Personal assistants like Muse, Instinct and Dot now book, buy and sign up on a person’s behalf. Agents act on sites more than ever, and that changed what a site needs for them to succeed.
Assistants like Muse open your site in a real browser and work through it for the person, which makes WebMCP essential: it lets your page hand the agent its actions as tools, so the agent calls them directly and skips clicking through the interface. Agents also act through an MCP server, apps that render inside the chat, and the docs they read to learn your product. Each of those protocols has moved fast. The latest WebMCP draft added hints that tell an agent when a tool does something it cannot undo. MCP servers now carry dozens of tools, and agents choose among them by description alone. MCP Apps put interactive views straight into the conversation. A presence check passes almost every site that adopts them. Far fewer get right the details agents depend on.
So we revisited what makes a site usable for an agent. We worked with leading AI labs on which details their agents rely on, and built Ranker 3.0 to grade each protocol on them. The layers and their weights stay the same; what earns the points inside them changes.
- WebMCP
- 2.0Passes when registerTool appears in the page
- 3.0Registered tools parsed and graded against the latest draft
- MCP
- 2.0Server found; tool quality as optional extras
- 3.0Every tool graded: descriptions, parameters, hints, output schemas
- MCP Apps
- 2.0At least one tool declares a view
- 3.0Every view a tool links to is actually served
- Pay
- 2.0Any payment protocol found earns the whole layer
- 3.0Graded on whether an agent can finish the payment
- Docs
- 2.0Homepage only
- 3.0Sampled docs pages: rendering, size, real 404s
- Checks
- 2.0Everything counts, extras only add
- 3.08 retired, about 20 shown but not scored
- Hard requirements
- 2.0None
- 3.0A bot wall, blocked agent crawlers or content that needs JavaScript caps the score at 59
- Scan time
- 2.0About 11 seconds
- 3.0About 11 seconds
What happens to your score
Most scores go down, because the bar moved with the protocols. Among the 22,492 sites scoring 60 or more on 2.0, the typical score moves from 67 to 48. Sites with a working MCP server and documented APIs lose about 6 points. Sites with neither lose about 26.
2.0
2% at 90+
3.0
1% at 90+
These numbers rescore the checks we already hold for every site. To test them, we rescanned 89 of those sites live with both versions. The projected change matched the real change to within a point at the median, and 3.0 scans took the same time as 2.0.
Layer by layer, Use moves the most. That is where MCP, WebMCP and APIs are scored, and all three are now graded on depth. Access goes up, because the checks we retired there were ones most sites failed. Pay still applies only to sites that take payments, and it now checks whether an agent can finish the purchase.
Discovery
68% → 61%
-7 pts
Access
72% → 77%
+5 pts
Use
66% → 25%
-41 pts
Pay
1,121 sites with payments
100% → 75%
-25 pts
An agent calls your tools and acts on what they return. 3.0 measures how well that works.
What 3.0 checks, protocol by protocol
WebMCP. The latest draft added consequentialHint, which marks a purchase, a booking or a payment so the agent confirms with the person before running it, and untrustedContentHint, which flags tools that return text other people wrote. 3.0 reads the tools you register out of your page’s JavaScript and grades them against that draft in ten checks, each its own line in your report: valid and descriptive schemas, clear names, no placeholder tools, annotations that match what each tool does, and the two new hints where they belong. As more assistants browse for the people they work for, these are the signals an agent in the browser uses to decide whether to act or to ask first.
MCP. An agent picks a tool from its name and description, and carries the whole tool list in its context on every turn. 3.0 grades each tool: a description that says when to use it, documented parameters, read and destructive hints that match what the tool does, and an output schema, which only about a third of servers declare. It also flags near duplicate tools and tool lists too large to load comfortably. A vague tool gets called wrong, and a bloated list crowds out the task.
MCP Apps. Every tool that renders a view must link to a view the server actually serves, and its visibility setting must use the values hosts understand. A broken link means the app never appears in the chat.
Pay. Muse and Instinct pay with Link by Stripe. On Stripe’s checkout pages the agent finds WebMCP tools that read the order and pay with a Link token, so it never types a card number. 3.0 credits Shop Pay, Google Pay and Link when a store’s UCP profile declares them on a checkout an agent can reach; for Link that is Stripe’s payment handler with Link turned on. Credit for checkout on Stripe comes next: on most sites it sits behind signup, where a scan of public pages can’t reach it yet. For agent payment protocols, finding the file is no longer enough. A UCP profile must declare checkout, a payment handler and a checkout an agent can reach. An ACP seller must publish the discovery file the April spec added and answer at its checkout endpoint. An MPP challenge that accepts Stripe lets agents that pay with Link pay you directly. AP2 no longer scores on its own: it approves a purchase but moves no money, so a UCP profile that declares it earns nothing for it. One working way to pay is still enough.
Docs. Agents read your docs more than any other page. 3.0 samples them and checks three things that break agents there: content that only renders with JavaScript, pages too large for an agent’s context, and missing pages that answer 200 instead of 404, so the agent reads an error page as an answer. An llms.txt over 100,000 characters now fails too, because agents cut it off.
What we retired, and what is now required
Eight checks scored things that do not change what an agent can do: batch endpoints, a Wikipedia page, sitemap dates, how many schema types you publish. They are retired. About twenty more, mostly draft standards and new discovery files, move to emerging: shown on your scorecard, not counted, until agents rely on them. A few extra checks, such as MCP output schemas, add points when they earn them and never cost any.
Two failures now cap the score at 59: blocking agents, with a bot wall or by turning away agent crawlers, and content that only appears with JavaScript. The AgentReady standard lists both as requirements. An agent that is turned away or sees an empty page cannot do anything else.
Where the points are now
Give your MCP tools descriptions that say when to use them, document every parameter, and mark which tools only read. Register WebMCP tools that map to real actions on your site, and mark the ones that spend money or cannot be undone. Serve your content and your docs without JavaScript. Publish an OpenAPI spec for an API that answers. If you sell, let an agent reach a checkout it can pay at, and declare how it pays.
Run a fresh scan at /#scan. The full check list and weights are at /methodology, and every check still shows the evidence that decided it.