For years, the standard online buying journey was simple: a person searched, clicked through to a website, compared options and checked out.
That journey still matters. But an increasing share of product research now begins in an AI assistant. The customer asks for “a durable carry-on under £200”, “accounting software for a small architecture practice” or “a gift for a keen cyclist”, and the assistant does much of the initial work: narrowing the field, comparing specifications, checking availability and recommending options.
That does not mean people have stopped making decisions. In many cases, the person still reviews the shortlist, visits the retailer and approves the purchase. The difference is that your website may first be assessed by software rather than read by a person.
This creates a useful distinction. Is your business built for a human-first experience, an agent-first transaction, or both?
The sensible answer for most businesses is both. But they need different things from your site.
The shift from browsing to delegated buying
AI shopping tools are not theoretical product-roadmap material any more. OpenAI says people are increasingly starting product discovery in ChatGPT, where they can compare products and refine choices conversationally. Its Agentic Commerce Protocol lets merchants share structured catalogue information, including product details, prices, stock and fulfilment data, for use in ChatGPT shopping experiences. OpenAI’s product discovery announcement is a useful indication of the direction of travel.
Shopify has made the same shift tangible for merchants. Eligible Shopify products can be shared through Shopify Catalog for discovery in AI channels, including ChatGPT. For ChatGPT purchases, Shopify says the customer currently completes the transaction in the merchant’s own checkout, opened in an in-app browser or a new tab. Shopify’s guidance on selling through ChatGPT makes an important point: agent-led discovery does not automatically remove your website from the journey.
So, do not think of this as “robots buying everything”. Think of it as delegated buying. The assistant researches and filters; the person remains the buyer, often giving final approval.
Human-first websites are built to persuade
A human-first website is designed around attention, trust and confidence. It answers the emotional and practical questions a customer has while deciding whether to buy.
That means:
- a clear brand and a credible visual identity;
- useful copy that explains benefits, not just features;
- product photography, video and demonstrations;
- reviews, testimonials and reassurance;
- navigation that makes browsing easy;
- a checkout that feels familiar, secure and low-friction.
None of this becomes less important because AI is involved. If an assistant recommends you, the eventual visitor may be more informed and closer to buying than a conventional search visitor. They will still want proof that you are a legitimate, reliable business.
The fundamentals remain the fundamentals. Before an online store launches, businesses still need to think beyond the template and ask how they will differentiate their offer, understand demand and convert visitors. Our guide to what to know before launching an online store remains sound on that point.
Human-first design is about making a good case for choosing you.
Agent-first websites are built to be understood and transacted with
An AI agent cannot infer a dependable price from a beautiful lifestyle photograph. It cannot safely guess whether a product is in stock, whether delivery is available to a particular area, or whether a discount applies.
It needs explicit, current information.
| Human-first priority | Agent-first equivalent | Why it matters |
|---|---|---|
| Persuasive product page | Structured product attributes | Lets systems compare like with like |
| Attractive price display | Machine-readable live price and currency | Prevents misleading recommendations |
| “Selling fast” messaging | Actual availability and lead-time data | Prevents unavailable items being surfaced |
| Reassuring checkout design | Stable, documented transaction path | Allows systems to pass correct order details |
For product businesses, this normally means clean data for each purchasable item or variant: title, brand, description, identifiers, images, specifications, price, currency, availability, delivery details, returns and relevant reviews.
This is not solely an AI-agent requirement. Google recommends Product and Offer structured data so it can understand product prices, availability, shipping, ratings and return information. It also recommends combining on-page structured data with a Merchant Center feed where relevant. Google’s Product structured data documentation explains why: consistent data gives platforms a better basis for displaying and verifying your offer.
OpenAI’s commerce documentation takes the same general approach. Its product feeds are designed to carry refreshed details such as identifiers, descriptions, pricing, inventory, media and fulfilment options. Its agentic commerce guide also describes an integration model in which the merchant keeps control of validation, tax, payment processing and order acceptance.
