Jake Young
Every website built in the last twenty-five years assumed the visitor had eyes. A growing share of first visits now come from software that reads your prices, checks your stock and reports back to a person who may never see your homepage. Our view is that agents won’t kill the website. They’ll split it in two: a data surface that machines judge on facts, and a brand surface that has to win the human before the agent is ever sent out.
Brands that win this shift will be legible to machines and memorable to people, and the second half matters more than most agent readiness advice admits. Here is what’s real, what isn’t yet, and how to get ready for both audiences.
What’s Live Right Now
The plumbing shipped in the last twelve months. In September 2025 OpenAI and Stripe released the Agentic Commerce Protocol, an open standard for an agent to place an order with a merchant. In January 2026 Google announced the Universal Commerce Protocol, developed with Shopify, Etsy, Wayfair, Target and Walmart. Both card networks built rails for agents to carry tokenised credentials, and the Model Context Protocol, which lets agents call tools and data, moved to the Linux Foundation’s Agentic AI Foundation in December 2025.
Traffic followed. Shopify says AI-driven visits to its stores grew eight times year over year in the first quarter of 2026, with orders from AI searches up nearly thirteen times. Adobe, measuring more than a trillion visits to US retail sites, found AI-referred traffic up 138% year over year in May 2026, converting 54% better than other traffic. A year earlier it converted at roughly half the rate. That reversal matters more than the growth curve. The assistants got better at sending people who were ready to buy.
Plenty didn’t stick. OpenAI pulled Instant Checkout, the feature for buying inside a ChatGPT answer, back into merchant apps in March 2026, after Shopify pointed to how much checkout has to handle. In July OpenAI said it was retiring its Atlas browser. And the legal ground is moving: a court blocked Perplexity’s Comet agent from shopping on Amazon in March, and the Ninth Circuit overturned that injunction in August.
What’s Still Speculation
People are not handing over their wallets. Gartner’s May 2026 survey found only 11% of consumers willing to let AI make purchase decisions, even in low-stakes categories like household supplies. More were happy to let it narrow the options. Of those who had used AI for a recent purchase, 54% said they had to double-check everything it told them. McKinsey’s widely quoted figure, up to $1 trillion of US retail revenue orchestrated by agents by 2030, is a scenario, and we’d treat it as one.
Our prediction for the next two to three years: agents do most of their work before the purchase. They research and shortlist, and a person still presses buy, often on your site. Fully delegated buying arrives first in replenishment, where the brand already won the first order. The first purchase is where brand persuades, and the repeat purchase is where legibility keeps you in the basket.
When the First Visitor Isn’t Human
Take a common errand. Someone asks an assistant for a carry-on under a certain weight that fits their airline’s limits and arrives by Friday. The agent fetches pages, reads structured data, perhaps queries a catalog through a protocol, and returns three options. Your brand film, your scroll animation and your hero photography did not load. What the agent saw was your weight in grams, if you published it, plus dimensions, stock, delivery promise and returns.
Researchers have started measuring how often that errand fails. In WebMall, a benchmark accepted at SIGIR 2026, the best web agents completed fewer than 65% of tasks such as finding the cheapest product across four shops. A July 2026 preprint by Elnaffar and Rashidi ran three models against the same shop built two ways, and agents succeeded 89.3% of the time on the agent ready version against 49.3% on the conventional one.
Much of the failure is invisible content. Vercel found that the major AI crawlers from OpenAI and Anthropic fetch JavaScript files without executing them, so a price that appears only after a script runs doesn’t exist for them.
The Legibility Ladder
We audit agent readiness with a four-rung model. Each rung depends on the one below, and most brands stall on the second.
1. Reachable: can an agent get in?
Start with your CDN and bot rules, because many sites now block AI traffic by default without the marketing team knowing. Make three separate decisions: search crawlers that cite you, user-directed agents acting for a shopper, and training crawlers. Cloudflare now verifies signed agents cryptographically through Web Bot Auth, so you can allow a shopper’s agent while still challenging anonymous scrapers. Then check that prices, stock and specifications arrive in the server-rendered HTML. The trade-off is scraping: anything a good agent can read, a competitor’s price bot can read too. We accept that for public offer data and keep account-level pricing behind authentication.
