Why AI visibility is not enough for website agents

AI visibility answers a useful but incomplete question: can an AI system find and quote your website? A business rarely wants the journey to end there. It wants the visitor, human or automated, to compare a plan, check availability, book a call, submit a claim, add an item to a cart, or finish some other task without getting lost.

That gap is easy to miss because visibility produces tidy metrics. You can count crawler visits, AI citations, referral sessions, and mentions. Task completion is harder. It crosses page structure, state, permissions, forms, APIs, errors, and confirmation. A page can perform beautifully in an AI answer while the product behind it remains unusable to the same agent.

The practical model is a five-stage funnel:

TextCopiedDiscoverable -> Crawlable -> Understandable -> Operable -> Confirmable

The first three stages earn attention. The last two decide whether that attention turns into a result.

Key takeaways

• AI visibility measures whether systems can find, access, and interpret a page.
• WebArena's 2023 baseline completed 14.41% of realistic tasks, despite using a capable GPT-4-based agent.
• Agent readiness must test actions and durable outcomes alongside content eligibility.

What does AI visibility actually prove?

Google's December 2025 guidance, AI features and your website, says pages need no special AI markup to appear as supporting links in AI Overviews or AI Mode. A page must be indexed, snippet-eligible, and meet ordinary Search requirements. That proves eligibility for retrieval. It says nothing about whether an agent can use the product.

Visibility work still matters. A blocked, unindexed, or badly structured page cannot explain the offer. Clear headings, crawlable text, canonical URLs, accurate schema, source-backed answers, and sensible internal links help search systems and assistants understand what the site knows.

But the evidence stops at the page boundary. A citation does not prove that a price selector exposes its current value. An indexed support article does not prove that the returns form preserves data after validation fails. A well-ranked booking page does not prove that its calendar can be operated with stable names and states.

This is the distinction between evidence about content and evidence about behavior. Both belong in an AI agent readiness program, but they need different tests.

Google requires indexed, snippet-eligible pages for possible inclusion as supporting links in its AI search features. These conditions establish visibility eligibility, not task completion. A separate operational test must show whether an agent can act, recover from errors, and verify the requested end state.

Why is being found different from finishing a task?

In 2023, the WebArena paper reported 14.41% end-to-end success for its best GPT-4-based baseline, compared with 78.24% for people. WebArena evaluates 812 long-horizon tasks on functional websites. Its evaluator checks whether the requested result happened, not whether the agent reached the right page or described the interface.

That is a far stricter standard than retrieval. Consider the request, "Find a red carry-on under $150, add it to my cart, and stop before checkout." The agent has to interpret the constraint, operate filters, inspect results, choose a qualifying item, select any required options, change cart state, and recognize the stopping boundary. A search citation might help it find the store. It does not complete any of those steps.

Mind2Web exposes the same problem from another angle. The NeurIPS 2023 dataset contains more than 2,000 tasks from 137 websites across 31 domains. The researchers note that real sites are dynamic and noisy, and raw HTML can be too large for direct model input. An agent has to decide which tiny part of that surface matters now.

Search systems optimize for relevant information. Task agents need relevant information plus correct state transitions. Confusing those outcomes leads teams to polish content while broken actions remain untouched.

WebArena's reported GPT-4-based baseline completed 14.41% of 812 realistic tasks, while people completed 78.24%. The benchmark's functional evaluator makes a useful distinction: reaching or understanding a page is intermediate progress; the requested state change is the outcome.

What are the five stages of an agent-ready journey?

The funnel is an analytical model from CanAgentUse, not an industry standard. Its purpose is diagnostic. When a task fails, the stages help a team locate the first broken contract instead of treating "the AI" as a single opaque visitor.

The stages are ordered because later strengths cannot repair earlier failures. A perfect checkout API does not help an agent that cannot discover it. Excellent schema cannot rescue a form with an unlabeled submit button. A successful network request is not enough when the agent cannot determine whether it created one order or two.

The model also prevents an audit score from becoming decorative. If a score says 92 but the primary quote flow fails at confirmation, the score is hiding the business outcome that matters.

Why is crawl permission not action permission?

RFC 9309, published in 2022, standardizes the Robots Exclusion Protocol and explicitly warns that robots.txt is not access control. The file communicates crawler preferences. It does not authenticate an actor, authorize a purchase, or establish what a user allowed an agent to do.

This boundary matters because "allow AI agents" can mean at least three things. A company may allow search crawlers to index public documentation, allow a user-directed fetcher to read an account-free page, and refuse automated changes unless an authenticated user approves them. Those are compatible policies.

