Agentic web design: what to change when agents do the buying
Gartner puts 90% of B2B buying through AI agents by 2028, over $15 trillion. Here is what changes in a product's data, feed and checkout.
In this piece
The agentic web is the part of the internet where software reads pages, compares vendors, and completes a purchase on behalf of the person who is buying. It already has a price attached. Gartner projects that by 2028, 90% of B2B buying will be intermediated by AI agents, moving more than $15 trillion through agent exchanges, a forecast presented at its IT Symposium in October 2025.
Most teams file this under marketing and hand it to whoever owns SEO. It lands on the product team. An agent that cannot read your price, confirm availability, or start a checkout without a browser session drops you from the shortlist, and the landing page a human would have liked never gets opened.
What fails first when the reader is a machine
Agents parse. They do not scroll, hover, or wait for a modal to close. Four things go wrong, roughly in this order.
Facts that live only in the layout. A price painted by client-side JavaScript after a currency lookup is visible to a person and absent to a parser. Same for availability badges, plan limits, and anything a third-party widget draws after hydration.
Facts that disagree with each other. The page says 49 euro, the feed says 59, the API still carries last quarter’s plan name. Adyen’s guidance for retailers is blunt on this point: consistency between the product feed and the site is a condition for being included in agentic shopping at all. A parser that finds two prices has no way to break the tie, so it drops the record or drops the vendor.
Checkout that assumes a human session. Cookie-bound carts, a CAPTCHA on the plan selector, a payment step that exists only inside your own interface. If a transaction cannot be created over an API with a signed authorization, an agent can research you thoroughly and still not buy.
Facts you hide on purpose. "Contact sales for pricing" is a legitimate position with a new cost. An agent asked to compare five vendors fills four price cells and leaves yours empty. Whether that trade is worth it is a commercial call, not a technical one, but it is now a call you are making rather than a default you inherited.
Volume on this channel is still small and the intent is dense. Ahrefs measured its own traffic in June 2025 and reported that AI-referred visitors were 0.5% of sessions and 12.1% of signups over 30 days, around 23 times the conversion rate of organic search. That is one company reading its own analytics, so treat it as a signal and not a benchmark. The mechanism behind it is easy to state: the agent already did the comparison, so the visit that survives it arrives qualified.
Why schema markup alone does not finish the job
Adding Product and Organization JSON-LD is the first move and the cheapest one. It is also where most teams stop. Markup describes. It does not hold inventory, it does not transact, and it does not carry an authorization that a payment provider will accept.
Three specifications are forming above it, and they do different jobs. ACP, the Agentic Commerce Protocol from OpenAI and Stripe, standardises the checkout exchange between an agent and a merchant, and ships a product feed specification alongside it. AP2, the Agent Payments Protocol, standardises authorization: who approved what, expressed as cryptographically signed mandates. Google donated AP2 to the FIDO Alliance on 28 April 2026, together with Mastercard’s Verifiable Intent framework, which moved it from a single vendor’s spec to community governance and added Human Not Present payments in v0.2. UCP, Google’s Universal Commerce Protocol, extends Merchant Center into cart, checkout, and post-purchase.
A complete agentic purchase generally needs two of them: one to transact, one to authorize. No amount of structured data covers that second half.
What to change in the build
Publish every commercial fact twice, from one source
Price, plan limits, availability, refund terms, and contract minimums each need a rendered version for people and a machine-readable version for parsers. Generate both from the same database record in the same request. When a designer edits copy on the pricing page and the JSON-LD comes from a separate CMS field, the two drift inside a quarter and nobody notices until an agent quotes the stale one. We build the structured data from the query that renders the page, so a price change stays one write.
Ship a feed, not only pages
Crawling is slow and lossy. A feed is a direct answer. The ACP feed spec defines the fields an agent expects, and a Merchant Center feed does the equivalent job on the Google side. Regenerate on write rather than nightly. Stale availability is worse than missing availability, because it produces a failed transaction instead of a skipped one.
For a SaaS with five plans this is a small file and a webhook. For a catalogue it is real engineering, and it is the work that decides whether you show up at all.
Decide a protocol posture before a platform decides for you
You do not have to adopt ACP, AP2, and UCP this quarter. The specifications are still moving and the platforms have already reversed course on parts of them. What is worth doing now is cheap and survives the churn: keep cart and checkout creation behind an API that does not require your own interface. Anything the front end can do to a cart, an authenticated API call should be able to do as well. Adding a protocol later then becomes an adapter instead of a rebuild. The same discipline pays off for WebMCP and the agent-facing browser APIs.