For service businesses, the equivalent may be structured services, locations, eligibility rules, opening hours, pricing ranges, appointment availability and a clear route to request or book.
Where human-first and agent-first design clash
Some conversion tactics work because they interrupt, delay or steer a human decision. Those same tactics can leave an agent unable to establish what is actually being sold or how to proceed.
Common examples include:
- a price visible only after entering an email address;
- “call for a quote” in place of any stated scope, starting price or process;
- mandatory account creation before basic product information is available;
- pop-ups that obscure pages or intercept every action;
- variant, stock or price information loaded only through fragile scripts;
- checkout steps with unclear form labels or unpredictable errors;
- CAPTCHA challenges at the wrong moment.
A CAPTCHA is not automatically wrong; fraud prevention is a real operational need. But it should not become the only viable route through authentication or checkout. The W3C’s accessibility guidance says that if CAPTCHA is used during authentication, there needs to be a route that does not rely on a cognitive test, subject to specified exceptions. WCAG 2.2 guidance on accessible authentication is relevant to people first, but the principle helps here too: predictable, properly labelled flows are easier for legitimate users, assistive technology and authorised software agents to complete.
Heavy JavaScript is another risk. Google can render JavaScript, but its own documentation notes that crawlers face differences and limitations, and that resources not considered essential may not be fetched. Google’s JavaScript search guidance is a useful reminder not to hide essential commercial information behind a complicated front end.
The practical model is an agent layer underneath and a human experience on top
You do not need to choose between a sterile machine interface and a compelling brand website.
Build the dependable commercial layer underneath:
- accurate catalogue and service data;
- public product or service pages;
- structured data and feeds;
- current price, stock and delivery status;
- clear policies;
- accessible forms and predictable checkout behaviour;
- APIs or integrations where your platform supports them.
Then build the human layer on top:
- brand;
- editorial judgement;
- useful imagery and explanations;
- comparison tools;
- reviews;
- confident, friction-light design.
The agent layer helps get you shortlisted. The human layer helps get you chosen.
This matters especially where the purchase involves taste, trust or meaningful trade-offs. An AI assistant may filter ten options to three. The customer may then decide based on design, ethos, customer service, warranty confidence or simply which business feels right.
The risk of building for only one audience
A human-only business risks becoming difficult to discover or compare when assistants are asked to do the initial research. If prices are hidden, stock is unreliable and terms are vague, an agent has little solid evidence to work with. It may recommend a competitor whose information is simply easier to verify.
An agent-only business has the opposite problem. Clean feeds and flawless data will not compensate for a weak brand, thin product explanation or an unconvincing checkout. People still need reasons to trust you.
The goal is not to automate every interaction. It is to remove avoidable ambiguity.
Five checks you can run on your own site
Start with a small audit before commissioning a major rebuild.
- Ask an assistant to compare you. Search for your category, location and key buying criteria in a mainstream AI tool. Are your products or services mentioned accurately? Are prices, availability and policies easy to find on your own site?
- Check one product page without clicking around. Can someone see the exact item, variant, price, stock status, delivery estimate and returns position on that page? If not, neither a rushed customer nor a software agent has a clear answer.
- Test your structured data. Use Google’s Rich Results Test and fix errors in product, offer, review, shipping and return markup. Do not add markup for information that is not genuinely visible or maintained on the page.
- Try the checkout with a critical eye. Use an incognito browser and a mobile device. Note forced registration, unexpected pop-ups, unclear errors, CAPTCHA interruptions and fields that are difficult to understand. If it frustrates a person, it is unlikely to help an agent-assisted journey.
- Make your policies specific. Publish delivery areas, dispatch times, return windows, exclusions, contact details and any eligibility conditions in plain language. “Terms apply” is not usable commercial information.
The next era of e-commerce will not replace customer experience with machine-readable feeds. It will reward businesses that can provide both: a site that makes sense to systems and feels right to people.
