point of the customer journey. It will become one surface within a broader retail environment shaped by search, retailer-owned agents, personal AI assistants, voice and dynamically generated interfaces.

That change has significant implications for where retailers invest.

Ecommerce scaled the shelf, but lost some of the salesperson

The physical store did something that ecommerce has never completely replicated.

A good salesperson could interpret an imprecise request, ask questions, understand the context and help a customer feel confident about a decision. They didn’t simply retrieve product information. They used judgement.

Ecommerce gave retailers reach, convenience and efficiency. Product pages, search functions, filters, recommendation engines and shopping carts allowed customers to browse and buy without the assistance of another person.

But in digitising the store, retail also removed much of the human touch.

Websites are usually designed around pages and categories rather than the customer’s complete intention. Even with personalisation, the experience remains largely fixed. The shopper has to know where to look, which filters to select and how to translate a need into the language of the catalogue.

Agentic AI offers the possibility of restoring some of what was lost.

Rather than requiring customers to navigate a retailer’s internal product structure, an AI sales agent can begin with the customer’s problem. It can ask questions, interpret preferences, compare suitable products and explain the differences.

That’s the promise. Building it has taught us why the transition will be more complicated than the current enthusiasm for agentic commerce suggests.

Our experiment building the retail experience beyond the website

When we began developing Vendio, our ambition was not to add another chatbot to a website.

We wanted to recreate the experience of dealing with a retailer’s best salesperson: someone with strong product knowledge who could listen, advise and dynamically adjust the experience around each customer.

Our original concept used voice, speech and a generative interface that changed throughout the interaction. Instead of moving through predetermined pages, the customer would see information, comparisons and actions generated according to the conversation.

It was an attempt to imagine the retail experience beyond the traditional website.

The early demonstrations were compelling. At their best, the models asked exactly the right questions, made unexpected connections and produced recommendations that felt genuinely useful.

But an AI sales experience can’t be judged only by what it does at its best.

The next interaction might be slower. The agent might ask the wrong question, misunderstand the shopper or offer to perform an action it couldn’t complete. An experience that was brilliant some of the time wasn’t dependable enough for a customer-facing retail environment.

We found that bringing agentic AI into production required us to answer three different questions:

1. What is technically possible?

2. What is commercially realistic?

3. And what will customers actually use?

The important lesson was that progress against one question didn’t guarantee progress against the others. An experience could be technically impressive but too slow, expensive or inconsistent to operate commercially. It could be commercially viable but still ask customers to behave in ways that felt unfamiliar or frustrating.

The challenge wasn’t simply to build something AI could do. It was to find the point where technical capability, commercial reality and customer behaviour aligned.

1. What is technically possible?

Our early work showed that AI could create a far more responsive retail experience than the conventional website allowed.

The agent could ask questions, interpret an imprecise need, identify relevant products and generate parts of the interface around the conversation. At its best, it made connections we hadn’t explicitly designed, and it produced recommendations that felt genuinely useful.

Voice and generative interfaces also offered a glimpse of a retail experience that wasn’t organised around fixed pages, categories and navigation paths. Instead of requiring shoppers to translate their needs into the structure of the catalogue, the experience could adapt around what they were trying to achieve.

There were moments when the technology appeared to deliver the digital equivalent of a highly capable salesperson.

But proving that an interaction was possible was only the first test. A customer-facing product couldn’t depend on the model performing at its best occasionally. It had to produce a reliable experience repeatedly.

That led to the second question.

2. What is commercially realistic?

The first commercial constraints were speed and cost.

The models that handled more complex tasks well were generally slower and more expensive. Faster, less costly models were more likely to misunderstand the customer, ask weaker questions or make mistakes.

Customers may tolerate a brief delay when the answer is valuable, but an AI-assisted journey can’t feel slower or more difficult than browsing the website. The quality of the recommendation needs to justify the time it takes to produce.

Reliability presented an even greater challenge.

An early version of the agent might conduct one exceptional interaction, but then approach the next customer differently. It could ask the wrong question, overlook an important requirement or offer to perform an action it couldn’t complete.

That inconsistency matters in retail. The agent is speaking on behalf of the brand. A plausible but incorrect statement about a product, price, feature or policy can quickly damage trust.

It was therefore not enough to connect a general-purpose language model to a product catalogue. The experience needed to prioritise authoritative brand information, including current products, specifications, pricing and policies, over the model’s general knowledge.

A commercially viable retail agent must be useful, fast and affordable. It must also be dependable enough to represent the retailer consistently because every inaccurate answer or failed interaction risks eroding the customer trust the brand has worked hard to build.

3. What will customers actually use?

Even when the technology worked, and the economics were manageable, there was another test: did the experience fit the way customers wanted to shop?

Voice illustrated the problem.

Once an interface begins speaking, customers expect it to behave like a natural conversation. They expect it to understand interruptions, background noise, accents, mispronunciations, changes of topic and incomplete thoughts.

When it couldn’t, the experience moved quickly from impressive to frustrating.

Our original interface also removed too many familiar elements of the website. We had imagined that the agent could generate the entire experience dynamically, without relying on established product pages, navigation patterns or conventional controls.

Technically, that was possible. From the customer’s perspective, however, it created unnecessary work.

