Every rating is out of 100. Measured evidence decides where it lands.
Agent Readiness is a 0–100 rating computed from a real crawl of your site, not a static checklist — Estimated from passive evidence, or Verified once an agent journey has run. It answers one question: can an AI agent actually find, understand, and act on this page?
Agent Readiness and Technical Readiness
Your headline number is Agent Readiness — labelled Estimated or Verified by the evidence behind it: how well an AI agent can find, read and complete a task on your site, including our agent simulations when they ran.
Beside it we show Technical Readiness: the passive checks alone — the same kind of evidence the open-source Scovant Core collects, so the two numbers stay comparable.
Each category counts only the checks we could actually measure. When too little could be measured, a scan shows Not enough evidence instead of a number; when the number rests on partial evidence, it carries an incomplete coverage note. When the cause is the site itself we say which: site refused our fetch when its firewall answered our scanner with an access refusal (for example HTTP 403) on every page, site returned errors when its pages answered with errors (404 or 5xx). The rating then covers what we could measure.
One R2 methodology, evaluated at different evidence levels: Technical Readiness from the open passive checks; Estimated Agent Readiness when only passive evidence exists; Verified Agent Readiness once at least one agent journey has run on the site. Every result says which it is (readiness.agent.evidence_state).
When every page we fetched was a bot challenge, public pages show Automated access refused — rating not computed instead of a number.
Rules apply by capability.
Rules apply by capability. Product, offer and price rules apply only to a site with a product catalog; shipping and returns only to a site that sells physical goods; checkout rules only to a site that checks out. Scovant sets a capability when a scanned page shows it (a Product with an offer, an add-to-cart control, shipping details in the schema, a declared checkout) and otherwise uses the default for the site type — for example, a hotel or restaurant has no physical goods. When neither is known, those rules are marked not applicable rather than failed. On a SaaS site an add-to-cart or "buy now" control sells a licence or a subscription, so only Product data with an offer establishes a catalog there.
| Site type | Product catalog | Physical goods | Checkout |
|---|---|---|---|
| commerce | yes | yes | yes |
| booking | no | no | yes |
| restaurant | no | no | yes |
| saas | no | no | not established |
| blog | no | no | no |
| news | no | no | no |
| docs | no | no | no |
| documentation | no | no | no |
| media | no | no | no |
| portfolio | no | no | no |
| education | no | no | no |
| real_estate | no | no | no |
| other, unknown, any unlisted type | not established | not established | not established |
Every model change since R2 became the headline score is listed in the methodology changelog, with its model ids.
R2 methodology details
The categories, weights and severity point values below are the R2 public policy. Technical Readiness is reproducible from this open formula and the passive evidence; the frozen R1 model is described in the appendix at the end of this page.
Formula R2.1 (Technical model R2.1@7b46d2ed7f8b). Each rule outcome scores PASS 1, WARN 0.5 or FAIL 0. Within a category, rules count with their severity weight: 40 (critical), 25 (high), 14 (medium), 6 (low) or 2 (info).
Category rating = 100 × (Σ weight × value + prior) ÷ (Σ weight + prior), over the measured rules of that category (a page rule's value is 1 − failed share × (1 − outcome)), where the prior is 14 (category_prior_weight). Category prior is a smoothing constant, not evidence. It affects the category rating but does not increase coverage. The prior is added only when at least one weighted rule in the category is measured; a category with nothing measured has no score.
| Category | Weight |
|---|---|
| discoverability | 17% |
| structured | 18% |
| consistency | 15% |
| actionability | 12% |
| trust | 8% |
| ucp | 12% |
| citability | 18% |
UCP applies to eligible transactional profiles. On an applicable site, an absent profile is a measured absence; on a non-applicable site the category is excluded and the remaining category weights renormalize. Citability applies to blog, docs, documentation, media, news, other, portfolio sites.
Coverage floors: a result needs at least 60% measured weight, otherwise it is Not enough evidence; below 85% the rating carries no grade or badge. The overall rating is the weight-averaged category rating over the categories that have a score. Grades: A from 90, B from 80, C from 60, D from 0. Badges: agent_ready from 80, agent_compatible from 60. Published Technical gate: agent_access_blocked → badge_ineligible.
One pipeline, four steps.
What we check, and why it matters to an agent.
