A growing share of B2B research is no longer done by a person clicking through your site. It is done by an AI agent acting on a buyer’s instructions: reading your pricing page, shortlisting you against competitors, filling in your demo-request form — and, in Google’s case, literally phoning your business to ask about pricing and availability. If your funnel silently breaks when the visitor is an agent, you don’t get an error report. You just get skipped.

This guide covers the site and funnel layer of agent readiness: whether an agent can actually traverse your forms and wizards, what machine-readable entry points you should expose, what agent traffic looks like in your server logs, and an honest maturity checklist. It is the sibling of our guide to writing for agentic parsing — that post covers how to structure the content agents read; this one covers whether agents can operate your site at all. Read both together.

Apply For Partnership

The buyer’s agent is already here — and it’s verified, not hype

Two developments make this concrete rather than speculative.

First, OpenAI’s ChatGPT agent. Its system card (July 2025) describes a model that combines deep research with “Operator’s capacity to execute tasks through a remote visual browser environment” — in plain terms, it operates a real browser: it navigates pages, clicks buttons and fills forms the way a human would, with user confirmations and a supervised “watch mode” applied in sensitive contexts. It does not need your API, your chatbot or your goodwill. It needs your site to work.

Second, Google now calls businesses on users’ behalf. In July 2025, Google announced that Search can call businesses to gather pricing and availability for a user — rolling out to US Search users, with higher limits for paid AI subscribers — and its later shopping guidance describes the same capability as an agentic calling feature that rings local businesses to check what’s available. Google notes businesses control participation through their Business Profile settings. The “buyer” on your phone line may already be software.

Be honest about the limits, too: these agents are early, sometimes slow, and are deliberately interrupted for confirmation before consequential actions. But the direction is unambiguous, and the sites that fail agents today are training tomorrow’s agents (and their users) to route around them.

Layer 1: Can an agent physically traverse your funnel?

Forms: labels, autocomplete attributes and error messages

Agents driving a browser parse your form the way assistive technology does — from the DOM. Everything that makes a form accessible makes it agent-operable:

• Real <label> elements bound to inputs, not placeholder text that vanishes on focus. A field whose only identity is a disappeared placeholder is a guessing game.
• Standard autocomplete attributes. The HTML autocomplete attribute exists precisely to let user agents provide automated assistance filling in form values — tokens like email, tel, given-name and organization declare unambiguously what each field wants. That declaration was designed for browsers autofilling; it is now a gift to any agent completing your form.
• Error messages agents can read. A red border with no text is invisible to a DOM-reading agent. Validation errors should be text, adjacent to the field, stating what is wrong and what format is expected. If your form fails silently, the agent abandons — and so does the buyer who sent it.

Multi-step wizards

Multi-step qualification wizards are where agents (and humans) quietly die. Common failure modes: steps that lose state on back-navigation, progress that depends on hover interactions, timeout-based session resets mid-flow, and “next” buttons that stay disabled with no explanation of what is missing. Every one of those is a place an agent stalls, burns its task budget and reports failure to its user. Keep wizards short, keep every requirement stated in text, and offer a low-friction fallback — a plain contact form or a phone number — for any flow longer than a few steps. This is the same logic behind speed to lead: every second and every step between intent and contact leaks pipeline, whoever (or whatever) is doing the clicking.

CAPTCHAs and anti-bot friction: the real trade-off

Here is the uncomfortable tension. In July 2025 Cloudflare announced it was changing the default to block AI crawlers unless they pay creators for content — a defensible stance against bulk scraping. But the same class of defences, applied bluntly, also blocks the on-demand agent your prospect just asked to research vendors. Blocking a training crawler costs you little pipeline; blocking a buyer’s agent at your demo form costs you the buyer.

There is no universally honest answer on CAPTCHAs. Agents operate real browsers, so crude bot checks may not stop them, while aggressive challenges stall legitimate agent sessions — and the friction lands on real buyers either way. The practical position: reserve hard challenges for endpoints that are actually abused, keep top-of-funnel pages and forms lightly defended, and rate-limit rather than wall off. If you run a chat widget as a funnel entry point, remember an agent may not execute your third-party widget script at all — never make chat the only path to contact.

