Page content, hidden data and a saved view
According to Expand’s documentation, Iris reads the rendered page and produces the main content as Markdown, retaining tables and icons. Navigation and footers move to a separate appendix; ads and banners are dropped.
X-Ray handles data behind the page, such as prices, stock levels or coordinates held in API responses rather than visible text. The docs say it keeps useful data as State JSON while dropping analytics and configuration.
Expand also saves a snapshot in its Wayback Machine feature. The documentation says each line of the main Markdown and State JSON links back to its location in that snapshot, which can be replayed or shared after the original site changes.
What Expand’s benchmark measures
On its MDBench Hard v1 test, Expand reports 74.1% of questions answered correctly, compared with 55.3% for Exa and 27.2% for Firecrawl. The test covered 1,094 questions across 326 hard pages and used Claude Sonnet 5.5 to answer from each API’s Markdown at default settings.
The company says it curated the cases with customers to target extraction problems, including pricing and competitive-intelligence pages. It dates the test to October 8, 2026, and notes that Expand was the only tested provider whose output included hidden JSON through X-Ray. Those are results from Expand’s own selected test set, not a general accuracy rate for every website.
The listed base price is $1 per 1,000 successful fetches. Expand says failed, blocked and empty requests cost nothing; requesting Highlights alongside a fetch adds another credit.
