KB Health Checker

Paste a help-center article and get an AI-readiness grade — how well it will perform when an AI copilot (Intercom Fin, Maven AGI, a RAG bot) is answering customer questions from it.

Why this exists: AI support agents don't read your article top-to-bottom — they retrieve chunks of it and generate answers from whatever a chunk says on its own. An article that reads fine to a human can still fail as AI source material: buried answers, sections that only make sense in context, "click here" links, walls of text. This tool encodes those failure modes into a rubric you can run on any article before your copilot does.
100% client-side · no data sent anywhere
–/ 100
–

Findings

Methodology — what gets checked and why

Structure & chunkability (25%). Retrieval systems split articles into chunks, usually along headings. No headings, giant sections, and wall-of-text paragraphs mean chunks get split mid-thought — the AI retrieves half an answer.

Answer clarity (25%). Generation quality tracks how directly the source states the answer. Answer-first openings, task-shaped headings, and numbered steps give the model something quotable; marketing throat-clearing gives it filler.

Self-containment (20%). A chunk gets retrieved alone. Sections that open with "This" or "As mentioned above," links labeled "click here," and undefined acronyms all break when the surrounding context isn't retrieved with them.

Machine readability (15%). Semantic HTML (real headings, lists, tables, alt text) survives ingestion pipelines; div-soup, styling-as-structure, and screenshot-only content don't.

Language quality (15%). Long sentences, passive voice, and hedge words ("simply", "just") degrade both retrieval matching and the tone of generated answers.

All checks are heuristics computed locally in your browser — a fast first-pass audit, not a substitute for reading the article. Plain-text input is supported but HTML input enables the full rubric.