A tactical guide to making B2B websites usable by AI browser agents: ARIA labels, open pricing, WebMCP, and agent testing.
An agent-ready website is a site an AI browser agent can read, navigate, and act on without a human taking over. For B2B teams, that means an agent sent by a buyer can find your pricing, compare your plans, and reach your demo form in one session. This guide shows you where agents fail on B2B sites today, which fixes recover the most completed tasks, and how to test your own pages in an afternoon.
An agent-ready website exposes its content and actions in a form that AI browser agents can parse and use: labeled buttons, open pricing, non-blocking consent banners, and forms an agent can submit.
AI search optimization so far has focused on retrieval. You want ChatGPT, Perplexity, or Google AI Overviews to quote your page. Agentic browsing adds a second job. A buyer types "compare the Team plans of these three vendors and book a demo with the cheapest one," and an agent opens your site, clicks, scrolls, and fills fields on the buyer's behalf.
Retrieval readiness and agent readiness overlap at the HTML layer. Both need server-rendered text and crawl access. Agent readiness then adds interaction: can the agent dismiss your cookie banner, find the "Pricing" link, open the plan comparison, and submit the form without hitting a wall?
If you already fixed rendering for crawlers (see our guide to JavaScript rendering for AI crawlers), you have the foundation. The steps below cover the interaction layer on top of it.
Agent traffic is small in absolute terms and growing faster than any other traffic class. HUMAN Security's 2026 State of AI Traffic report (published March 26, 2026) measured 7,851% year-over-year growth in agentic traffic during 2025.
The same report found overall AI-driven traffic grew 187% from January to December 2025, and that 2.3% of agentic activity already lands on checkout pages. Agents already reach the pages where money changes hands.
Readiness lags behind that growth. ora's April 2026 State of Agent Readiness scan ran real agent sessions against thousands of products and scored each site on 110 checks. The median score was 44 out of 100, and about 73% of products graded D or F. ora sells the scanner, so treat the grading scale with care, but the access numbers are direct observations: ChatGPT-User reached 81% of scanned sites and ClaudeBot reached 82%.
For a B2B marketer, the read is this: one in five sites already blocks a major agent at the door, and most of the rest let agents in without helping them finish anything.
Most AI browser agents read a page through the accessibility tree plus screenshots, and they act on elements by role and label. OpenAI's Publishers and Developers FAQ states that ChatGPT Atlas "uses ARIA tags," the same labels and roles that screen readers use.
The accessibility tree is a stripped-down copy of the DOM. It keeps headings, text, links, buttons, and form fields, and it drops styling. Sitebulb's June 23, 2026 guide calls it "an SEO surface" for this reason. Whatever your tree says about a button is what the agent believes about that button.
Three consequences follow for your site:
Put your key links and buttons in the first two screens, and give each one a text label.
AI agents fail on B2B websites at three gates: consent modals, bot challenges, and login or email walls in front of pricing. MaxAEO's July 2026 test of 42 B2B SaaS websites traced 76% of failed agent runs to those three causes.
The test ran one standard task across six agent surfaces, 252 runs in total, between May 12 and June 20, 2026. Agents completed 154 runs (61%). Of the 98 failures:
Academic benchmarks show the same pattern. ClawBench, a 2026 benchmark from researchers at the University of British Columbia and the Vector Institute, ran 153 everyday tasks on 144 live platforms. The best model, Claude Sonnet 4.6, finished 33.3% of tasks. In a sample of 26 of its failures, anti-bot walls caused 8, early stops before submission caused 7, and login or signup walls caused 5.
The worst finding sits in the MaxAEO data. In 31 of the 98 failed runs, the agent answered the user anyway without reporting the failure. In 9 of those, it quoted a competitor's pricing as the target brand's price.
Make your consent banner non-blocking, and scope bot challenges to abuse-prone endpoints. These two changes address over half of agent failures in the MaxAEO dataset.
In that test, sites whose consent banner locked page scroll until dismissed completed 38% of agent runs. Sites with non-blocking banners completed 79%. Region mattered too: on EU IP addresses, where consent dialogs fire by default, completion dropped to 52%, against 70% on US addresses. MaxAEO also reports that iframe-hosted consent platforms failed about twice as often as banners rendered in the same document.
Your fix list:
Monitor your progress with Nobori: track whether AI answers about your pricing start citing your own page after the fix.
