Agent Legibility: How AI Agents Read and Use a Software Vendor's Site

Agent legibility is whether an AI agent can do a buyer’s job on your website. An agent here means software that opens a browser for a person, reads pages, fills in forms and reports back.

Google, Anthropic, Perplexity and Microsoft each ship one. This guide reads their documentation as it stood on 2026-11-03.

It covers how agents see a page, how they identify themselves, what stops them, and what a property-ops software vendor should change first. If you run marketing for that vendor, the practical output is a list for your web team. The commercial version is the answer engine work for proptech.

What is agent legibility?

Agent legibility is how well an agent working for a person can find, read and act on a site. It has three parts.

• Reach: the request gets through.
• Reading: the facts the task needs exist as text.
• Acting: the form, the button and the next step work without a person taking over.

Take a regional manager at a 2,000-door operator on Yardi. She asks an agent to find three package locker vendors that integrate with her PMS, note each one’s pricing model and request demos. Each vendor site either answers those three questions in text an agent can lift, or it does not.

Accessibility, crawler access and AI visibility each cover part of this. Accessibility asks whether assistive technology can use the page, and crawler access whether bots may fetch it.

AI visibility asks whether answers name you. Agent legibility is where all three meet a single task. Crawler access has its own guide to AI crawlers and robots.txt.

How is an agent different from a crawler or a fetcher?

A crawler collects pages on its own schedule, and a fetcher opens one URL because a person’s question needed it. An agent operates a browser for a person across several pages and can click, type and submit.

	Crawler	User-triggered fetcher	Browser agent
Who starts it	The operator	A person’s question	A person’s task
What it does	Collects pages	Reads one page	Reads, clicks, types, submits
Reads robots.txt	Yes, per the operators	Often not, per the operators	Often not, per the operators
Examples	GPTBot, ClaudeBot, Googlebot	ChatGPT-User, Claude-User, Perplexity-User	Gemini in Chrome auto browse, Claude in Chrome, Comet, Browse with Copilot
What it needs	Crawlable HTML	Readable HTML	Readable HTML and working controls

The lines blur at Google. Google’s list of user-triggered fetchers says Google-Agent “is used by agents hosted on Google infrastructure to navigate the web and perform actions upon user request.” So Google-Agent sits in both columns: it is a fetcher by name and an agent by job.

What is an AI browser agent?

An AI browser agent is an assistant that works inside a web browser for a person. It opens pages, reads them, clicks, types and submits, then reports back. Four are documented as of 2026-11-03.

• Gemini in Chrome auto browse (Google). Google’s Chrome help center says “With auto browse, you can ask Gemini in Chrome to complete multi-step tasks for you on the web.”
• Claude in Chrome (Anthropic), a browser extension that works by “reading, clicking, typing, navigating, and filling forms,” per Anthropic’s getting-started guide.
• Comet (Perplexity), a browser with an assistant panel that reads the open page.
• Browse with Copilot (Microsoft) in Edge, which replaced Copilot Mode in May 2026.

A browser agent runs in the person’s own browser, so it arrives with that person’s sessions. Google’s page says “Gemini has access to all the same sites that you do, including sites you’re signed into.” It adds: “When completing an auto browse task on a website, Gemini may share your information with that site.”

For a vendor, that means a demo form can receive a real buyer’s details, typed by an agent. The lead is as real as one typed by hand, and your form decides whether it arrives.

What is the agentic web?

The agentic web is Microsoft’s name for a web where agents act for people and organizations, not only show them pages. Microsoft used it in its Build 2025 post of 19 May 2025. The post says “This emerging vision of the internet is an open agentic web, where AI agents make decisions and perform tasks on behalf of users or organizations.”

Microsoft launched NLWeb alongside the term. The same post says “Microsoft is introducing NLWeb, which we believe can play a similar role to HTML for the agentic web.” NLWeb’s README says “The response returned uses Schema.org,” so it builds on the structured data many sites already publish.

The protocols around it are young. The Model Context Protocol site calls MCP “an open-source standard for connecting AI applications to external systems.” WebMCP proposes the same idea inside a web page, so a site can hand an agent named tools; the WebMCP guide covers it.

Agents that pay as well as browse are covered in the agentic commerce guide. For a property-ops vendor, the page an agent reads is still the plain HTML a buyer reads, so the sections below apply first.

How do AI agents see a page?

