homepageshop.lab.webforagents.dev/
Checking page code against public standards: can agents recognise and read it?
Browser agents100
We check the code of the task pages against public standards (WCAG, axe, Lighthouse): do controls have names, do forms have labels, is anything in the way. We only open and read pages; we never click or type.
B1Reachable in a browser
What we found
- The pages load in a browser.
Why it matters
Browser agents have to load the page before they can do anything with it.
Check it yourself
Open https://shop.lab.webforagents.dev/ in a private browser window.B2Interactive elements identifiable
BlocksAccessibility-tree agentsDOM agents
What we found
- 100% of 58 interactive elements can be identified by an agent.
Why it matters
Browser agents build their picture of a page from the accessibility tree. A control without a name, or a clickable element that is not a control, is invisible or ambiguous to them.
If this check fails, this part of the score can be at most 69 (grade D).
Check it yourself
npx @axe-core/cli https://shop.lab.webforagents.dev/ --rules button-name,link-name,labelB4Key facts after rendering
BlocksAccessibility-tree agentsDOM agents
What we found
- "Understand what the site does": the information is visible on https://shop.lab.webforagents.dev/.
- "Check pricing plans": the information is visible on https://shop.lab.webforagents.dev/pricing.
- "Look something up in docs": the information is visible on https://shop.lab.webforagents.dev/docs/getting-started.
Why it matters
Once an agent reaches a task's page, the information it came for (a price, the main text, contact details, what the site does) has to be visible text on the page. If it sits only in collapsed content, hidden data or an image, the agent cannot read it or needs extra steps.
Check it yourself
Open the task's page (for example the product or contact page of https://shop.lab.webforagents.dev/) in a browser without expanding anything, and check that the information is shown on the page.B10Real links
BlocksDOM agentsAccessibility-tree agents
What we found
- Every link on the pages has a real address.
Why it matters
Agents that read the DOM follow a link by its address. A div or span that navigates with a script has no address, so they have to click it, and some tools that read the accessibility tree do not list it at all. In our runs this cost one to five extra steps.
Check it yourself
curl -s https://shop.lab.webforagents.dev/ | grep -Eo '<a[^>]*href="(#|javascript:)[^"]*"|onclick="[^"]*location[^"]*"' | headB6Forms recognisable
BlocksAccessibility-tree agentsScreenshot agents
What we found
- Forms on 5 pages can be recognised.
Why it matters
To sign up, get in touch or book, an agent first has to recognise what each field is and where the submit button is. We never fill or submit forms; we only check whether they can be recognised.
Check it yourself
Open an action task's entry page (sign-up, for example) and check in the browser's accessibility panel that every input has a name.B3Not obstructed
BlocksScreenshot agentsAccessibility-tree agents
What we found
- No page is covered by an overlay after loading.
Why it matters
Screenshot agents cannot see what an overlay covers, and agents that read the accessibility tree find the elements behind it but cannot click them. Gemini's and Microsoft's documentation both name unexpected pop-ups as a cause of errors.
Check it yourself
npx lighthouse https://shop.lab.webforagents.dev/ --only-audits=cumulative-layout-shiftB11Content in the page's own interface
BlocksAccessibility-tree agentsDOM agents
What we found
- Content and controls are in the pages' own HTML.
Why it matters
Text and controls drawn on a canvas, or a form inside a page embedded from another site, do not appear in the page's accessibility tree. Claude's browser tool documentation names both as cases where it gets no elements and falls back to screenshots; an agent that reads only the tree gets nothing.
If this check fails, this part of the score can be at most 69 (grade D).
Check it yourself
Open https://shop.lab.webforagents.dev/ with the browser's accessibility panel and check that the content appears in the tree.B12States stated
BlocksAccessibility-tree agents
What we found
- Menus and controls state whether they are open, selected or checked.
Why it matters
Agents that read the accessibility tree tell whether a menu is open, a tab selected or a box checked from its state. Without it they click again or click the wrong thing.