Label every interactive element with a visible text label and a correct ARIA role and state. OpenAI tells site owners to follow "WAI-ARIA best practices by adding descriptive roles, labels, and states to interactive elements."
Start with the demo request form, the pricing toggle, and the main navigation. Kick Point's Atlas ARIA guide (Dana DiTomaso, updated December 2025) lists the patterns that matter most:
Use less ARIA where native HTML works. The WebAIM Million analysis has found that home pages using ARIA average about twice as many detected accessibility errors as pages without it. A native <button> beats a div with role="button" every time.
Keep forms short. ClawBench logged early stops before submission as the second-largest failure cause. Every extra field gives the agent another chance to quit or invent a value.
Yes, publish at least starting prices and plan limits on an open HTML page. Gated pricing caused 19% of agent failures in the MaxAEO test, and agents that hit the gate sometimes filled the gap with a competitor's numbers.
A buyer's agent compares vendors by reading pricing pages side by side. If your page asks for an email first, the agent has three options: give up, guess, or quote a third-party source. None of those outcomes puts your real number in front of the buyer.
You don't need to publish enterprise rates. A page that works for agents includes:
If legal or sales leadership blocks public numbers, publish the pricing model in words (per seat, per tracked prompt, annual only) and the plan limits. That still gives an agent something true to report.
Monitor your progress with Nobori: check how ChatGPT, Gemini, and Perplexity describe your pricing before and after the page goes live.
WebMCP is a proposed browser standard that lets a page declare its actions as tools an AI agent can call directly. Google announced it at I/O 2026, and it entered origin trials in Chrome 149, according to InfoQ's June 13, 2026 report.
The spec has two surfaces. The declarative API adds attributes to existing HTML forms: toolname, tooldescription, and toolautosubmit. The imperative API registers tools in JavaScript through navigator.modelContext, each with a name, a description, and an input schema. Limits are tight: 500 characters per tool description, 150 per parameter description, and 30 per name.
Adoption is near zero. ora's April 2026 scan found WebMCP on 1% of products, against 34% with any MCP endpoint and 48% with llms.txt.
My read for B2B teams: annotate your demo form with the declarative attributes once your ARIA labels and consent banner are fixed. It takes an hour and costs nothing if no agent uses it. Skip the imperative API until a production browser ships it outside a trial. As with llms.txt for AEO, adoption alone proves nothing about results.
Verify agents through cryptographic request signatures where they exist, and through user-agent strings and referral tags where they don't. Cloudflare launched signed agents in August 2025, with ChatGPT agent in the first cohort.
Signed agents use Web Bot Auth, which attaches HTTP message signatures to each request so your edge can confirm the agent's identity. Cloudflare lists ChatGPT agent, Goose from Block, Browserbase, and Anchor Browser among the first signers, and Enterprise customers can write security rules for signed agents as a group. A verified signature lets you allow a buyer's agent through while still challenging unknown scrapers.
Three checks to run this week:
Compare those paths with your money pages. If agents request pricing but never reach the demo form, you've found your drop-off point.
Give two or three AI browser agents the same buyer task on your site, record where each one stops, and repeat from an EU and a US connection. MaxAEO's data shows a median abandonment after just two interaction attempts, so failures surface fast.
A one-afternoon test protocol:
Expect agents to work hard on a single page. Tory Gray told Sitebulb she watched one client's agent attempt about 2,000 actions in a ten-minute window. All that clicking still stalls on a button with no label.
Ship the fixes in order of failure share: consent banner, bot challenge scope, pricing gate, then labels and forms. That sequence follows the 76% of failures MaxAEO traced to the first three causes.
Put a quarterly re-test on the calendar. Consent tools, CMS themes, and security rules change without warning, and any one of them can reintroduce a blocking modal.
Agent readiness and AI visibility share an outcome you can measure: what AI engines say about your brand when a buyer asks. Nobori tracks your brand across ChatGPT, Google AI Overviews, Perplexity, Gemini, and Claude, so you can see whether answers about your pricing, plans, and integrations match your site after each fix.
Use Nobori to build a prompt set from your three agent test tasks, record which sources each engine cites for your pricing today, and spot answers that quote competitor numbers or stale plan limits. When you ship an open pricing page or a non-blocking banner, Nobori shows whether your own URL starts appearing as the cited source.
For deeper context on the crawler side, read our breakdown of crawl-to-refer ratios and agentic traffic.
See if AI engines are citing you → nobori.ai