Keep the human path first-class
The agent builds the shortlist. A person still signs. Gartner’s May 2026 survey of B2B buyers found that 69% go to a sales representative to validate AI-generated insights before they decide. That has a direct design consequence. The page an agent cites has to hold up when a sceptical human opens it two minutes later, which rules out thin pages written for machines. The second read is where the contract is.
This is also the practical answer to teams asking whether to write for agents or for buyers. The same page does both jobs, and the constraint that binds is the human one, because it is stricter. Our notes on building interfaces a non-human user can read go deeper on the interface side of that.
Measure agent traffic as its own channel
Referrals from ChatGPT, Perplexity, Claude, and Gemini arrive in analytics as ordinary referrals and vanish into the aggregate. Split them at the referrer, log agent user agents separately at the edge, and carry the source through to the CRM. Until that is wired, every argument about whether the work paid for itself is anecdote. The difference between GEO and AEO matters here, because the two produce different numbers and teams routinely report one while measuring the other.
What this looks like on a real build
Take a B2B SaaS with three plans, usage-based overages, and a self-serve trial. The first agentic pass is about three days of work.
Day one: move pricing out of the page component into a table, render both the page and the Product plus Offer JSON-LD from it, and express plan limits as structured properties rather than prose in a feature list. Day two: expose a read-only plans endpoint and generate a product feed from the same table on write. Day three: confirm that robots.txt does not block GPTBot, ClaudeBot, PerplexityBot, and Google-Extended, add an llms.txt that points at the plans endpoint and the docs, and split agent referrers in analytics.
What we would leave out of that first pass: AP2 mandates and any rebuild of checkout for ACP. Those earn a spike when a real agent-sourced transaction shows up in the logs, and not before. The three days above pay off whichever protocol wins, because all of them read the same underlying facts.
One limit is worth stating plainly. None of this creates demand. If an agent compares four vendors and yours is the expensive one with the shorter integration list, clean structured data gets you accurately excluded faster. The work makes a product legible to the machine doing the comparison. It does not make it the answer.
Sources
- Gartner Unveils Top Predictions for IT Organizations and Users in 2026 and Beyond
- Gartner Survey Finds 69% of B2B Buyers Turn to Sales Reps to Validate AI-Generated Insights
- Ahrefs: 0.5% of visitors drove 12.1% of signups
- Google: donating the Agent Payments Protocol to the FIDO Alliance
- FIDO Alliance: building the trust layer for agentic payments with AP2 and Verifiable Intent
- Agentic Commerce Protocol: product feed specification
- Adyen: agentic commerce and product feeds, a guide for retailers
Frequently asked questions
Do I need to implement ACP or AP2 to show up in AI shopping and research answers?+
No. Appearing and transacting are separate stages. To appear, an agent needs to read consistent facts from your pages and, ideally, a product feed, plus crawl access for the AI user agents. ACP and AP2 only matter at the moment an agent tries to complete a purchase with no human in the browser. For most B2B SaaS that moment has not arrived yet, and the readable-facts work is what pays today.
What does a first agentic pass cost for a small SaaS?+
Two to five engineering days for a product with a handful of plans and a single currency. The variable is where pricing lives today. If it is already one record in the database, the work is rendering it twice and adding a feed. If prices are hardcoded in page components, in a CMS, and in Stripe with no single source, expect the high end, because the real task is consolidating them first. A catalogue business with thousands of SKUs is a different project.
If I block AI crawlers to protect my content, can agents still buy from me?+
In practice, no. The same user agents that read your pages for training and for answers are the ones that read them for comparison. Blocking GPTBot or ClaudeBot at the edge removes you from the shortlist an agent builds before any transaction is considered. The workable position is selective: allow crawl access to commercial pages, pricing, and docs, and restrict what you genuinely do not want reproduced, such as gated research or customer data.
Does agentic web work replace SEO?+
It sits on top of it. Agents still reach most sites through search indexes, so crawlability, speed, and internal linking remain the entry condition. What changes is what happens after the page is fetched: a human skims and a parser extracts, and only the second one fails silently when a fact is missing. Teams that already do technical SEO well are usually two or three days away from being agent-readable.
Related services
Studio
Start a project.
We write about what we build. Tell us what you want to build.