Check it yourself
npx @axe-core/cli https://shop.lab.webforagents.dev/ --rules aria-required-attrB13No native dialogs
BlocksAccessibility-tree agentsScreenshot agents
What we found
- No page opens a native dialog while loading.
Why it matters
alert, confirm and prompt dialogs stop the page until someone answers. Claude in Chrome's documentation says they block it until a person closes them, and Amazon's Nova Act cannot handle them either.
Check it yourself
Open https://shop.lab.webforagents.dev/ with the console open and watch for dialogs while it loads.B14Layout stable
BlocksScreenshot agentsAccessibility-tree agents
What we found
- Layouts stay stable (largest layout shift 0.004).
Why it matters
Screenshot agents click by position, so a shift moves the target away from the click; agents that read the accessibility tree find their element references go stale. Chrome's Lighthouse lists layout stability among its agentic browsing checks.
Check it yourself
npx lighthouse https://shop.lab.webforagents.dev/ --only-audits=cumulative-layout-shiftB15No new tabs
BlocksScreenshot agentsAccessibility-tree agentsDOM agents
What we found
- Links within the site open in the same tab.
Why it matters
A link that opens a new tab makes every kind of agent find and switch to that tab. In our runs Codex took one more step each time.
Check it yourself
curl -s https://shop.lab.webforagents.dev/ | grep -o 'target="_blank"' | wc -lB16Pages readable in one go
BlocksAccessibility-tree agents
What we found
- Every page can be read in one go (largest snapshot 1,084 tokens).
Why it matters
Agents read a page's snapshot as one tool result, and some cut it off: Codex at 10,000 tokens, Claude Code at 25,000 (about 21,000 in our count). Claude in Chrome reads 50,000 characters by default but can ask for more, so it is not counted.
Check it yourself
Open https://shop.lab.webforagents.dev/ and see how long its accessibility snapshot is.B17No hover-only content
BlocksScreenshot agentsAccessibility-tree agents
What we found
- No hidden navigation menus.
Why it matters
Screenshot agents only see a menu after hovering over it, and a menu hidden with CSS is not in the accessibility tree. A menu that opens with a button can be found and opened by every kind of agent.
How to fix
Open menus with a <button aria-expanded> that also works on click, and keep important links visible.
Check it yourself
Check whether the navigation menus on https://shop.lab.webforagents.dev/ open with a click or the keyboard.B7Targets easy to hit
BlocksScreenshot agents
What we found
- Text contrast (shown, not scored): 100% of 78 text elements pass.
Why it matters
Screenshot agents click by position, so small or crowded targets get missed; Claude's computer use documentation notes that small text and icons need zooming in. Text contrast is shown but not scored.
Check it yourself
Run Lighthouse's accessibility check in Chrome DevTools and look at target size and contrast.B18Direct addresses
BlocksDOM agentsScreenshot agentsAccessibility-tree agents
What we found
- No "load more" buttons.
Why it matters
Agents move between pages by address, and Gemini's and OpenAI's agents can type an address directly. Items loaded only by a "load more" button have no address to come back to.
How to fix
Add page links (or rel="next") alongside "load more" and infinite scroll, and give tabs and filters their own URLs.
Check it yourself
Check whether later items on https://shop.lab.webforagents.dev/ can be reached through a page link.B19Same name, same place
BlocksAccessibility-tree agents
What we found
- Link names are distinct: each name goes to one place.
Why it matters
An agent reading the accessibility tree sees a list of link names. When several "Learn more" links go to different places, it cannot tell which one to follow and may pick the wrong one.
Check it yourself
Look for links on https://shop.lab.webforagents.dev/ with the same name that point to different addresses.B20Visible text matches the name
BlocksScreenshot agentsAccessibility-tree agents
What we found
- Every control's visible text is part of its name.
Why it matters
Many agents find a button on a screenshot and then look it up by name in the accessibility tree. If the button says "Buy now" but its name is something else, the two do not match.
