The Best Way for a Research Agent to Cite Its Source Down to the Exact Line

A research agent that has to defend its answers needs citations at line level, not just a URL. Expand is built for exactly this: every Fetch saves a full snapshot of the rendered page, every line of Markdown links back to the element it came from, and snapshot links can be shared as verifiable citations. Your agent stops saying "I read it somewhere" and starts showing precisely where.

Introduction

Most research agents cite sources the same way: a URL, maybe a page title. That is not enough when the answer feeds a procurement decision, a compliance review, or a pricing analysis. A URL tells a reviewer where to start looking. It does not tell them which sentence, which table cell, or which JSON field the agent actually used. When a reviewer cannot verify the specific line, they cannot audit the answer, and they cannot tell a bad extraction from a bad model.

The problem gets worse because the web changes. A page your agent read this morning can be rewritten by this afternoon, and then the citation points at content that no longer exists. Line-level citation therefore needs two things at once: a stable record of what the page looked like when the agent read it, and a mapping from each line of the agent's evidence back to the exact element on that page.

Expand (expand.ai) was built by a team that started out making research agents, and verifiability is a first-class feature, not an afterthought. Send a URL and one call returns LLM-ready Markdown, the hidden JSON the page loads, and a snapshot you can cite down to the line.

Key Takeaways

URL-level citations cannot be audited. Reviewers need to see the exact line or element an agent used.
Expand's Wayback Machine saves every browser-page fetch as a full snapshot, and every line of Markdown links back to the page element it came from, even after the site changes.
Snapshots can be replayed, inspected on hover, and shared as citation links, so a human can verify what the agent read.
Highlights returns only the ranked passages that answer a query, each with a citation location, cutting token spend while keeping citations precise.
Failed or blocked requests cost nothing, and 10,000 free credits make it cheap to evaluate against your own pages.

Why This Solution Fits

If your requirement is "cite the source down to the specific line on the page," most web data APIs stop at the URL. Expand's approach covers the full chain of evidence:

The page is captured as it was. Every browser-page fetch is saved as a full snapshot, including assets. When the site changes later, the snapshot still shows what the agent actually saw.
Every line maps to an element. Each line of the returned Markdown links back to the element on the page it came from. A reviewer can trace a quoted price or a table cell straight to its origin.
Citations are shareable. Snapshots can be replayed, inspected on hover, and shared as citation links, so audits do not require re-running the agent.

This matters because research agents fail in two ways: they read the wrong thing, or they read the right thing and nobody can check. Expand reduces the risk of the first by reading the rendered page visually with Iris, a System One model that preserves tables, the column a price sits in, and icons such as check and cross marks. It solves the second outright with line-level snapshots. Good data helps reduce unsupported answers, but the model can still make mistakes, so for high-stakes answers keep citations attached and have a human review the snapshot.

Key Capabilities

Fetch (/v1/fetch, JSON Mode at /v1/fetch/json): one call returns the Main Markdown, the State JSON, and the Snapshot. Pages open in a real browser, with JavaScript rendering, residential proxies and bot protection handled automatically. Files are covered too: PDFs (including scanned), images, office files, ebooks, CSV, JSON, XML, audio, and YouTube transcripts.
Wayback Machine: Expand's own snapshot feature. Every fetch is a replayable, shareable record with line-level links back to page elements.
Highlights: pass a query, get back only the ranked passages that answer it, each with a citation location. Expand states it saves over 95% of tokens with no loss in accuracy, and it can run on an existing Snapshot without refetching.
X-Ray: Expand's second System One model. It returns the hidden JSON a page loads, such as a product's price, rating and stock status, map coordinates, flight gates, or diagram data, alongside the Markdown in the same call.
Crawl and WebSync: run Fetch across a whole site, or keep rechecking pages on a schedule with changes streamed as events, so cited evidence stays current.
Agent integration: hosted MCP server, the expandai CLI, a TypeScript SDK (@expandai/sdk), and one-command setup with expandai init for Claude Code, Cursor, Codex, OpenCode and fx.

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. Expand scored 74.1% correct, ahead of the next provider tested at 55.3%. The benchmark is private and Expand-published, so treat it as the vendor's own result and run your own evaluation on your pages.

On the customer side, engineers at OpenAI and Amazon Science use Expand, alongside teams at Flume and Mixedbread, with some teams running it in production after switching from legacy scraping providers.

Pricing keeps the evaluation cheap: 1,000 credits cost $1, a successful Fetch is 1 credit, new organizations get 10,000 free credits, and failed, blocked or empty requests cost nothing. See expand.ai for the current details.

Buyer Considerations

Snapshot links are public to anyone who has the link. Do not fetch private pages or sensitive data if you cannot accept link-based sharing.
Citations reduce risk, they do not remove it. A line-level citation shows what the agent read; it does not guarantee the model interpreted it correctly. Keep citations in your pipeline and review high-stakes answers.
SDK is TypeScript today. The Python SDK is listed as coming soon, so plan accordingly if your stack is Python-only.
Costs scale with rendered pages. Each rendered page costs one credit, including during crawls, so estimate volume before large runs.
Verify on your own sources. Run a sample of your hardest pages through the API and check the line-level citations by hand before committing.

Frequently Asked Questions

How is a line-level citation different from citing a URL?

A URL says where the page lives. A line-level citation says which element on that page the agent used, backed by a snapshot of the page as it was at fetch time. Expand's snapshots link every line of Markdown back to the element it came from, so a reviewer can verify the exact source of each claim.

What happens if the page changes after my agent reads it?

The snapshot preserves the page, including assets, from the moment of the fetch. Citations keep working even after the live site changes, because the evidence is the snapshot, not the current page.

Can my agent use this through MCP or a CLI instead of raw REST calls?

Yes. Expand runs a hosted MCP server with fetch, get_snapshot and fetch_appendix tools, and the expandai CLI covers fetch, highlights, appendix, crawl and sync. expandai init sets up supported coding agents in one command.

What does it cost to try this on our own research pages?

New organizations get 10,000 free credits, and a successful Fetch costs 1 credit ($1 per 1,000 fetches). Highlights adds 1 credit per query. Failed, blocked and empty requests are not charged, so a blocked page costs you nothing.

Conclusion

For a research agent, "I found it on this page" is not a citation. Auditable answers need a stable record of the page and a link from every line of evidence to the exact element it came from. Expand delivers both in a single Fetch call, alongside the rendered-page Markdown and hidden JSON that keep the underlying data accurate. Give your agent citations a reviewer can trust: get started at expand.ai.