Exa Snapshot Lets AI Search the Web at Specific Past Dates

Search API startup Exa has introduced Snapshot, a feature that allows developers to query and retrieve web pages as they existed at a specific point in time. By indexing 400 billion webpage versions spanning two decades, the tool addresses a major issue in AI development: web leakage. When AI agents are evaluated on benchmarks, they often access modern search results containing solutions published after the benchmark was created, which contaminates the evaluation.

Developers can access this historical data by adding the snapshotAsOf parameter to raw JSON requests, or snapshot_as_of in the Python SDK, targeting the existing /search or /contents endpoints. The /search endpoint discovers past URLs, while /contents retrieves the archived text, titles, authors, and summaries. However, the feature only restricts the content itself, not the ranking algorithm, meaning historical search results may not perfectly reconstruct a past search engine results page.

The preview is available with a pay-as-you-go tier that offers 10 queries per second and a rolling five-month archive window. Users get 100 free requests before needing to contact sales for extended access. The feature supports Exas auto, fast, and instant search modes, but deep-lite, deep, and deep-reasoning are currently unavailable. Additionally, developers cannot combine the snapshot parameter with livecrawl, livecrawlTimeout, maxAgeHours, or subpages without triggering an INVALID_REQUEST error.

Beyond preventing benchmark contamination in reinforcement learning, Exa expects Snapshot to aid quantitative financial backtesting, where researchers require point-in-time web data to match historical stock prices. It also enables policy audits, allowing teams to run diffs on documentation or pricing pages across different dates, and lets developers test historical agents against strict content boundaries.