RXed AI News

AI news with a spine.
@RXed_EU
LatestAgent VersusMethodologyUpcoming & requests

How we score agent readiness

Crawler-based checkers (Cloudflare's Agent Readiness score, AgentGrade, Fern) score one URL on what they can detect over HTTP. The ARI scores whether an AI agent can actually use a service end to end — including its terms, its pricing behaviour toward automation, and whether an agent can transact. Editorial, sourced, dated.

Ten criteria in three groups, 0–10 each with one decimal. The overall is the plain unweighted mean of the applicable criteria, shown as a whole number — computed by the generator, never set by hand. No weighting: weights are arguable, and the group scores already tell you where a service is strong.

Bands: ≥9 Agent-ready benchmark · ≥7 Strong · ≥5 Capable, with friction · below that, Agent-hostile.

Evidence bar: at least 4 sources per service, official documentation and terms mandatory, every source dated. Point-in-time assessment — services change terms, prices and APIs often. That volatility is itself scored under Reliability & continuity. Re-verified at least quarterly.

FIND & RETRIEVE

Can an agent get correct information?

Sf — Machine-readable surface

Is the content reachable and parseable by a machine — clean HTML or markdown negotiation, sitemap, schema.org, feeds, llms.txt where useful?

10: Structured, negotiable content: sitemap, schema.org, RSS/JSON feeds, markdown variants, no bot wall on public pages.
2: JS-only rendering, CAPTCHA on public pages, facts only inside PDFs or images.

Iq — Information retrieval quality

Are the facts an agent needs actually published and current — prices, availability, hours, stock, terms — rather than gated or stale?

10: Live, specific, dated facts on the public surface; an agent can answer a customer question without guessing.
2: Real answers behind 'contact sales', stale prices, marketing copy where data should be.

Dx — Documentation for machines

Can an agent or its developer find auth, endpoints, limits and error semantics without a sales call?

10: Executable examples, per-field reference, changelog, errors documented as a taxonomy.
2: Docs behind a login, no changelog, documentation drifted from reality.
N/A rule: N/A only where the service exposes no programmatic surface at all AND scores accordingly on B1/B2.
ACT & TRANSACT

Can an agent do the job?

Ac — Action coverage

Does a public API cover the service's core actions, measured against what a human can do in the UI?

10: Every meaningful UI action has an endpoint; test mode mirrors production.
2: Read-only API, or the core action (book, order, invoice, pay) is UI-only.

Pt — Agent protocol support

Official MCP server, WebMCP, Agent Skills, API catalog or OAuth discovery. Community-maintained MCP counts at half credit.

10: Official, maintained agent protocol surface (MCP/WebMCP/Agent Skills) that covers the service's CORE actions, with scoped auth and discoverable capability metadata.
2: Nothing official; agent access exists only via scraping or third-party wrappers.
N/A rule: Never N/A — absence is a score of 0–2.

Au — Access & auth friction

Self-serve credentials? Agent-compatible OAuth? Scoped/restricted keys? Or manual review and enterprise gates?

10: Keys in minutes, restricted/scoped credentials, OAuth flows an agent can complete.
2: Partner programme, manual approval or enterprise contract before the first call.

Py — Agent payment capability

Can an agent transact — pay the service, or pay/check out on the customer's behalf — under delegated, spend-capped authority? Checked against ACP, AP2, MPP, x402 and documented programmatic checkout.

10: Supports a published agent payment protocol (ACP / AP2 / MPP / x402) or a documented programmatic checkout with delegated, spend-limited credentials and an audit trail.
2: Money only moves when a human clicks a hosted payment page.
N/A rule: N/A ONLY if the service never takes money from a customer at all. 'No agent payment path exists' is a LOW SCORE, not N/A.

Cf — Cost & rate fairness

Usable rate limits at accessible tiers; is agent traffic priced like human traffic or punished?

10: Documented, generous limits on paid-but-affordable tiers; no agent-vs-human price discrimination.
2: Agent traffic penalised — the reference case is X's 13x price premium on posts containing links, with replies locked behind a ~$42K Enterprise tier.
TRUST & DURABILITY

Will it still work next quarter?

Ts — Permission & ToS stance

Is automation explicitly permitted, silent, or prohibited — including enforcement history, bot rules, Web Bot Auth and Content Signals.

10: Automation explicitly welcomed, bot rules published and honoured, no hostile enforcement history.
2: Automation prohibited or selectively enforced; litigation history against agents and scrapers.

Rl — Reliability & continuity

Uptime and status transparency for programmatic consumers, plus the API/pricing deprecation track record over the last ~24 months.

10: Public status page with programmatic status feed, date-versioned API, years of backward compatibility.
2: Endpoints sunset with minimal notice, repeated pricing overhauls, no status transparency.
AGENT PAYMENT RAILS WE CHECK FOR
WHAT THIS IS NOT

It is not a crawl. Automated checkers report what a single URL exposes over HTTP — robots.txt, sitemaps, bot rules, capability manifests. Useful, and we read them. But no crawler tells you that replies moved behind a $42K tier, that a link in a post costs 13× more to publish, or that the checkout an agent needs exists but is granted case by case. Those answers need a human reading the terms, the pricing page and the changelog — and citing them.