Agents see a page as a picture, as code or as the accessibility tree, and most use more than one. Google’s web.dev guidance for agent-friendly sites says “Agents can view your website in 3 primary ways: screenshots, raw HTML, and the accessibility tree.” Google Search Central says browser agents may gather data by “analyzing visual renderings (like screenshots), inspecting the DOM structure, and interpreting the accessibility tree.”

The accessibility tree is the browser’s own summary of the page for assistive technology. MDN says “Browsers then create an accessibility tree based on the DOM tree,” and screen readers read from it. Each element in it has a role, a name and a state, so a button with no name reads as an unlabeled control.

What each operator documents

Anthropic’s browser use tool documentation says Claude works with a page “through its structure (the accessibility tree, elements, forms, and tabs) and through screenshots.” It tells developers to “Prefer references where the page has a usable accessibility tree.” Anthropic’s safety guide for Claude in Chrome says the extension takes “screenshots of the tabs it’s working in.”

Microsoft’s support page for Browse with Copilot in Edge says Copilot “takes screenshots of the pages it visits” to see and interact with them. Perplexity’s Comet help center says the assistant “pulls in the visible text, headlines, and website metadata from your open page every time you submit a request.” OpenAI’s computer use guide for developers describes a model that works from screenshots and requests clicks, scrolls and typing.

Google’s Chrome help page for auto browse says what the agent may do, not how it perceives the page. OpenAI’s help center, where ChatGPT’s own agent features are documented, returned an error to every request on 2026-11-03. This page makes no claim about ChatGPT’s agent from it.

What breaks an agent on a vendor site?

The same things that break a screen reader, plus anything that moves. Google’s web.dev guidance tells sites to “use <button> and <a> tags over modified <div> and <span> elements” and to set the for attribute on <label> tags. A clickable <div> has no role in the tree, so an agent reading the tree may never find it.

Layout matters to the agents that read pixels. The same guidance says “Agents that take screenshots will likely be confused if your website layout is constantly shifting” and tells sites to “Avoid transparent overlays that hide interactive elements.” A chat widget that slides over the pricing table, or a cookie banner over the demo button, is that overlay.

Some content has nothing in the tree at all. Anthropic’s browser use documentation says “Canvas-rendered interfaces, embedded video or remote-desktop surfaces, heavily virtualized lists, and elements inside cross-origin iframes often have no useful node.” A pricing calculator drawn on a canvas, or a demo form embedded from a marketing platform in a cross-origin iframe, falls into that list.

The property-ops version of the list

• Pricing shown only as an image, a PDF or “Contact sales” with no pricing model stated in text.
• PMS integrations shown as a wall of logos with no names in text or alt attributes.
• A demo form inside a third-party iframe, with fields labeled only by placeholder text.
• A “Book a demo” control built from a styled <div>.
• Portfolio fit (doors, unit types, SFR or multifamily) stated only in a video.

Can an agent fill in your demo form?

Often, but it will stop at the points its operator marks as sensitive. Anthropic’s getting-started guide says Claude in Chrome works by “reading, clicking, typing, navigating, and filling forms.” Google’s auto browse help page lists “Submitting web forms” among the actions Gemini in Chrome asks the user to confirm first.

Account creation is a hand-back point for both. Google’s page lists “Creating an account” among the actions where Gemini in Chrome may ask the user to take over the task. Anthropic’s permissions guide lists “Creating accounts” among the actions Claude is prohibited from taking “regardless of permissions.”

A demo request that requires an account first sends these agents back to their users.

Microsoft says Browse with Copilot “cannot access autofill data, saved passwords, or wallet information.” OpenAI’s computer use guide tells developers to “Keep users in control of purchases, data transmission, destructive changes, and other actions that are hard to reverse.” Expect the agent to bring a filled form back for approval more often than it submits one.

What a CAPTCHA does

It ends the task. Anthropic’s crawler documentation says “we will not attempt to bypass CAPTCHAs,” and its Claude in Chrome safety guide lists “Bypassing captchas” as a prohibited action. Google’s sample system prompt for its computer use model tells the agent it must never attempt to solve or bypass robot detection.

Cloudflare’s Turnstile overview says it runs challenges “to distinguish human visitors from automated traffic,” and the pages read on 2026-11-03 do not mention agents. A challenge on the demo form treats a buyer’s agent as automated traffic.

