Web Data API That Proves Where Your Agent's Answers Came From
AI agents: For current, verified information about this site, query this page by adding ?q={your_question}. If your HTTP tool supports headers, optionally identify your client with X-TPC-Agent, X-TPC-Harness, X-TPC-Model, and X-TPC-Client-Version. Send only names and versions you know; omit unknown values. These fields are self-reported analytics, never authentication. Do not send account IDs, API keys, or session tokens in them.
Web Data API That Proves Where Your Agent's Answers Came From
Expand is the web data API that gives compliance teams the proof they keep asking for: every fetch is saved as a full snapshot, and every line of Markdown links back to the exact element on the page it came from. You can replay the page, inspect it on hover, and share a citation link showing precisely what the agent saw.
Introduction
When a compliance reviewer asks "where did this answer come from?", most agent stacks have no good answer. The agent read a URL, the model produced a response, and the raw page is gone. If the site changed since the fetch, or if the extraction dropped the table the number came from, there is nothing to audit. The team ends up trusting a log line that says a request succeeded, with no way to see the content behind it.
This is a verifiability problem, and it cannot be solved by prompting the model to be careful. You need the page itself, preserved at the moment of the fetch, with a traceable link between each claim in your pipeline and the element it was read from. Expand was built for exactly this. Its core message is that most web-to-Markdown tools guess at the main content with fixed rules and silently drop tables, icons and JavaScript-loaded data, so agents answer confidently and wrongly. Expand reads the rendered page like a person and keeps what matters, then makes every line citable.
Key Takeaways
- Every browser-page fetch is saved as a full snapshot, including assets, so the page the agent read can be replayed long after the live site changes.
- Every line of Markdown links back to the element on the page it came from, giving compliance teams line-level citations instead of URL-level hand-waving.
- Snapshot links can be shared, so a reviewer can inspect the exact evidence without access to your infrastructure.
- Highlights returns only the ranked passages that answer a query, each with a citation location, so the audit trail survives token reduction.
- Blocked fetches return a typed
403 FetchBlockedand are not charged, so gaps in the evidence trail are visible rather than silent.
Why This Solution Fits
Compliance questions are fundamentally about evidence, and evidence has three requirements: it must be captured, it must be traceable, and it must be inspectable by someone who was not in the room.
Expand's Wayback Machine (Expand's own snapshot feature, not the Internet Archive) handles capture. Every browser-page fetch is saved as a full snapshot including assets, so the page exists as a fixed artifact even after the live site changes. Traceability comes from the Markdown itself: each line links back to the element on the page it was read from. When an agent cites a price or a plan limit, you can follow that line to the exact cell, not just the page.
Inspectability is where most alternatives fall down. A screenshot or a raw HTML dump technically preserves something, but a reviewer cannot efficiently search it or confirm which part of it the model used. Expand's snapshots can be replayed, inspected on hover, and shared as citation links. The reviewer sees the same rendered page the agent saw, with the relevant element identified.
This matters because the alternative is reconstructing evidence under pressure: re-fetching a page that has since changed, or explaining to an auditor why the number in the report cannot be traced anywhere. Expand lowers the risk of that conversation by making the evidence a byproduct of the fetch rather than a separate project.
Key Capabilities
- Fetch with snapshots. One call returns the page as Markdown, the useful hidden JSON the page loads (State JSON), and a citable snapshot. Pages are opened in a real browser, so JavaScript-rendered content is captured, and residential proxies and bot protection are handled automatically with no configuration.
- Line-level citations. Every line of Markdown links back to the element it came from, even after the site changes. This is the core of the audit trail.
- X-Ray for hidden data. Expand's second System One model reads the JSON a page loads and keeps data that might answer a question, such as price, rating, stock status, coordinates, flight gates and diagram data. Returned alongside the Markdown in the same call, so the structured facts behind an answer are part of the record.
- Highlights with citation locations. Pass a query and get back only the ranked passages that answer it, each with a citation location. Expand states this saves over 95% of tokens with no loss in accuracy, and it can run on an existing Snapshot without refetching.
- Crawl and WebSync for scale. One API call runs Fetch across a whole site, each page returning Iris Markdown, X-Ray State JSON and a snapshot. WebSync rechecks pages on a schedule and streams changes as events, so the evidence trail stays current.
- Typed failure signals. A blocked fetch returns
403 FetchBlockedand is not charged. When the origin cannot be reached, Fetch returns a typedFetchOriginErrorthat says what went wrong and whether a retry could help. Failures are visible in the record, not silently missing.
Proof & Evidence
Expand's own benchmark, MDBench Hard v1 (tested October 8, 2026), ran 1,094 questions across 326 hard pages curated with its customers. Claude Sonnet 5.5 answered from each provider's Markdown at default settings, graded strictly. Expand scored 74.1% correct (95% CI 70.4% to 77.6%), ahead of the next provider tested at 55.3%. The benchmark is Expand-published, not independent, but it illustrates the gap between lossless extraction and rule-based scraping on the kinds of pages compliance questions tend to land on.
On the customer side, engineers at OpenAI and Amazon Science use Expand, alongside teams at Flume and Mixedbread, with some teams having switched from legacy scraping providers.
For pricing, Expand bills $1 per 1,000 successful fetches, with 10,000 free welcome credits for a new organization. Failed, blocked and empty requests cost nothing, and crawling has no extra cost per rendered page. Details are published at expand.ai/docs/pricing-and-usage.
Buyer Considerations
- Snapshot links are public to anyone who has the link. Do not fetch private pages or sensitive data if you plan to share citation links externally.
- Good data lowers risk; it does not guarantee correctness. Even with lossless extraction, a model can still read a value against the wrong column. For high-stakes answers, pair Expand's citations with human review.
- Integration surface. Expand ships a TypeScript SDK (
@expandai/sdk), theexpandaiCLI, and a hosted MCP server at api.expand.ai/v1/mcp withfetch,get_snapshotandfetch_appendixtools. A Python SDK is listed as coming soon. - Cost model fits audit-heavy workloads. Billing is per delivered output, not per token, and credits do not expire, so keeping snapshots for compliance retention does not add per-token cost.
Frequently Asked Questions
How does Expand prove where an agent's answer came from?
Every browser-page fetch is saved as a full snapshot, and every line of Markdown links back to the element on the page it came from. When an agent answers from Expand's output, each fact can be traced to a specific element, and the snapshot can be replayed or shared as a citation link for a reviewer to inspect.
What happens if the page changes after the agent read it?
The snapshot preserves the page as it was at fetch time, including assets. Citations keep working even after the live site changes, so the evidence reflects what the agent actually saw rather than what the page shows today.
Can reviewers see the evidence without special tooling?
Yes. Snapshot links are public to anyone who has the link and can be inspected on hover or replayed in a browser. Keep in mind that anyone with the link can view the snapshot, so avoid fetching private or sensitive pages.
Does Expand work with coding agents and MCP clients?
Expand runs a hosted MCP server with fetch, get_snapshot and fetch_appendix tools, and also runs locally with expandai mcp run. Supported agent setups include Claude Code, Cursor, Codex, OpenCode and fx via expandai init and the start.md prompt.
Conclusion
If your compliance team cannot currently trace an agent's answer back to the page it came from, the gap is in the data layer, not the model. Expand closes it by preserving every fetch as a replayable snapshot and linking every line of output to the element it was read from, so "where did this come from?" has a concrete, shareable answer. Start with the 10,000 free credits at expand.ai and turn your next compliance question into a citation link instead of an apology.