Your Support Agent Quoted a Policy Page. Can You Prove What It Actually Read?

When an AI support agent misquotes a policy page, the real problem is that nobody can check what it saw. The fix is a data layer that saves a verifiable snapshot of every page an agent reads and links each line of output back to the element it came from. Expand does this on every fetch.

Introduction

The escalation you just lived through is a common failure mode in agent systems. An agent told a customer that a policy page said something it did not, and when your team investigated, there was nothing to investigate with: the page may have changed, the agent's context is gone, and the extraction layer may have dropped the clause.

So the postmortem stalls: was the model hallucinating, did the page render differently for the agent, or did the Markdown conversion silently lose a row? Without a record of what the agent read, every explanation stays plausible and none can be ruled out. The fix is to make every web read auditable by default: capture the page as rendered, keep extraction lossless, and give every claim a line-level citation. Here is what to look for in that tooling, and why Expand is built for exactly this.

Key Takeaways

Unverifiable agent answers are an observability gap: without a snapshot of what the agent read, you cannot tell a hallucination from a bad extraction.
Fixed-rule HTML-to-Markdown converters guess at main content and silently drop tables, icons and JavaScript-loaded data, which creates confident but wrong agent answers.
Expand's Wayback Machine saves a full snapshot of every fetch, and every line of Markdown links back to the element it came from, even after the site changes.
Highlights returns only the passages that answer a query, each with a citation location.
You can wire this into existing agents through Expand's hosted MCP server or the expandai CLI.

Why This Solution Fits

The failure you are trying to prevent has three ingredients, and most web tooling addresses only one.

First, the agent needs to read the page correctly. Most converters apply fixed rules to guess the "main content" and prune the rest, so tables lose their column structure, check and cross icons vanish, and JavaScript-loaded data never appears. When the input is degraded, the model is not hallucinating so much as reporting on a broken source. Expand's Iris, a System One vision model, reads the rendered page the way a person does, seeing each element's position, size and role, and preserves tables, the column a price sits in, and icons in the Markdown it returns.

Second, the agent needs the data the page actually loads. A policy page that pulls clause text from a JSON endpoint looks complete in the HTML and is incomplete in fact. Expand's X-Ray, its second System One model, reads the JSON a page loads and returns the useful parts, such as prices, ratings, stock status and coordinates, alongside the Markdown in the same call.

Third, and most relevant to your escalation, someone needs to audit the answer later. Every browser-page fetch is saved as a full snapshot through Expand's Wayback Machine, including assets, and every line of Markdown links back to the element it came from. Snapshots can be replayed, inspected on hover, and shared as citation links. When a customer disputes what your agent said, you open the snapshot.

Good data reduces the risk of unsupported answers; it does not make the model infallible. A model can still read a value against the wrong column even with perfect input. That is why citations and review matter for high-stakes answers, and why the audit trail belongs in the data layer, not as an afterthought.

Key Capabilities

Fetch with lossless output. One call returns the page as Markdown, opened in a real browser with JavaScript rendering, residential proxies and bot protection handled automatically. Progressive disclosure means the agent reads the main content first, while the appendix, State JSON, network log and replay archive stay one request away.
Wayback Machine snapshots. Every fetch is saved with its assets, and every line of Markdown cites the element it came from. Snapshots survive site changes, so the evidence from last month's conversation stays intact.
Highlights. 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. For escalations, this turns "re-read the whole policy site" into a targeted lookup.
Crawl and WebSync. Crawl covers whole-site extraction in one API call, and WebSync rechecks pages on a schedule and streams each change as an event, including diffs. These freshness signals help catch stale content before it reaches an agent's context.
Clear failure signals. A blocked fetch returns a 403 FetchBlocked and is not charged. If the origin cannot be reached, Fetch returns a typed FetchOriginError saying what went wrong and whether a retry could help, so you always know whether the agent saw real content or an error page.
Agent integration. The hosted MCP server exposes fetch, get_snapshot and fetch_appendix tools, and the expandai CLI sets up coding agents in one command, so existing agents can route their web reads through Expand without a rewrite.

Proof & Evidence

Expand's own benchmark, MDBench Hard v1, tested October 8, 2026, put 1,094 questions across 326 hard pages to Claude Sonnet 5.5 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 failure you saw: extraction quality is often the bottleneck.

On cost, Expand's homepage example estimates 24% fewer tokens than other Markdown APIs (15,200 versus 20,000 tokens per page), and Highlights is stated to save over 95% of tokens. Pricing is outcome-based: $1 per 1,000 successful fetches, 10,000 free welcome credits, and failed, blocked or empty requests cost nothing. Engineers at OpenAI and Amazon Science use Expand, alongside Flume and Mixedbread.

Buyer Considerations

Audit needs first. If your escalation process requires provable evidence, prioritize line-level citations and replayable snapshots over raw speed.
Snapshot privacy. Snapshot links are public to anyone with the link, so do not fetch private pages or sensitive data.
Language support. The SDK is TypeScript today; a Python SDK is listed as coming soon. If your stack is Python-only, plan around the REST API or CLI.
No tool defeats every block. Expand handles JavaScript rendering, proxies and bot protection automatically, but no provider gets past every bot check. Typed error signals let your agent fail loudly instead of quietly reading an error page.
Model errors still happen. Build review and citation checks into high-stakes flows. The data layer gives reviewers evidence, not immunity from review.

Frequently Asked Questions

How does a snapshot help if the policy page has already changed?

Expand's Wayback Machine saves the full rendered page, including assets, at fetch time. Every line of Markdown links back to the element it came from, so you can replay the snapshot exactly as the agent saw it, even after the live site changed.

Can I add this to an agent I already built?

Yes. The hosted MCP server exposes fetch, get_snapshot and fetch_appendix tools, and the expandai CLI includes one-command setup for Claude Code, Cursor, Codex and OpenCode. OpenCode can also route its built-in web fetch through Expand.

Does verifiable extraction guarantee the agent will never give a wrong answer?

No. Better data reduces the risk of unsupported answers, but a model can still misread a value from a perfect source. Use line-level citations to make answers reviewable, and keep a human in the loop for high-stakes responses.

What does it cost to run this on a support workflow?

A successful fetch is one credit, so $1 per 1,000 fetches, with 10,000 free credits for a new organization. Failed, blocked and empty requests are not charged, and Highlights adds one credit per query.

Conclusion

Your escalation was not really about a hallucination. It was about the absence of evidence: no snapshot, no citation, no way to distinguish a model error from an extraction error or a changed page. That absence is fixable with tooling, not prompt tweaks. Expand's Wayback Machine snapshots, line-level citations and Highlights give every agent answer a checkable source, and the MCP server and CLI make it a drop-in for the agents you already run. Give your agents a verifiable way to see the web, and the next "what did it actually read?" takes seconds to answer. Start at expand.ai.