Check it yourself
Check the buttons and links with aria-label on https://shop.lab.webforagents.dev/: does the name contain the text they show?B5Page structure
BlocksAccessibility-tree agents
Why it matters
Agents that read the accessibility tree find the main content through landmarks and headings, and read tables through their headers.
Check it yourself
npx @axe-core/cli https://shop.lab.webforagents.dev/ --rules bypass,landmark-one-main,page-has-heading-one,document-title,heading-order,th-has-data-cellsB8CAPTCHA
What we found
- No action-task entry page was rendered.
Why it matters
Browser agents usually stop or hand control back to the user at a CAPTCHA. A CAPTCHA is a site's normal protection, so this is shown and never penalised.
How to fix
Nothing to fix. If you want agents to get through, consider showing the CAPTCHA only for suspicious traffic.
Check it yourself
Open an action task's entry page and look for a reCAPTCHA, hCaptcha or Turnstile widget.B9Load speed
What we found
- These are load times we measured on our own server. Images, media and fonts are not loaded and our network location differs from a real agent's, so the numbers are for reference only and not scored.
- home (https://shop.lab.webforagents.dev/): DOMContentLoaded 821 ms, load event 822 ms, network quiet 825 ms.
- pricing (https://shop.lab.webforagents.dev/pricing): DOMContentLoaded 790 ms, load event 790 ms, network quiet 793 ms.
- docs (https://shop.lab.webforagents.dev/docs/getting-started): DOMContentLoaded 837 ms, load event 838 ms, network quiet 839 ms.
- page declared by the site (https://shop.lab.webforagents.dev/tools/resize?img=img&width=1280&format=webp): DOMContentLoaded 784 ms, load event 785 ms, network quiet 788 ms.
- page declared by the site (https://shop.lab.webforagents.dev/tools/compress?img=img&kb=500): DOMContentLoaded 823 ms, load event 824 ms, network quiet 826 ms.
Why it matters
A browser agent waits for each page to load before it can read or click. The slower the page, the more time the agent spends, and past the tool's wait limit the step simply fails. This item is shown, not scored, for now.
How to fix
Cut the scripts and requests the first view waits for: defer non-essential third-party scripts, compress and cache static files, and let the main content render first.
Check it yourself
Open https://shop.lab.webforagents.dev/ in the Performance panel of Chrome DevTools, or run Lighthouse, and see how long the page takes to load.
Fetch agents83
Checks of the raw HTML against fetch agents' documentation, code and measured behaviour: content without JavaScript, search crawlers allowed, key facts in the visible text, pages not too large. We only read; we run no scripts.
F1Reachable
What we found
- The homepage and the key pages we fetched returned 200 with HTML.
Why it matters
A blocked or challenged response leaves a fetch agent with nothing. This has the strongest measured effect on agent success.
Check it yourself
curl -sI -A 'Claude-User' https://shop.lab.webforagents.dev/Evidence
- 200https://shop.lab.webforagents.dev/homepage HTML
- 200https://shop.lab.webforagents.dev/pricingpricing page HTML
- 200https://shop.lab.webforagents.dev/docs/getting-startedkey page
F2Content without JavaScript
BlocksWhole-page fetchersWindow readers
What we found
- 100% of the text shown in a browser is already in the raw HTML.
Why it matters
The fetch tools of Claude Code, Codex and Gemini CLI, and the major assistants' fetchers, do not run JavaScript.
If this check fails, this part of the score can be at most 69 (grade D).
Check it yourself
curl -s https://shop.lab.webforagents.dev/ | sed -e 's/<[^>]*>//g' | wc -wEvidence
- 200https://shop.lab.webforagents.dev/homepage HTML
- 200https://shop.lab.webforagents.dev/pricingpricing page HTML
F4Reading cost for fetch agents
BlocksWhole-page fetchersWindow readersIndex readers
What we found
- {url}: within every truncation limit (about {tokens} tokens, shown only).
