How the AI Readiness Score works
Five observable website pillars add up to 100 points. Each pillar uses public website evidence and clearly defined criteria, so cloud partners can understand their score and identify the improvements that matter most.
| Pillar | Weight | Primary question |
|---|---|---|
| Search access | 20 points | Can answer-engine crawlers fetch, index, and quote the page? |
| Server-delivered content | 25 points | Does useful page content exist in the fetched HTML before client-side rendering? |
| Extractable structure | 20 points | Can a parser segment the page into meaningful, quotable sections? |
| Entity clarity | 20 points | Do metadata, structured data, and visible copy agree on who the organization is? |
| Site discovery | 15 points | Can a crawler move from the homepage to the rest of the useful site? |
Search access
Can answer-engine crawlers fetch, index, and quote the page?
Server-delivered content
Does useful page content exist in the fetched HTML before client-side rendering?
Extractable structure
Can a parser segment the page into meaningful, quotable sections?
Entity clarity
Do metadata, structured data, and visible copy agree on who the organization is?
Site discovery
Can a crawler move from the homepage to the rest of the useful site?
Hard caps
A page cannot earn a high readiness score when an answer engine cannot index it or the useful copy is absent from the fetched HTML. These conditions cap the total at 40, even if other checks pass.
The page tells search systems not to index it.
The robots policy blocks OpenAI's documented search crawler.
The fetched HTML contains fewer than 50 visible words.
Search bots and training bots are different
The rank uses answer-engine access signals. Policies for GPTBot, ClaudeBot, and Google-Extended are recorded for transparency but add zero points. A company can restrict model training without losing leaderboard points.
Benchmark process
- 1. Define the cohort. Capture public AWS Partner Finder profiles and separately labeled AWS Marketplace seller profiles, then deduplicate organization websites by domain and normalized legal name.
- 2. Fetch public evidence. Request the listed website, robots.txt, sitemap.xml, and llms.txt with timeouts and retries.
- 3. Apply the formula. Parse the raw response and assign points from the rulebook above.
- 4. Publish the measurement date. Keep unreachable sites visible but unranked so a network failure never becomes a zero.