Layer 2: Machine-readable entry points

Traversal is the floor. The next layer is making the right path obvious to software:

• llms.txt. The llms.txt proposal is a markdown file at your site root that curates your key pages for LLMs, expected by its authors to be useful mainly at inference time — exactly when a buying agent is deciding which of your pages to read. It is a proposal, not a standard with committed engine adoption, but it costs almost nothing to maintain. We publish one at zian.ai and regenerate it whenever content ships.
• XML sitemap and clean information architecture. Boring, load-bearing, frequently stale. An agent with a task budget follows the shortest credible path; a current sitemap and crawlable navigation shorten it.
• Schema markup on product and pricing pages. Structured data (Product, Offer, FAQPage, Organization) states machine-readably what a page is. For how far to take machine-readability — including whether you need an MCP server so agents can query you directly — see our analysis of MCP for AI buying agents.
• Structured contact channels. A tel: link and visible phone number, a contact page linked from every page footer, accurate business hours, and — given Google’s agentic calling pulls from it — a current Google Business Profile. If an agent phones you, whoever (or whatever) answers is your funnel; our whole product thesis is that an AI agent should be picking up that call the moment it rings.

Human vs agent: how each funnel element actually performs

Funnel element	How a human buyer experiences it	How an agent experiences it	Readiness fix
Demo-request form	Skims labels, autofills from browser, tolerates vague errors	Parses DOM; unlabelled fields and silent validation cause abandonment	Real labels, standard autocomplete tokens, text error messages
Multi-step wizard	Mild annoyance; may persist if motivated	Stalls on lost state, disabled buttons, unexplained requirements	Fewer steps, requirements in text, plain-form fallback
CAPTCHA / anti-bot wall	Solves in seconds, minor friction	May stall or fail the whole task; behaviour varies by challenge	Challenge only abused endpoints; rate-limit instead of wall
Chat widget	Convenient live channel	Third-party script may never execute; effectively invisible	Never the only contact path; mirror answers in indexable HTML
Phone line	Calls when urgent, hates hold queues	Google’s agentic calling asks structured questions about price and availability	Answer fast, keep Business Profile current, brief whoever answers
llms.txt / sitemap	Never sees them	Cheap, curated map of your best pages at inference time	Publish both; regenerate on every content change

Layer 3: What agent traffic looks like in your logs

Agent readiness is measurable, but only if you read your logs with the right taxonomy. The major AI vendors document distinct user agents with distinct jobs:

• Search-index crawlers — OAI-SearchBot (surfaces sites in ChatGPT’s search features), PerplexityBot, and Claude-SearchBot. These crawl proactively so you can be cited later.
• On-demand fetchers — ChatGPT-User, Perplexity-User and Claude-User. These fire because a specific user asked a question right now and the assistant went to your page to answer it. OpenAI notes that because these fetches are user-initiated, robots.txt rules may not apply to them; Perplexity documents the same user-initiated distinction for Perplexity-User.
• Training crawlers — GPTBot and ClaudeBot, which collect content for model development and honour robots.txt.

The distinction matters commercially. Index-crawler traffic is about future visibility; on-demand fetcher traffic is a live buyer conversation happening somewhere you can’t see. In our own server logs at zian.ai, we routinely see on-demand fetchers pulling exactly the pages a shortlisting buyer would: product and comparison pages, and — tellingly — compliance and regulatory explainers. Someone is asking an assistant hard diligence questions, and the assistant is coming here for answers. (We’ll describe that qualitatively only; log-derived counts are noisy, not least because scanners spoof agent user-agent strings, so verify with status codes and published IP ranges before trusting any number.)

Standard analytics barely captures any of this — fetchers execute no JavaScript and send ambiguous referrers. If you want to instrument it properly, we’ve documented the approach in measuring AI referral traffic, and the strategic context in our guides to answer engine optimisation and zero-click AI answers.

An honest agent-readiness maturity checklist

Score yourself against four levels. Most B2B sites we look at are at level 1.