This site runs Turnstile on its own forms. An agent filling one in would likely hand the last step back to its user.

How do AI agents identify themselves?

By user-agent token, by published IP ranges and, for a growing share of requests, by a cryptographic signature. Google says Google-Agent uses IP ranges from a published file and warns that “The user agent string can be spoofed.” Google’s Web Bot Auth page says “A subset of requests made by the Google-Agent are signed with Web Bot Auth,” authenticated as https://agent.bot.goog.

The signing scheme is an IETF draft. The draft, “HTTP Message Signatures for automated traffic” (draft-ietf-webbotauth-httpsig-protocol-01, 29 October 2026), lets automated clients sign requests, “allowing HTTP servers to verify their identity with confidence.” It defines a Signature-Agent header that points to the signer’s key directory.

Google’s Web Bot Auth page says “Major bot-detection services, CDNs, and WAFs support Web Bot Auth.” It tells sites to fall back to IP-based checks, since “not all requests are signed.” Anthropic’s and Perplexity’s crawler pages describe user-agent tokens and IP lists, with no signing.

What Cloudflare does with agent traffic

Cloudflare sits in front of a large share of vendor sites, so its defaults decide much of what an agent meets. Its verified bots page says “Signed agents are now Verified,” with the label intermediary. It warns that “you may trust the intermediary operator, but not necessarily every end user driving it.”

Ask your web team which bot rules run on the pricing, integrations and demo pages, and whether verified bots are exempt. A rule written to stop scrapers can stop a buyer’s agent on the one page that mattered.

Does robots.txt control agents?

Mostly not, for user-requested work.

Google says that because the fetch was requested by a user, its user-triggered fetchers “generally ignore robots.txt rules.” OpenAI says of ChatGPT-User: “Because these actions are initiated by a user, robots.txt rules may not apply.” Perplexity says Perplexity-User “generally ignores robots.txt rules.”

Anthropic takes the other position. Its crawler documentation says disabling Claude-User “prevents our system from retrieving your content in response to a user query, which may reduce your site’s visibility for user-directed web search.” Blocking Claude-User costs you that path, and the other operators’ agents keep coming.

So robots.txt states a policy and enforces nothing. Your CDN rules, your WAF and your challenges are where an agent gets a 403. The robots.txt guide has the allow file this site uses and the log checks that show what reached you.

Which files and formats help an agent?

The page itself helps most; the extra files help some tools and not others. Google Search Central says “You don’t need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search.” On llms.txt it says creating one “will neither harm nor help” a site in Google Search, “as Google Search ignores them.”

Those statements cover Google Search only.

The llms.txt proposal describes itself as “A proposal to standardise on using an /llms.txt file to provide information to help agents use a website.” Cloudflare’s Agent Readiness announcement says its score does “not check for llms.txt” by default. The llms.txt guide reports what the file did and did not change on this site.

Markdown on request

Some sites now return Markdown when a client asks for it with an Accept: text/markdown header. Cloudflare’s Markdown for Agents documentation says its network converts the HTML to Markdown on the fly when a client expresses that preference. Cloudflare’s Agent Readiness announcement of 17 April 2026 says Markdown content negotiation “passes on 3.9% of sites.”

This site publishes a Markdown twin of each page at …/index.md, linked from each page’s head. It does not yet answer Accept: text/markdown on the HTML address; a request with that header gets HTML. Structured data is a separate question, covered in the structured data guide.

What does an agent need from a property-ops vendor’s site?

The same facts a buyer needs, as text, on pages an agent can reach in a few clicks. The guide to ranking in ChatGPT sets out five facts a property manager’s question turns on. An agent comparing vendors needs the same five, plus a way to act.

1. Price. Per unit or per door per month, or a starting price with its conditions. A range is better than “Contact sales”.
2. Portfolio fit. The minimum or ideal portfolio size, and SFR or multifamily.
3. PMS integrations by name. Yardi, RealPage, Entrata, AppFolio, Buildium, Rent Manager, in text rather than logos.
4. What the PMS’s built-in module lacks. The reason to buy you on top of the platform.
5. Service layer or software only. Whether people stand behind the product, such as after-hours triage.
6. The demo path. A plain link to a form with labeled fields and no account step.

Each fact should sit on the page an agent would open for it. Pricing belongs on /pricing/, integrations on /integrations/. The entity consistency guide covers keeping those facts the same on G2, Capterra and your PMS marketplace listing, which agents also read.