Authentication proves an identity or session. Authorization decides which resources and actions it may use. A user mandate narrows that power to the current instruction, perhaps a maximum price or a rule to stop before payment. The interface then needs a clear confirmation boundary for any consequential action.

Treating crawler policy as transaction permission is unsafe. Treating every automated request as hostile is not much better. The site needs separate, explicit controls for retrieval, identity, authority, and consent. Our signed agent access guide covers the identity side in more detail.

What makes a website operable by an agent?

Operability comes from explicit contracts. In the browser, that means controls with stable roles and accessible names, inspectable values, visible selected state, errors linked to the affected field, and focus that follows dialogs and multi-step flows. Our browser-agent UX guide shows how those details change forms, filters, and modals.

Some tasks also deserve a structured action surface. OpenAPI can describe a conventional HTTP API. MCP can expose named tools to configured clients. Chrome's proposed WebMCP work can expose page tools through JavaScript or annotated forms. Each option reduces interpretation, but none removes the need for authentication, authorization, validation, and user review.

The right interface depends on the task:

An agent-ready site does not replace its interface with an API. It gives each important task a dependable path and makes the consequences readable.

Why does confirmation break weak agent journeys?

Confirmation answers two questions: did the action finish, and what exactly changed? A green toast saying "Done" is poor evidence. It lacks an object, identifier, time, and durable place to inspect the result. If it disappears, an agent may repeat the action and create a duplicate.

Good confirmation is proportional to risk. Saving a display preference may only need a persistent selected value. Creating a support ticket should return a ticket number and a link. Booking an appointment should show the date, timezone, attendee, and cancellation path. A payment step needs an amount, merchant, line items, and a deliberate final boundary.

Idempotency also matters for APIs and tools. If a retry can create a second order, the interface should accept an idempotency key or expose a safe lookup before creation. If a browser form cannot offer that contract, the resulting receipt still needs enough detail for an agent to detect a duplicate.

This is why "the button clicked" is a useless success metric. The observable end state is the metric.

How should teams measure more than AI visibility?

Start with a small set of valuable tasks, not a generic claim that the whole site is agent ready. Write each task as a constrained outcome: starting state, allowed actions, required result, forbidden side effects, and proof of completion.

For each task, record stage-level evidence:

1. The agent found the intended page or tool without a hidden hint.
2. The intended crawler or session could retrieve the required resources.
3. The agent identified the relevant controls, values, constraints, and current state.
4. It completed the allowed action and recovered from realistic validation errors.
5. It verified the result from a durable artifact and stopped at the correct boundary.

Track task success rate beside citation and referral metrics. Also track assisted completion, where a person had to clarify or take over, and unsafe attempts, where the agent exceeded the instruction or approached a consequential boundary without adequate review.

The useful dashboard joins both halves of the journey. Visibility explains how an agent arrived. Completion evidence explains whether the product worked.

Frequently asked questions

Is AI visibility the same as AI agent readiness?

No. Google's 2025 guidance says indexed and snippet-eligible pages may appear as supporting links in AI features. Agent readiness extends beyond that eligibility. It tests whether an agent can understand state, perform an approved action, handle errors, and verify the result.

Does an AI citation mean the cited website is usable by agents?

No. WebArena evaluated 812 realistic tasks and reported 14.41% success for its best GPT-4-based baseline. Citation is evidence that content was retrievable and useful. It does not test a form, cart, account workflow, permission boundary, or completion artifact.

Do websites need special AI markup for visibility?

Google says no special AI file or schema is required for its AI search features. Ordinary technical SEO still applies. Separate machine-readable interfaces may help agents act, but they solve an operational problem rather than a visibility requirement.

What should we test first?

Test one valuable, low-risk task from discovery through confirmation. A quote request, support ticket, or catalog search is usually more revealing than another homepage crawl. Once the evidence model works, add higher-risk account and commerce tasks.

A citation is the beginning of the journey

AI visibility deserves investment. It helps the right systems find the right evidence. But visibility is a doorway, not a completed task.

The sharper operating question is simple: after an agent arrives, can it understand the current state, act within the user's authority, recover when something goes wrong, and prove the outcome? A site-wide agent-readiness audit should answer that with page and task evidence, not a single flattering number.

Research sources

• Google Search Central, AI features and your website, updated December 10, 2025, retrieved 2026-07-12.
• Zhou et al., WebArena: A Realistic Web Environment for Building Autonomous Agents, ICLR 2024, retrieved 2026-07-12.
• Deng et al., Mind2Web: Towards a Generalist Agent for the Web, NeurIPS 2023, retrieved 2026-07-12.
• IETF, RFC 9309: Robots Exclusion Protocol, September 2022, retrieved 2026-07-12.