1. Level 0 — Blocked. Anti-bot rules challenge or block AI user agents site-wide; chat widget is the primary contact path; forms rely on placeholders and colour-only errors. Agents fail and report failure to your prospects.
2. Level 1 — Passable. Agents can reach and read key pages, but the funnel is fragile: unlabelled form fields, a long wizard with no fallback, stale sitemap. Some agent sessions convert; many silently die.
3. Level 2 — Legible. Labelled forms with standard autocomplete attributes and text errors; schema on product and pricing pages; current sitemap and llms.txt; visible phone number and contact page; anti-bot friction scoped to abused endpoints only.
4. Level 3 — Agent-ready. Everything in level 2, plus: you monitor on-demand fetcher activity in logs, you’ve walked your own funnel with an agent and fixed what stalled it, phone enquiries — including agentic calls — are answered immediately, and every lead an agent submits gets sub-minute follow-up. That last mile is where we live: an agent-submitted lead decays exactly like a human one, just faster to notice.

The single highest-leverage exercise: give a browsing agent your own demo-request task and watch where it struggles. It is the cheapest funnel audit you will ever run. And treat agent-facing surfaces as untrusted-input surfaces too — see our piece on prompt injection risks for autonomous agents for the security half of this story.

Apply For Partnership

FAQs

Should I block AI bots from my site?

Not indiscriminately. Distinguish the three documented bot classes first: training crawlers (GPTBot, ClaudeBot), search-index crawlers (OAI-SearchBot, PerplexityBot, Claude-SearchBot) and on-demand fetchers (ChatGPT-User, Perplexity-User, Claude-User). Blocking training crawlers is a legitimate content-rights choice — Cloudflare moved to block AI crawlers by default in July 2025 for exactly that reason. But blocking index crawlers removes you from AI search results, and blocking on-demand fetchers turns away an assistant acting for a live prospect. For a B2B vendor that wants to be found and shortlisted, blanket blocking is usually self-harm.

Will CAPTCHAs stop AI buying agents?

Sometimes, and that is precisely the problem. Agents like ChatGPT agent operate a real remote browser, so simple bot heuristics may not catch them, while harder challenges can stall an agent mid-task — and the failed task is reported back to your prospect. There is no published guarantee in either direction, so assume CAPTCHAs are unreliable both as a defence against agents and as a filter that spares them. Scope hard challenges to endpoints with demonstrated abuse and leave your top-of-funnel lightly defended.

What is llms.txt and does it help?

llms.txt is a proposed convention — a markdown file at your site root listing your most important pages with short descriptions, intended per the llmstxt.org proposal to help LLMs find relevant content at inference time. Honestly: no major AI engine has publicly committed to consuming it, so treat it as a low-cost bet rather than a proven channel. It takes minutes to generate, costs nothing to maintain alongside your sitemap, and curates exactly the shortlist you would want an agent to read.

Can AI agents actually complete B2B demo-request forms today?

Yes, with caveats. OpenAI’s ChatGPT agent system card describes a model that executes tasks through a remote visual browser environment — clicking, typing and submitting forms — with user confirmations before consequential actions and a supervised watch mode in sensitive contexts. In practice that means an agent can complete a well-built form, will pause to ask its user before committing to anything binding, and will struggle or bail on fragile multi-step flows. Reliability varies by site, which is exactly why testing your own funnel with an agent is worth an afternoon.

How can I tell whether AI agents are visiting my site?

Check server logs, not analytics dashboards — agent fetchers don’t execute JavaScript. Look for the documented user-agent tokens (ChatGPT-User, OAI-SearchBot, GPTBot, PerplexityBot, Perplexity-User, ClaudeBot, Claude-User, Claude-SearchBot) and note which pages they request. Be sceptical of raw counts: scanners spoof popular AI user-agent strings, so filter to successful responses and, where vendors publish them, verify requests against official IP ranges before treating a spike as real buyer-agent interest.

Your buyers are already delegating the boring parts of buying. When their agent lands on your site, it will either glide to a submitted form and a booked call — or stall, fail and shortlist someone else. Walk your funnel as an agent would, fix what breaks, and make sure something intelligent answers the moment a lead (human or agent) comes through.

Apply For Partnership