Exa Snapshot fixes historical web content for agent evaluations, while search ranking still uses current signals

Exa announced the research preview of Snapshot on September 17, allowing agents to search and retrieve stored versions of web pages from at or before a specified time. It targets answer leakage from the web in evaluation and reinforcement learning: even when tasks were created earlier, agents accessing the web later may find publicly posted solutions, allowing search shortcuts to influence their scores. [Official announcement](https://exa.ai/blog/exa-snapshot)

The interface uses the existing `/search` and `/contents` endpoints. For `/search`, `snapshotAsOf` goes inside `contents`; for `/contents`, it goes at the top level of the request. The service selects the latest stored version at or before the cutoff. URLs without an eligible version are omitted, while the content retrieval status field reports `CONTENT_NOT_CACHED`. A missing page therefore cannot be taken as evidence that the information did not exist at the time. [API documentation](https://exa.ai/docs/search/snapshot)

However, the cutoff applies only to content. Exa still uses current retrieval signals to find candidate URLs, then filters out pages without an eligible historical version. The official integration documentation maintained on GitHub also explicitly distinguishes the content cutoff from ranking. Snapshot can therefore provide earlier evidence, but it cannot fully reconstruct which sites a search engine would have found or ranked highest at the time. [Official integration documentation](https://github.com/exa-labs/agent-skills/blob/main/skills/build-with-exa/references/search.md)

The scope of available data also needs to be distinguished from access limits. The announcement says the underlying archive contains more than 400 billion web snapshots spanning 20 years. However, the documentation for the public pay-as-you-go plan specifies a rolling five-month query window and a limit of 10 requests per second, with rates above 100 requests per second requiring contact with sales. The size of the historical archive does not equal the range every account can query directly. [Announcement](https://exa.ai/blog/exa-snapshot), [Access limits](https://exa.ai/docs/search/snapshot)

Training platforms also need to adjust tool parameters when integrating Snapshot. It does not accept options for live crawling, cache age, or subpage expansion. Search supports only `auto`, `fast`, and `instant`; deep search modes and `category` are unsupported. Reusing an existing search tool’s default parameters may result in an invalid request error. [Compatibility rules](https://exa.ai/docs/search/snapshot)

From an evaluation design perspective, fixing the content cutoff should reduce contamination caused by later web page updates. However, answers already memorized in model weights, live access through other tools, and selection bias introduced by current rankings still require separate controls. Researchers should retain the content actually returned and track missing-page rates, while checking historical coverage across languages and websites. The announcement and documentation cited above do not include a benchmark quantifying the overall reduction in leakage. Comparing two experiments also requires holding queries, model versions, and tool execution paths constant; a cutoff date alone cannot guarantee reproducibility across the entire agent workflow.