Exa launches a historical web index with 400 billion snapshots

• A historical query returns Exa's newest stored page version at or before its timestamp, rather than a current revision hidden behind a date filter, as WilliamBryk explained and the Snapshot documentation specifies.
• One snapshotAsOf field pins both search and content extraction to a datetime. Exa's Snapshot announcement describes the 400 billion-snapshot index as search "as if it's the past."
• The feature targets two leakage-sensitive workloads: RL agents whose answers may have appeared online after a task cutoff, and web-data backtests, according to WilliamBryk's launch explanation and the official launch post.
• Access is a research preview, with 10 QPS and a rolling five-month window on pay-as-you-go. Exa's availability post says Snapshot is available to try now.

The launch post puts snapshotAsOf in different places for /search and /contents. The API docs add two less obvious boundaries: content is time-pinned but ranking is not, and a URL without an eligible stored copy comes back as CONTENT_NOT_CACHED.

Exa says Snapshot holds more than 400 billion webpage snapshots across two decades. WilliamBryk's explanation identifies the failure mode it is meant to avoid: a normal date filter can still expose a present-day revision of a page that changed after the requested date.

The documented cutoff is a stored version, not a page publication date. For an eligible URL, the docs say title, author, publication date, text, highlights, and summaries all come from the selected historical copy.

The request shape differs by endpoint:

• /contents: put snapshotAsOf at the top level beside ids and text.
• /search: put snapshotAsOf inside contents, beside options such as highlights.

Both endpoints select the newest stored version at or before the supplied datetime, per the endpoint examples. Exa's Snapshot announcement says the index is already being used for prediction-model backtesting, RL work, and exploration of the pre-AI web.

A Snapshot search is not a reconstruction of the ranked results a user would have received on that date. The docs say Exa uses current retrieval signals to discover candidate URLs, then retains only candidates with a stored version at or before snapshotAsOf.

That leaves the returned page content bounded by the requested time while the candidate ordering remains current. Exa's Snapshot announcement presents the feature as a way to search the past, but its documentation defines the narrower retrieval semantics.

The public preview exposes only a rolling five months of the index despite the much larger historical store described in the launch post. Exa's availability post marks the feature as available now, while the docs send users seeking full index coverage, rate limits beyond the public tier, or ZDR details to Exa's team.

The documented limits are concrete:

• Pay-as-you-go allows 10 QPS; requests more than five months old are rejected, and continued use after 100 requests requires contacting sales.
• Historical requests reject livecrawl, livecrawlTimeout, maxAgeHours, and subpages, since each can reach beyond stored content.
• Snapshot search supports auto, fast, and instant, but not deep-lite, deep, deep-reasoning, or category.