Access — can they fetch you at all?
- robots.txt names GPTBot, OAI-SearchBot, ChatGPT-User, PerplexityBot, Perplexity-User, ClaudeBot, Google-Extended, Applebot-Extended and Bytespider explicitly.
- No CDN, WAF or bot-management rule returns 403 or a challenge to those agents.
- Key pages return 200 with full HTML to a plain curl, no JavaScript required.
- No consent wall or interstitial blocks the main content server-side.
- sitemap.xml is current, reachable and referenced from robots.txt.
- /llms.txt exists and lists your canonical pages.
Extractability — can they lift the answer?
- Every page has exactly one H1 that states the topic as a question or claim.
- Each section opens with a 40–70 word self-contained answer paragraph.
- Headings are phrased the way people phrase prompts, not as nav labels.
- Entities are named explicitly rather than referenced by pronouns.
- Tables and lists carry the comparative facts; prose carries the nuance.
- Critical content is not hidden inside tabs or accordions that render client-side only.
Trust — will they attribute it to you?
- Organization schema sitewide with accurate sameAs profiles.
- Article or BlogPosting with author, datePublished and dateModified on content pages.
- FAQPage and HowTo where genuinely applicable, matching visible content.
- Named authors with real credentials and an author page.
- Outbound citations to primary sources for every statistic.
- Consistent boilerplate description across every off-site profile.
Freshness and measurement
- dateModified reflects real substantive edits, not a nightly build timestamp.
- Roundups, pricing and comparison pages are reviewed at least quarterly.
- Server logs are monitored for AI-bot hits as a leading indicator.
- A fixed prompt panel is run on a schedule across at least three engines.
- Competitor appearance is logged alongside your own.
- Findings are turned into one content or technical fix per cycle, not a backlog.
Work top to bottom. Access problems invalidate everything below them, extractability problems waste good schema, and measurement without the first three just documents the failure. Most sites we scan fail somewhere in the first two sections, which is good news — those are the cheapest fixes available.