The 19 site readiness signals every brand needs.

BeCited classifies AI bots as training (GPTBot, Google-Extended, ClaudeBot, CCBot) versus retrieval (OAI-SearchBot, ChatGPT-User, PerplexityBot, Claude-SearchBot). Blocking training bots is a defensible policy choice. Blocking retrieval bots makes you invisible to AI search.

The most common anti-pattern: brands block GPTBot to "protect" content but accidentally block OAI-SearchBot too. Per a BuzzStream analysis, 71% of sites have this misconfiguration.

An emerging standard for handing LLMs a curated index of your most quotable pages. No major answer engine reads it for search today. Its value is in agents and coding tools that read it directly, so the grade weighted it low, as cheap insurance.

The classic sitemap is still load-bearing for AI crawlers. We check it exists, that it is reachable from robots.txt, and that lastmod values are filled in, since those dates feed the freshness check below.

The check looked for topic clusters (pages grouped under shared URL paths), readable URL slugs that match the page’s subject instead of opaque IDs, and a sensible priority order in the sitemap. AI retrieval rewards sites with clear hierarchy; orphan pages and flat URL structures lose ranking against well-clustered competitors.

A forward-looking check based on the IETF Web Bot Auth draft (Cloudflare-led, 2025), which lets a site verify which AI bot is at the door. It asked two things: can the CDN or firewall in front of your site verify signed bot requests, and which AI companies publish their signing keys. The standard is still early, so the check warned but never failed a site.

The biggest single item in our earlier site grade. Independent studies put the citation uplift from proper structured data at 1.8–3.2x. The check looked for type-appropriate schema (LocalBusiness or Service for local; SoftwareApplication for SaaS; Product for consumer goods) plus Organization, FAQPage, AggregateRating, and Review where relevant.

Title, description, og:title, og:description, og:image. AI engines often quote the meta description word for word when the page is cited.

One H1 per page, descriptive H2s, and a consistent hierarchy. Headings phrased as questions tend to earn 40% more citations because they map directly to user prompts.

Self-contained 50–150-word chunks with answer-first structure. AI engines pull blocks, not paragraphs. A page full of conversational prose with no extractable claims will lose to a competitor with one well-formed answer block.

Use main, article, section, lists and tables instead of div soup. Engines parse semantic elements faster and trust them more. We count semantic tags vs unstructured div containers and flag the imbalance.

FAQPage schema, native HTML <details> elements, or Q&A-style headings. FAQs are over-represented in AI citations because their structure aligns with the prompt-and-answer pattern of generative search.

Person/author schema, byline markup, credential language ("certified", "licensed", "N years of experience"). Per AI Overview research, 96% of cited pages have strong E-E-A-T signals. A position-6 page with E-E-A-T markup beats a position-1 page without.

Last-Modified HTTP headers, JSON-LD dateModified, and sitemap lastmod. Perplexity weights content under 30 days old at roughly 3.2x. Half of all AI citations come from content less than 11 months old.

A composite of paragraph length distribution, answer-first pattern detection, statistic density, and self-contained chunk count. Engines reward content that is easy to lift, and this score measured how easy yours is.

The Princeton GEO paper identifies originality as the primary differentiator between content that gets cited and content that gets filtered out as paraphrase. We measure first-person research markers ("in our research", "we tested"), original data tables, and case-study evidence, and we discount paraphrase markers ("according to", "studies show") that signal derivative reporting. Pages with high information gain dominate AI citations even when their authority signals are weaker than competitors.

The check sorted every script tag (render-blocking, deferred, async, inline, JSON-LD), looked for empty app shells (<div id="root"></div>), confirmed that critical content (an H1 and at least 300 body words) is in the initial HTML, and measured image alt-text coverage. An optional second fetch rendered the page with JavaScript and compared word counts, to catch pages whose static HTML looks fine while the visible content is added by JavaScript.

Wikipedia presence, Wikidata entry, Organization schema with sameAs links to authoritative profiles, and consistent brand naming across the web. AI engines are reluctant to recommend entities they cannot disambiguate.

AGENTS.md, OpenAPI spec, public API documentation, and MCP manifest. As Anthropic's Model Context Protocol, Google's UCP, and Visa's Agentic Ready standards mature, this tells AI engines that agents can call your product directly.

BeCited's earlier site check measured the three Core Web Vitals with Google PageSpeed Insights: Largest Contentful Paint (LCP), Interaction to Next Paint (INP), and Cumulative Layout Shift (CLS). Pass thresholds (per Google web.dev) are LCP under 2.5s, INP under 200ms, and CLS under 0.1. Warn thresholds are 4s, 500ms, and 0.25; above those, fail.