Why it matters
Every token an agent reads costs its user money and time. Fetch agents convert your HTML to markdown and read all of it; tools also cut pages longer than 1 MiB of HTML or 100,000 characters.
Check it yourself
curl -s https://shop.lab.webforagents.dev/ | wc -cEvidence
- 200https://shop.lab.webforagents.dev/homepage HTML
- 200https://shop.lab.webforagents.dev/pricingpricing page HTML
- 200https://shop.lab.webforagents.dev/docs/getting-startedkey page
F5Markdown available
What we found
- https://shop.lab.webforagents.dev/ has no markdown version: a request for text/markdown returned HTML and no alternate is advertised.
- https://shop.lab.webforagents.dev/pricing has no markdown version: a request for text/markdown returned HTML and no alternate is advertised.
Why it matters
Markdown cuts the tokens an agent spends on a page. Claude Code, Cursor and OpenCode ask for it.
How to fix
Return text/markdown when the request's Accept header prefers it, and send Vary: Accept. Or advertise a .md version with <link rel="alternate" type="text/markdown">.
Check it yourself
curl -sI -H 'Accept: text/markdown, text/html, */*' https://shop.lab.webforagents.dev/Evidence
- 200https://shop.lab.webforagents.dev/homepage as markdown
- 200https://shop.lab.webforagents.dev/pricingpricing page as markdown
F7HTTP semantics
BlocksWhole-page fetchers
What we found
- Status codes, redirects and content types behave as agents expect.
Why it matters
Agents rely on status codes and redirects. A missing page that answers 200 looks like content.
Check it yourself
curl -sI https://shop.lab.webforagents.dev/does-not-existEvidence
- 404https://shop.lab.webforagents.dev/wfa-probe-cph1rmtt5wjkc9dfrandom missing page
- 200https://shop.lab.webforagents.dev/homepage HTML
F8robots.txt and user-triggered agents
BlocksClaude appsDuckDuckGo
What we found
- No robots.txt: all agents are allowed.
- robots.txt allows the user-triggered agents that honour it.
Why it matters
robots.txt is the one access standard every major crawler honours; Claude-User honours it even for fetches a user asks for.
Check it yourself
curl -s https://shop.lab.webforagents.dev/robots.txtEvidence
- 404https://shop.lab.webforagents.dev/robots.txtrobots.txt
F10Allowed into search indexes
BlocksIndex readers
What we found
- https://shop.lab.webforagents.dev/: search crawlers disallowed in robots.txt: ; robots directives on the page: noindex,nofollow,noarchive.
Why it matters
ChatGPT, Perplexity, Copilot and Google's AI Overviews answer most questions from their own index instead of fetching the page. If robots.txt keeps their search crawlers out, or the page says noindex or noarchive, these agents cannot read the site. Keeping them out is your choice; the consequence is that they cannot read you.
How to fix
Allow OAI-SearchBot, PerplexityBot, Bingbot, Googlebot and Claude-SearchBot in robots.txt, and remove noindex, noarchive, nosnippet and nocache from key pages.
Check it yourself
curl -s https://shop.lab.webforagents.dev/robots.txt; curl -sI https://shop.lab.webforagents.dev/ | grep -i x-robots-tag; curl -s https://shop.lab.webforagents.dev/ | grep -io '<meta[^>]*name="\(robots\|googlebot\|bingbot\)"[^>]*>'F11Key facts not only in the head or structured data
BlocksWhole-page fetchersWindow readers
What we found
- No page has JSON-LD structured data, so this check does not apply.
Why it matters
When fetch agents convert a page to text they drop the head and every script, JSON-LD included. A price, phone number or address written only in structured data never reaches them.
How to fix
Keep the structured data, and also write prices, phone numbers, addresses and product names in the visible text.
Check it yourself
curl -s https://shop.lab.webforagents.dev/ | grep -o '<script type="application/ld+json">[^<]*' | headF12Not too many links before the main content
BlocksWindow readers
What we found
- All 1 pages have at most 19 links before the main content.