2. Readable: can it understand the offer?
Agents answer attribute questions, so publish attributes. For products that means schema.org Product with an Offer carrying price, priceCurrency and availability, plus gtin or sku, brand, OfferShippingDetails for delivery windows and MerchantReturnPolicy for the returns window. For services, use Service with provider, areaServed and an offer that describes how pricing works, even when the number itself lives in a quote. Google’s Shopping Graph refreshes more than 2 billion listings an hour, and in January Google added dozens of Merchant Center attributes for conversational discovery, such as product Q&As, compatible accessories and substitutes. That list is a good map of the questions agents get asked.
On llms.txt, our position: ship one if it takes an afternoon, generate it from the same source as the site, and don’t count it as progress. A hand-written summary that drifts out of date becomes one more contradictory source.
3. Reliable: does your data agree with itself?
An agent that finds one price on your page, another in your feed and a third on a marketplace has a reason to recommend someone else. The fix is architectural. Page copy, JSON-LD and feeds should all render from one product source, so they can’t diverge, and Google’s structured data policies already require markup to match what’s visible. Put a date on anything that changes. Link certifications and warranties to their terms. Reliability is where brand trust becomes something a machine can check.
4. Transactable: can it finish the job?
Design checkout and booking so an agent can complete them and a person can approve them: guest checkout, stable URLs for every variant, real labels and autocomplete attributes on fields, errors written as text next to the field, and no action that only exists on hover. Keep a human confirmation step for payment and anything irreversible. Turn on what your commerce platform already supports, since Shopify now serves catalog and cart tools to agents through UCP. For software, expose read only tools first, such as search, availability and quotes, then put write actions behind OAuth, rate limits and idempotency keys so a retried request can’t book twice.
The ladder has a trap at the top. A brand that is legible and nothing else becomes a row in a comparison table, and rows compete on price.
The Website Splits in Two
When agents handle the comparing, the website takes on two jobs that pull in opposite directions. One is a data surface: fast, plain, exhaustive. Spec pages, policies, feeds and endpoints, boring in the best sense. The other is a conviction surface, the place a person visits before sending the agent, when deciding which brands deserve a shortlist, and after, when checking the choice before paying.
Agents shortlist by attribute. People shortlist by name. “Find me running shoes for flat feet” produces a filtered list. “Reorder my usual Hokas in a 10” produces a sale. Brands requested by name skip the comparison, and that preference is still built through a clear idea and an experience worth remembering. Adobe’s data points the same way: AI-referred shoppers spent 53% more time on retailer sites than other visitors. The people arriving after an agent are late in the decision and paying attention.
So the design brief changes. Less of the homepage should explain specifications, because the agent already has them. More of it should carry what an agent can’t bring back in a table: a point of view, proof, taste and the feeling of dealing with people who care.
The Next Two to Five Years
These are predictions. Comparison pages lose value, because agents will run the comparison from everyone’s data; primary evidence such as tests and published methodology gains it. Policies become marketing: return windows, delivery promises and warranties are filters an agent applies, and a generous, clearly published policy will win shortlists the way a hero image used to win attention.
Agent traffic becomes a reported channel. Most analytics setups filter automated visits as noise. We expect brands to report agent sessions, agent-completed tasks and agent-assisted revenue, the way Shopify already attributes orders by AI channel. And branded demand becomes the leading indicator, since the number of people asking for you by name, in search and in assistants, predicts whether agents bring you customers.
What to Do This Quarter
Run an agent errand test. Give ChatGPT, Gemini, Perplexity and Copilot five realistic tasks in your category, phrased the way customers phrase them, and record where each one fails on your site. Repeat it monthly, because the results move as fast as the models do.
Audit your bot rules with engineering and marketing in the same room, then fetch your key pages without JavaScript. If price, stock or specifications disappear, move them into the server rendered HTML first.
Generate structured data from your product source of truth, including shipping and returns, and clear the path to buy or book: guest checkout, labelled fields and a human confirmation step for payment.
Give the homepage back to the brand. Move specifications to where agents and diligent humans look for them, and spend the prime space on the idea you want people to ask for by name.
We design and build websites for both audiences, the machine that compares and the person who decides. If you’d like to see how your brand looks to an agent today, and what it would take to become the name people ask for, get in touch. We’ll start with the errand test.
