Expand

AI agents that read web pages via existing Markdown extraction tools lose crucial details during conversion, such as icons indicating checkmarks or x marks, leading to confidently wrong answers (as shown in the uptime SLA example where dropped icons caused a bot to say both plans had SLAs when only one did). Important data like coordinates, flight gates, customer stories and pricing details often live in hidden JSON the page loads rather than in visible HTML, and hard-coded conversion rules miss this entirely. Developers also face a tradeoff between sending agents the full page (wasting tokens on navigation, footers and promotional banners) or a trimmed version that may cut needed content. At scale, inefficient token usage is costly: the site calculates that 5 million requests a month on Gemini 4 could cost $200,000 versus $152,000 with Expand, a $48,000 monthly or $576,000 annual difference.

Expand offers an API and SDK (TypeScript, with Python coming) that fetches any URL and returns page.md (accurate Markdown via its vision model Iris), state.json (hidden JSON extracted by its model X-Ray), and a snapshot for citation, all in one call. Iris reads the rendered page like a human, preserving tables and icons while moving navigation and footers to a separate appendix and dropping cookie banners and promotional noise. X-Ray extracts meaningful data from a page's hidden JSON payloads without needing a browser, as demonstrated by pulling customer data from Linear's pricing page JSON. The Wayback Machine feature links every extracted line back to its source element for verification, Highlights lets agents query for only relevant passages to save over 95% of tokens, Crawl processes entire sites at one credit per page, and WebSync subscribes to a site to stream changes (pages added, changed, or removed) as a durable, resumable event feed synced to a local file system. The company also published a benchmark, MDBench Hard v1, claiming 74.1% accuracy versus competitors Exa (55.3%) and Firecrawl (27.2%) on 1,094 questions over 326 hard pages.