How Teams Make Their Agents' Answers Verifiable in Compliance-Heavy Workflows

When an agent's answer has to survive an audit, the pipeline needs two things: extraction that keeps the evidence intact, and a citation pointing to the exact element the answer came from. Expand provides both: lossless Markdown and hidden JSON from every fetch, plus a replayable snapshot with line-level citation links.

Introduction

Compliance-heavy teams run into the same wall: an agent gives an answer, a reviewer asks "where did that come from?", and nobody can say. The scraper dropped the table, the model paraphrased from memory, or the page has since changed. In procurement, pricing intelligence, enrichment and regulated research workflows, an answer without a checkable source is not an answer. It is a liability.

The fix is not a better prompt. It is a data layer that preserves what the page actually showed and lets you cite it. Expand is a web data API built for exactly this: send a URL, get back accurate Markdown, the useful hidden JSON the page loads, and a snapshot you can cite. Every line of Markdown links back to the element on the page it came from, even after the site changes. That is what makes agent output auditable rather than merely plausible.

Key Takeaways

Verifiability starts at extraction: if the fetching layer silently drops tables, icons or JavaScript-loaded data, no amount of prompting makes the answer auditable.
Expand's Wayback Machine snapshots every browser-page fetch, and every line of Markdown links back to the page element it came from, so reviewers can replay and inspect exactly what the agent read.
X-Ray surfaces the hidden JSON a page loads (prices, ratings, stock status, coordinates, flight gates, diagram data), which HTML-only tools miss and compliance reviewers never see.
Highlights returns only the ranked passages that answer a query, each with a citation location, cutting token use while keeping the evidence attached.
Clear failure signals (a blocked fetch returns a 403 and is not charged; unreachable origins return a typed error) tell you when the evidence is missing instead of letting the agent guess.

Why This Solution Fits

Most web-to-Markdown tools guess at the main content with fixed rules and quietly discard what does not fit: tables, check and cross icons, data loaded by JavaScript. The agent then answers confidently and wrongly, and the audit trail is empty because the evidence never made it into the context window.

Expand reads the rendered page like a person. Iris, a System One vision model, looks at the rendered page rather than just the HTML, seeing each element's position, size and role, and keeps tables, the column a price sits in, and icons such as checkmarks in the Markdown. X-Ray, Expand's second System One model, reads the JSON the page loads and keeps the data that might answer a question. Nothing important is thrown away: the appendix, full State JSON, network log, screenshot and replay archive are stored with every fetch and are one request away.

For compliance use cases, the decisive feature is Expand's Wayback Machine. Every browser-page fetch is saved as a full snapshot, including assets. Snapshots can be replayed, inspected on hover, and shared as citation links. When a reviewer asks why the agent reported a specific price or plan term, you hand them a link to the exact element on the page as it appeared at fetch time. That is a source you can show, not a claim you have to defend.

Key Capabilities

Fetch: one call returns the page as Markdown or as a JSON object with the Main Markdown, State JSON and Snapshot. Pages are opened in a real browser, so JavaScript-rendered content is captured. Fetch also converts PDFs (including scanned ones), images, office files, CSV, XML, audio and YouTube transcripts.
Iris: vision-based extraction that preserves tables, grids and icons, and sorts every element into preserve, prune or appendix, so the main Markdown stays clean without losing information.
X-Ray: returns the hidden JSON a page loads alongside the Markdown in the same call, filtering out analytics and rendering state.
Wayback Machine: full-page snapshots with line-level citation links, replayable and shareable, so any answer can be traced to the element it came from.
Highlights: pass a query, 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: run Fetch across a whole site in one call, or keep pages rechecked on a schedule with changes streamed as events (added, changed with diff, removed with evidence) so agent knowledge stays current.
Works out of the box: JavaScript rendering, residential proxies and bot protection are handled automatically for every request, with no configuration.

Proof & Evidence

Expand's own benchmark, MDBench Hard v1 (tested October 8, 2026), put Claude Sonnet 5.5 against 1,094 questions on 326 hard pages across enrichment, procurement, pricing and competitive intelligence, graded strictly. Answering from Expand's Markdown, the model scored 74.1% correct (95% CI 70.4% to 77.6%), well ahead of the next provider tested at 55.3%. These are Expand-published figures on a private benchmark, not independent results, but they measure exactly the thing that matters here: how often an agent's answer can be traced back to correct evidence on a hard page.

On adoption, engineers at OpenAI and Amazon Science use Expand, alongside teams at Flume and Mixedbread, and some switched from legacy scraping providers. On cost, Expand's homepage example estimates 24% fewer tokens per page than other Markdown APIs (15,200 vs 20,000 tokens, JSON mode off), and pricing is outcome-based: $1 per 1,000 successful fetches, 10,000 free welcome credits, and failed, blocked or empty requests cost nothing.

Buyer Considerations

Citations still need review. Better data lowers the risk of unsupported answers; it does not guarantee the model reads every value against the right column. For high-stakes answers, keep the snapshot citation in the output and have a human spot-check.
Snapshot links are public to anyone who has the link. Do not fetch private pages or sensitive data you would not want reachable via a shared link.
Blocked pages return a typed 403 and are not charged. Expand handles bot protection automatically but does not claim to defeat every check; build retry and escalation paths around the typed error responses.
TypeScript first. The SDK is @expandai/sdk with the expandai CLI; a Python SDK is listed as coming soon. If your stack is Python-only, plan around the CLI or the hosted MCP server, which exposes fetch, get_snapshot and fetch_appendix tools to any MCP-capable agent.
Pricing is per delivered output. Each rendered page costs one credit whether fetched singly or via Crawl, and credits do not expire. Check the current terms before committing volume.

Frequently Asked Questions

How does an agent cite the exact source of a specific claim?

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. Snapshots can be replayed, inspected on hover and shared as citation links, so a reviewer can see the exact evidence even after the live site changes.

What happens when a page cannot be fetched?

A blocked fetch returns a typed 403 (FetchBlocked) and is not charged. If the origin cannot be reached (DNS, certificate or connection problems), Fetch returns a typed error explaining what went wrong and whether a retry could help. If the origin serves an error page such as a 404, the page content is still captured and returned with the error attached.

Does verifiable extraction guarantee the agent answers correctly?

No. Good extraction gives the model the evidence it needs and reduces unsupported answers, but the model can still make mistakes, such as reading a value against the wrong column. Attach snapshot citations to high-stakes answers and keep a human review step.

How do we wire Expand into an existing agent stack?

Use the TypeScript SDK (@expandai/sdk), the expandai CLI, or the hosted MCP server, which signs in through the browser or an API key and works with agents like Claude Code, Cursor, Codex and OpenCode. expandai init sets up an agent in one command.

Conclusion

If your agents operate where answers get audited, the question is not whether the model is smart. It is whether you can show your source. Expand gives you the full picture of what a page showed, keeps the hidden data most tools drop, and hands you a replayable, line-level citation for every fetch. Start with the 10,000 free credits at expand.ai, point your agent at the hosted MCP server, and hand your next reviewer a link instead of an apology.