Shoppers already understand how to browse products, select options, type a query and use buttons. Removing those conventions meant they had to learn a new interface at the same time as they were deciding whether to trust an AI agent.

We were asking customers to change too much, too quickly.

The lesson was not that shoppers rejected AI assistance. It was that they wanted the assistance to fit naturally into familiar behaviour. They needed the freedom to type, touch, browse or use voice according to the situation.

Generative interfaces were most effective when they appeared selectively to improve the decision, not when they replaced the entire shopping experience simply because the technology made that possible.

Where the three answers led us

The three questions forced us to narrow the gap between the future we could imagine and the experience we could responsibly deliver.

What was technically possible wasn’t always commercially sustainable. What was commercially achievable wasn’t always something customers found natural to use. And a compelling demonstration was not the same as a reliable retail product.

The answer was to retain the ambition while changing the route.

We introduced a more familiar conversational interface. We allowed shoppers to type when the suggested choices didn’t reflect what they needed. We used generative UI where it added clear value, rather than asking it to replace the entire website. We began reintroducing voice as an option instead of making it the primary way to interact.

Most importantly, we concentrated on helping the customer make a better decision rather than trying to make the decision for them.

The near-term opportunity is assisted decision-making

That distinction between helping customers decide and deciding on their behalf matters.

Much of the excitement around agentic commerce assumes that consumers will soon delegate significant purchasing authority to personal AI agents. An agent might anticipate a need, select the product and complete the transaction with little involvement from the customer. That future is possible. However, it’s not yet the most credible starting point for most retailers.

Purchasing decisions often involve trade-offs that customers haven't fully articulated. In higher-value or more considered categories, people want to understand why one option is more suitable than another. They may welcome assistance without surrendering control.

The more immediate role for retail AI is therefore not autonomous purchasing. It’s decision support.

An effective retail agent can:

• understand intent that doesn’t fit neatly into search keywords

• ask questions that reveal the customer’s real priorities

• recommend and compare suitable products

• explain differences in language the customer understands

• surface reliable, current information from the retailer

• help customers move forward without taking control away from them

This isn’t simply a more conversational version of site search. It changes the organising principle of the experience from “navigate our catalogue” to “tell us what you’re trying to achieve”.

What happens when every customer has an agent?

The longer-term change may be more profound.

Today, retailers design journeys on the assumption that the customer will enter an environment the retailer controls. In the future, the customer may arrive with an agent that already knows their preferences, budget, past purchases and current objective.

That agent may research products across multiple retailers before the customer visits any of them. It may request information, test availability, or compare offers. The retailer’s website may no longer be the first place where consideration occurs.

Several futures could develop simultaneously.

Retailers may operate their own agents within their websites and apps. Customers may use personal assistants as their primary interface to multiple brands. Retailers may supply specialist product interfaces and actions that appear inside third-party agent environments. Websites may evolve into richer brand destinations for customers who want exploration, inspiration or reassurance rather than simple information retrieval.

Voice, visual interaction, conventional browsing and autonomous actions will each take different portions of the journey. Their importance will vary by customer, category and occasion.

The likely future is not one interface replacing every other interface.

It’s a fragmented environment in which the retailer’s website is demoted from the front door to one of several doors.

Retailers should prepare the business, not bet on a single interface

Retail leaders don’t need to predict exactly which agent, protocol or platform will dominate.

They do need to prepare their organisations to participate in a retail environment where customers and machines both need to understand their products.

That begins with the foundations.

Can your product information be accessed beyond the pages on which it currently appears? Is it complete, current and structured consistently? Can an agent reliably distinguish between authoritative product data and generic information? Can your systems expose availability, pricing, service policies and useful actions without compromising security or control?

Retailers should also identify the customer problems that justify an agentic experience.

Where do shoppers struggle to choose? Which categories require explanation? What questions repeatedly reach customer service teams? Where does conventional search fail because customers describe a problem rather than a product? Where does uncertainty cause abandonment?

Start there.

Point AI at a real customer or operational problem. Test whether it can improve the outcome. Measure accuracy, response quality, customer behaviour, cost and speed. Keep people in control while the system earns greater responsibility.

The goal isn’t to purchase an ‘AI agent’ and declare the business ready. It’s to develop the organisational ability to test, learn and adapt as the technology changes.

Our experience building Vendio reinforced a simple reality: creating a convincing AI demonstration can happen relatively quickly. Making the experience reliable enough for production requires far more work.

That is where retailers can separate genuine AI capability from hype.

The website Isn’t dead. Its monopoly is.

The website will remain essential for many years. It will continue to carry brand content, product information, customer accounts, transactions and service experiences.

But it will no longer be safe to assume that every meaningful customer journey begins there, or that the website alone should carry the responsibility for digital selling.

Retailers should prepare for customers to move between personal agents, retailer agents and conventional digital experiences. They should expect discovery, evaluation and purchase to separate across different surfaces. And they should design their data, systems and operating models so that trusted retail knowledge can travel with the customer.

The winners won’t necessarily be the retailers that predict the future interface correctly.

It will be the ones that can keep adapting as the front door moves again.

This editorial was adapted from the Tech Talk conducted by Vervio at the Online Retail Exhibition on 23 July 2026