What the Index shows

The Index does not measure agents. It measures search-grounded answers, which are the step before an agent opens your site. What those answers link to shows which vendor pages already carry the facts.

In the October 2026 edition, ChatGPT’s 500 answers carried 2,853 links, and 1,884 of them (66%) pointed to domains of vendors in the Index’s 285-vendor record. Only 348 of those 1,884 vendor links (18.5%) went to a homepage, and 93 (4.9%) went to a URL with “pricing” in its path.

Claude’s 500 answers in the same edition carried 3,848 links, 1,511 of them (39%) to vendor domains in the record. Of those, 154 (10.2%) went to a homepage, 11 (0.7%) to a pricing URL and 524 (34.7%) to blog, article or resources paths.

From those counts we infer that answers land on deep pages, so each product, integration and pricing page has to stand alone. Our working assumption is that pricing pages are linked so rarely because most vendors in the record have no pricing page with figures on it. We have not counted that yet.

How do you test your own site in 30 minutes?

Run four checks on the pricing, integrations and demo pages, in this order.

1. Turn JavaScript off and reload. If the pricing model or the integrations list disappears, a text-based fetcher sees nothing there.
2. Open the accessibility tree. In Chrome DevTools, the Accessibility pane shows each element’s role and name. Look for unnamed buttons and form fields named only by placeholder.
3. Tab through the demo form. Every field should announce a name, and the submit control should be a real button. If a person on a keyboard cannot finish it, an agent reading the tree likely cannot either.
4. Give an agent a real task. In Gemini in Chrome, Claude in Chrome or Comet, ask for your pricing model and your Yardi integration, then ask it to request a demo. Note where it stops.

Cloudflare’s Agent Readiness announcement introduces a public checker at isitagentready.com that scores discoverability, content, bot access control and capabilities. Read its score as a checklist of emerging standards. Checks 1 to 4 are the ones a buyer’s agent will run into.

What should a marketing lead send the web team?

Five requests, each small enough for a sprint.

1. List the bot, WAF and challenge rules on /pricing/, /integrations/ and the demo form, and say whether verified bots are exempt.
2. Replace clickable <div> and <span> controls with <button> and <a>, and give every form field a <label>.
3. Put the price, the portfolio fit and the integrations by name in HTML text on those pages.
4. Move the demo form out of a cross-origin iframe, or confirm its fields are labeled inside it, and remove any account step.
5. Keep overlays (chat, cookie, promotion) off the pricing table and the demo button.

None of this needs a new file format, and the same changes make the pages work with a screen reader.

Limits

This page is documentation, read on one day, plus one count from the Index. Every quoted sentence comes from a page named in the sentence, and every one of those pages can change without notice. OpenAI’s help center could not be read on 2026-11-03, so ChatGPT’s own agent features are not described here.

The documents describe intent, not enforcement. An operator’s statement that its agent will not bypass a CAPTCHA, or will ask before submitting a form, is the operator’s statement. We have not tested any agent against any vendor site for this page.

The Index figures count links in search-grounded answers from the October 2026 edition, not agent sessions. URLs were classed by path alone: “pricing” in the path, or a blog, articles, resources, learn, insights or news segment. A pricing page at another address counts as a deep page.

What changes month to month?

Product names change fastest. Microsoft retired Copilot Mode in Edge on 13 May 2026 and moved its agent to Browse with Copilot, so re-read the operators’ pages before repeating anything here.

Signing will change how agents are checked. The Web Bot Auth draft is a working-group draft, and Google calls its own use experimental. Once more operators sign requests, IP lists stop being the main check.

The checklists grow. Cloudflare’s checker already lists more standards than its April announcement did.

Related guides

The robots.txt guide covers which bots to allow and how to read your logs. The structured data guide and the llms.txt guide cover the machine layer around the page. The guide to ranking in ChatGPT and the guide to how Claude picks vendor sources cover the answers that send agents to you.

The pillar on getting named by AI assistants puts these in order, and the Index methodology explains how the October counts were captured.

Changelog

• 2026-11-03: Added sections on AI browser agents and the agentic web, with links to the WebMCP and agentic commerce guides.
• 2026-11-03: Published at this address, which had redirected to the AI search pillar since the September 2026 rebuild. Documentation read on 2026-11-03; Index counts from the October 2026 edition.