Only the categories that apply to your site type count toward your overall score, and the weights are public — the categories, weights, rules and severities are published (see the full rule catalog), so Technical Readiness is reproducible from the open R2 formula. Where a category doesn't apply, it is excluded and the remaining weights renormalize.
- Robots.txt allows AI crawler access
- llms.txt published
- Sitemap.xml present and valid
- Pages render as clean markdown for agents
- schema.org Product, Offer & availability markup
- Sensible heading structure
- Page weight an agent can parse quickly
- Displayed price matches the price in markup
- One canonical entity per page, not duplicates
- Schema blocks on the same page tell one story
- Key actions reachable without a JS-only wall
- Stable deep links to product, cart & checkout
- A declared machine interface (e.g. MCP) for known tasks
- Return policy easy to find
- Shipping costs & timelines published pre-checkout
- Clean citation metadata (OpenGraph)
- /.well-known/ucp profile published
- Profile schema is valid
- Checkout capability declared
- Self-contained passages
- Answer-first structure
- Concrete figures over vague claims
Not every issue costs the same.
The severity weights are part of the open methodology: inside its category each rule counts with weight 40 (critical), 25 (high), 14 (medium), 6 (low) or 2 (info), and its measured outcome scores PASS 1, WARN 0.5 or FAIL 0; a rule that does not apply is left out, and one we could not measure lowers coverage instead of the rating. The two evaluations below differ in what may change the result.
Technical Readiness. Technical Readiness may be ineligible for a badge under a published badge-block condition. The public Core policy does not apply a hidden numeric cap to Technical Readiness.
Cloud Agent Readiness. Cloud Agent Readiness may be changed by a proprietary calibrated overlay. Applied conditions are explained, while thresholds and numeric effects remain private.
Every Scovant score is explainable: the categories, their weights, every rule in the public catalog, its severity, and the severity point values above are all published — enough to understand any score, act on any finding, and cite a result like a benchmark. Every scan also records the exact scoring-model version that produced it, so a score change from a model update is never mistaken for a website regression. 32 of the 66 rules are experimental: their findings are reported, but they contribute zero points until promoted. The Agentic Security & Trust checks are reported alongside the score and never change it.
Technical Readiness is reproducible from the open R2 formula and Core evidence. Cloud Agent Readiness uses the same public rule taxonomy plus observed execution and a proprietary calibrated overlay. Every applied gate is explained, but the overlay is not fully reproducible outside Cloud. What we deliberately keep private is that overlay's calibration — its gate thresholds and the anti-gaming heuristics that keep checklist tricks from inflating a rating — so that a high Scovant rating keeps meaning what it claims to mean.
Two thresholds, evidence-qualified.
Scores also map to a grade — A at 90 or higher, B at the top badge (80), C at Agent Compatible (60), D below — the same ladder your scan reports show.
For commerce-tier sites the badge also depends on the real agent journey: when too few of the agents we run complete the purchase, the badge is withheld whatever the number says. A badge means the rating and the real agent journeys agree.
For commerce-tier sites, we don't just read the markup.
For commerce-tier sites (online stores, booking services, restaurants), we run several AI agents that each attempt the purchase journey in a live browser — finding a product, adding it to the cart, and working through checkout — stopping just before payment. They never enter real payment details or create an account. How many of them complete the journey feeds directly into your score: when the agents succeed it supports the Actionability category, and when too few can complete the purchase the overall score drops and the badge is withheld, as described above. Readiness is a pass-rate over the agents you ran (75% or more is ready). A failing agent is a note, not a penalty, as long as at least 75% of the agents you ran complete the purchase; with two or three agents a single failure drops readiness to partial.
Legacy R1 methodology (frozen)
R1 was the headline model until 2026-10-02 and is frozen. It used the same seven categories, weights and severity point values, subtracted findings from 100 into one number, capped that number when a site blocked AI agents or when purchase simulations failed, and added an Actionability bonus when a purchase simulation passed. R1 results stay as immutable history and are never compared with R2 results.
Common questions.
My score changed and I didn't touch the site — what happened?
How do I actually raise my score?
How often should I rescan?
Why does a category on my report show N/A?
See where your site stands.
Run a scan and get your estimated or verified Agent Readiness result, category by category, with the exact issues to fix.