Why it matters
Agents that read a page in windows, such as ChatGPT, see a few thousand characters at a time, navigation first. With fewer than 20 links the content fills about 78% of a window; with 60 or more, about a third. The bands come from a third-party measurement and are to be retested.
Check it yourself
curl -s https://shop.lab.webforagents.dev/ | sed -n '1,/<h1\|<main/p' | grep -o '<a [^>]*href=' | wc -l
robots.txt and AI agents
Whether each agent is allowed to fetch your homepage and pricing page. Only user-triggered agents that honour robots.txt are scored.
F8robots.txt and user-triggered agents
BlocksClaude appsDuckDuckGo
What we found
- No robots.txt: all agents are allowed.
- robots.txt allows the user-triggered agents that honour it.
Why it matters
robots.txt is the one access standard every major crawler honours; Claude-User honours it even for fetches a user asks for.
Check it yourself
curl -s https://shop.lab.webforagents.dev/robots.txtEvidence
- 404https://shop.lab.webforagents.dev/robots.txtrobots.txt
| Agent | Vendor | Homepage | Pricing page |
|---|---|---|---|
| User-triggered agents | |||
Claude-UserScored | Anthropic | Allowed | Allowed |
ChatGPT-User | OpenAI | Allowed | Allowed |
Perplexity-User | Perplexity | Allowed | Allowed |
Google-Agent | Allowed | Allowed | |
Meta-ExternalFetcher | Meta | Allowed | Allowed |
Amzn-User | Amazon | Allowed | Allowed |
MistralAI-User | Mistral | Allowed | Allowed |
DuckAssistBotScored | DuckDuckGo | Allowed | Allowed |
| Search crawlers | |||
Claude-SearchBot | Anthropic | Allowed | Allowed |
OAI-SearchBot | OpenAI | Allowed | Allowed |
PerplexityBot | Perplexity | Allowed | Allowed |
Googlebot | Allowed | Allowed | |
Bingbot | Microsoft | Allowed | Allowed |
Applebot | Apple | Allowed | Allowed |
Amzn-SearchBot | Amazon | Allowed | Allowed |
Meta-WebIndexer | Meta | Allowed | Allowed |
MistralAI-Index | Mistral | Allowed | Allowed |
| Training crawlers | |||
ClaudeBot | Anthropic | Allowed | Allowed |
GPTBot | OpenAI | Allowed | Allowed |
Google-Extended | Allowed | Allowed | |
Applebot-Extended | Apple | Allowed | Allowed |
Meta-ExternalAgent | Meta | Allowed | Allowed |
Amazonbot | Amazon | Allowed | Allowed |
CCBot | Common Crawl | Allowed | Allowed |
Bytespider | ByteDance | Allowed | Allowed |
Declared
Files and metadata a site can publish for agents. Shown for reference; they don't affect the score.
- A2A agent cardAbsentNo A2A agent card.
- Agent Skills indexAbsentNo Agent Skills index.
- agents.jsonAbsentNo /.well-known/agents.json.
- AI catalogAbsentNo AI catalog.
- API catalog (RFC 9727)AbsentNo API catalog (RFC 9727).
- Agent Resource DiscoveryAbsentNo Agent Resource Discovery file.
- Content usage preferencesAbsentNo Content Signals or AIPREF usage preferences.
- JSON-LD structured dataAbsentNo JSON-LD structured data.
- llms.txtAbsentNo llms.txt.
- MCP server cardAbsentNo MCP server card at a well-known path (current discovery goes through the AI catalog).
- NLWebAbsentNo NLWeb endpoint found. NLWeb has no discovery file, so it is found only when the site links to it.
- OAuth / OpenID metadataAbsentNo OAuth or OpenID authorization server metadata.
- OAuth protected resource metadata (RFC 9728)AbsentNo OAuth protected resource metadata (RFC 9728).
- SitemapAbsentNo sitemap.xml.
- WebMCPAbsentNo WebMCP tools.