The Pulse
Exa Launches Snapshot to Search the Web as It Once Existed
Exa’s Snapshot research preview lets developers search and retrieve web content as it existed on a specified date, supporting temporal AI evaluations and financial backtests.
AI.info Team ·
Exa Adds a Time Dimension to Web Search
Exa launches Snapshot on September 17, 2026, giving developers a way to search the web as it appeared on a selected date. The research preview draws on more than 400 billion webpage snapshots collected across two decades.
“There's growing demand to test models in more complex environments and on increasingly difficult tasks.”— The Exa Team, Exa blog authors
The company says Snapshot is designed for temporal testing: researchers can ask whether an AI system could have found an answer using only information that existed at the time. Instead of searching the current web and filtering results afterward, developers provide a date and receive results from that historical point.
Exa exposes the capability through its existing /search and /contents API endpoints. The date is supplied through a snapshotAsOf parameter, which tells the service when the search or page content should be evaluated.
Preventing Web Leakage in Agent Training
Exa presents reinforcement-learning evaluations as one of Snapshot’s main uses. A research team may write tasks and answers in June, then train an agent in September. By that point, answers or closely related solutions may have appeared in a research paper, a GitHub pull request, or a blog post.
An agent with access to live search could retrieve that material instead of solving the task independently. Exa says researchers can set Snapshot to a date before the solutions were published, reducing the chance that an agent benefits from information that was unavailable when the evaluation was designed.
The approach also supports more reproducible comparisons between systems. Every agent can be tested against a defined version of the web rather than a changing collection of pages, although Exa describes the product as an early research preview with ongoing development.
Historical Web Data for Financial Backtests
Exa also points to quantitative finance as a second application. Financial backtesting asks how a prediction model would have performed using only information available on a particular day. Existing point-in-time datasets commonly track prices and company fundamentals, but web pages introduce a separate problem: their contents change, disappear, or become available only after the period being tested.
Snapshot gives researchers a versioned web index for those experiments. A trading system could retrieve news, company updates, or other web material as it existed at a specified date, then measure its signals without accidentally importing later information.
Exa says collecting that material manually would otherwise require months of work. The launch post does not publish performance results for financial strategies, so Snapshot’s value in trading remains a use case proposed by the company rather than an independently demonstrated result.
How Developers Use Exa Snapshot
The company’s Python example searches for “latest stable Python release notes” while setting the snapshot date to May 1, 2026. A second example requests the historical contents of a specific Python documentation page using the same date-based control.
Snapshot can be accessed through Exa’s dashboard by selecting a date, or through the API for programmatic evaluation workflows. Exa says developers seeking full index coverage, higher rate limits, or zero-data-retention arrangements should contact the company.
The launch places historical retrieval alongside Exa’s existing live search and content-extraction services. Its practical distinction is not simply finding old pages; it is allowing an AI system, evaluator, or data pipeline to operate against a defined past state of the web.
A Search Index Built for Reproducibility
AI evaluations often depend on information that changes after a benchmark is created. A live search result can make a test easier over time, while a deleted page can make the same test impossible to repeat. Snapshot addresses that problem by tying retrieval to a date rather than leaving the web state implicit.
Exa has not disclosed the full composition of the 400 billion snapshots, its retention policy for every page, or a detailed coverage breakdown by domain and date. Those details will matter to researchers judging whether a historical result represents the available web broadly enough for a particular experiment.
For now, the product is a research preview with a clear interface: choose a date, send a search or content request, and receive results constrained to that point in time.