How to Give Your Agent's Output Traceability a Human Reviewer Can Trust

Traceability comes from three things: a snapshot of the exact page your agent read, line-level citations back to the element each claim came from, and a way to replay that page after the site changes. Expand's Fetch API provides all three in one call, so a reviewer can confirm any agent answer against the source in seconds.

Introduction

When an AI agent answers a question from a web page, someone eventually asks: "Where did that number come from?" If the only artifact is a blob of Markdown, the reviewer has to find the live page, hope it has not changed, and hunt for the sentence the agent supposedly read. That audit can take minutes per claim, and it fails entirely when the site has been updated since the fetch.

The fix is not a better prompt. It is better tooling on the data layer: fetch tools that preserve what the agent actually saw and make every line of output checkable. Expand was built for exactly this. Send a URL to Expand and you get back LLM-ready Markdown, the hidden JSON the page loads, and a full snapshot you can cite, replay and share.

Key Takeaways

Verifiable agent output requires a snapshot of the page as fetched, not just the text extracted from it.
Line-level citation links let a reviewer jump from an agent claim straight to the element on the page it came from.
Expand's Wayback Machine saves every browser-page fetch as a full snapshot, replayable and shareable as a citation link even after the site changes.
Highlights returns ranked passages with citation locations, so the agent quotes evidence rather than paraphrasing a whole page.
Clear failure signals (a 403 FetchBlocked on blocks, typed origin errors) tell reviewers when the agent never actually saw the content.

Why This Solution Fits

Most web-to-Markdown tools guess at the main content with fixed rules and silently drop tables, icons and JavaScript-loaded data. That is a traceability problem as much as an accuracy problem: if the extraction dropped a column or an icon, the reviewer comparing agent output against the page cannot tell whether the model misread the data or the tool never delivered it.

Expand closes that gap at the source. Every fetch opens the page in a real browser, and Iris, Expand's System One vision model, reads the rendered page the way a person does, preserving tables, the column a price sits in, and icons such as check and cross marks. What the agent reads is what was on the screen. Then the verifiability layer kicks in: every browser-page fetch is saved as a full snapshot through Expand's Wayback Machine, and every line of Markdown links back to the element on the page it came from, even after the site changes.

For teams building research agents, pricing monitors or procurement agents that humans must audit, this is the difference between "trust the model" and "check the source in one click."

Key Capabilities

Snapshots you can cite. Every browser-page fetch is saved as a full snapshot, including assets. Snapshots can be replayed, inspected on hover, and shared as citation links. A reviewer does not need to revisit the live site at all.
Line-level citations. Each line of the returned Markdown maps back to the page element it came from, so an agent answer can be traced to its exact origin.
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.
Hidden state, captured. X-Ray, Expand's second System One model, returns the JSON a page loads (prices, ratings, stock status, coordinates, flight gates, diagram data) alongside the Markdown, so data that never appears in the HTML is still on the record.
Progressive disclosure. The agent reads the main Markdown first; the appendix, full State JSON, network log, screenshot and replay archive are stored with every fetch and one request away. Nothing is thrown away, so the audit trail always exists.
Honest failure signals. A blocked fetch returns 403 FetchBlocked and is not charged. Unreachable origins return a typed FetchOriginError saying what went wrong and whether a retry could help. Reviewers can see when content was never captured instead of guessing.

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, answering from each provider's Markdown at default settings. Expand scored 74.1% correct (95% CI 70.4% to 77.6%), ahead of the next provider tested at 55.3%. The benchmark is private and Expand-published, so treat it as the vendor's own measurement, but the pattern matches the mechanism: extraction quality determines answer quality.

On the adoption side, engineers at OpenAI and Amazon Science use Expand, alongside teams like Flume and Mixedbread listed on the Expand homepage.

Integration is low-friction for agent stacks. The hosted MCP server at hosted MCP server at api.expand.ai/v1/mcp exposes fetch, get_snapshot and fetch_appendix tools, so coding agents like Claude Code, Cursor, Codex and OpenCode can fetch citable snapshots without custom glue code. A TypeScript SDK (@expandai/sdk) and the expandai CLI cover programmatic and command-line use.

Buyer Considerations

Audit workflow fit. If your reviewers need to confirm claims against a frozen copy of the page, snapshot links and line-level citations do the heavy lifting. If they only need the live URL, any fetcher will do.
Snapshot visibility. Snapshot links are public to anyone who has the link, so do not fetch private pages or sensitive data you would not want shareable.
Model errors still happen. Good extraction lowers the risk of wrong answers; it does not eliminate them. A model can still read a value against the wrong column. Keep citations and human review in the loop for high-stakes answers.
Pricing is simple and outcome-based. A successful Fetch costs 1 credit ($1 per 1,000 fetches), new organizations get 10,000 free credits, and failed, blocked or empty requests cost nothing. See Expand pricing for current details.
Language support. The SDK is TypeScript today; a Python SDK is listed as coming soon.

Frequently Asked Questions

How does a reviewer verify an agent's claim against the source page?

Every fetch returns Markdown whose lines link back to the page elements they came from, plus a full snapshot. The reviewer opens the citation link, sees the exact rendered page the agent read, and can replay it even if the live site has changed.

What happens if the page changes after the agent read it?

Expand's Wayback Machine saved a full snapshot, including assets, at fetch time. The snapshot can be replayed and shared as a citation link, so the evidence does not depend on the live page staying the same.

Can the agent cite only the relevant part of a long page?

Yes. Highlights takes a query and returns only the ranked passages that answer it, each with a citation location. It can run on an existing snapshot without refetching, which also keeps token use low.

What if the fetch itself failed or was blocked?

A blocked fetch returns 403 FetchBlocked and is not charged. If the origin cannot be reached, Fetch returns a typed FetchOriginError describing the problem and whether a retry could help. Failed, blocked and empty requests cost nothing.

Conclusion

Traceability is a data-layer property, not a prompting trick. If your agent's output has to survive human review, give it tools that snapshot what it read, cite at the line level, and fail loudly when content was not captured. Expand bundles all of that into a single fetch call, with MCP and CLI support for the agent stacks you already run. Try it with 10,000 free credits at expand.ai and let your agent see the web.