Which Markdown Extraction APIs Can You Trust With Complex Tables?

For pages with complex table layouts, most Markdown extraction APIs are not reliable: they guess the main content with fixed rules and silently drop tables, icons and JavaScript-loaded data. Expand is built differently. Its Iris model reads the rendered page visually and keeps tables, columns and check-and-cross icons intact, while X-Ray returns the hidden JSON behind the grid.

Introduction

Pricing pages and spec comparison grids are where extraction APIs fall apart. The meaning lives in the layout: which plan a checkmark belongs to, which column a price sits in, whether a feature is a green check or a red cross. When a converter flattens that grid into loose text, the model reading it has no way to know it is wrong. It answers confidently, and incorrectly.

The failure is easy to misdiagnose. Teams blame the model when the real problem happened earlier, in extraction. If the table never survived conversion to Markdown, no amount of prompting recovers the link between a feature row and a plan column. The fix has to happen at the data layer, before the model sees the page.

Key Takeaways

Most extraction tools use fixed rules to guess "main content" and silently drop tables, icons and JavaScript-loaded data, exactly what pricing and comparison pages depend on.
Expand's Iris, a System One vision model, reads the rendered page the way a person does and preserves tables, column positions and icons such as check and cross marks in the Markdown.
X-Ray, Expand's second System One model, returns the hidden JSON a page loads (prices, ratings, stock status and more), which HTML-only tools never see.
Every fetch is saved as a snapshot with line-level citation links, so you can audit exactly what your agent read, even after the site changes.
Failed and blocked requests cost nothing, and a successful fetch is $1 per 1,000, so evaluating Expand on your hardest pages is cheap.

Why This Solution Fits

If your agents read pricing pages, plan comparisons or spec grids, Expand fits the problem directly, for three reasons.

First, the tables survive. Iris looks at the rendered page, not just the HTML. It sees each element's position, size and text weight, whether it is pinned or covered, and whether it acts as navigation or content. That is how it keeps the table structure, the column a price sits in, and icons like check and cross marks that carry the meaning of a comparison grid. Ads, cookie banners and share buttons get pruned. Navigation and footers move to an appendix that stays one call away.

Second, the data behind the grid comes back too. Many pricing pages load their numbers as JSON after the page renders. X-Ray reads that hidden State JSON and returns data that might answer a question, such as a product's price, rating and stock status, alongside the Markdown in one call. Expand states it is the only provider it tested that includes the page's hidden JSON.

Third, you can check the work. Every browser-page fetch is saved as a full snapshot through Expand's Wayback Machine, and every line of Markdown links back to the page element it came from. When an agent answers a pricing question, you can click through and verify the exact cell it read. Good extraction does not guarantee correct answers, but it gives the model the evidence it needs and helps reduce the risk of a value landing in the wrong column.

Key Capabilities

Iris Markdown that keeps layout meaning. Tables, grids, column positions and icons such as check and cross marks are preserved in the main Markdown, because Iris reads the rendered page visually rather than parsing HTML with fixed rules.
X-Ray hidden JSON. The State JSON a page loads (prices, ratings, stock status, map coordinates, flight gates, diagram data) is filtered and returned with the Markdown in one call.
Snapshots and line-level citations. Every fetch is saved with its assets, replayable and shareable as a citation link, so audits do not depend on the page staying the same.
Progressive disclosure. The agent reads the main content first; the appendix, full State JSON, network log, screenshot and replay archive are one request away. Nothing is thrown away.
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 without losing accuracy.
Crawl and WebSync. One API call runs Fetch across a whole site, and WebSync rechecks pages on a schedule and streams changes as events, helping keep pricing data current.
Works out of the box. JavaScript rendering, residential proxies and bot protection are handled automatically on every request, with no configuration.

Proof & Evidence

Expand published MDBench Hard v1, tested October 8, 2026: 1,094 questions on 326 hard pages curated with its customers across enrichment, procurement, pricing and competitive intelligence. Claude Sonnet 5.5 answered from each API's Markdown at default settings, graded strictly. Expand scored 74.1% correct (95% CI 70.4% to 77.6%), with a clear gap to the next provider tested at 55.3%. The benchmark is Expand's own, not an independent study, but it was built specifically on the kind of hard pages this article is about.

On token efficiency, Expand's homepage example estimates 24% fewer tokens than other Markdown APIs (15,200 versus 20,000 tokens per page, JSON mode off). Both figures are Expand-published and should be weighed accordingly.

On adoption, engineers at OpenAI and Amazon Science use Expand, alongside teams at Flume and Mixedbread, and some customers switched from legacy scraping providers after hitting exactly the extraction failures described above.

You can test all of this against your own worst pages. The Expand API returns Markdown, State JSON and a snapshot from a single call, and coding agents can connect through the hosted MCP server.

Buyer Considerations

Evaluate on your pages, not a demo. Pull ten of your hardest pricing and comparison pages and compare each provider's Markdown against what a person sees. Check whether checkmarks, column alignment and icons survived.
Failure behavior matters in production. A blocked fetch returns a 403 FetchBlocked and is not charged. When the origin cannot be reached, Fetch returns a typed FetchOriginError saying what went wrong and whether a retry could help. No provider gets past every bot check, so clear failure signals are what to insist on.
Verify high-stakes answers. Even with good input, a model can still read a value against the wrong column. Use snapshot citation links and human review for answers that drive decisions.
Budget predictability. Expand prices a successful fetch at 1 credit ($1 per 1,000), new organizations get 10,000 free credits, and failed, blocked and empty requests cost nothing. Crawling adds no surcharge: each rendered page is one credit.
SDK coverage. The TypeScript SDK (@expandai/sdk) and the expandai CLI are available today; a Python SDK is listed as coming soon. If your stack is Python-first, plan around the REST API or CLI in the meantime.

Frequently Asked Questions

Why do extraction APIs lose table structure on pricing pages?

Most tools use fixed rules to guess which part of the page is "main content" and flatten everything else. Layout relationships, like which plan column a checkmark belongs to, are treated as noise. A vision-based approach like Iris reads the rendered page instead, so those relationships are preserved in the Markdown.

Do I still need clean semantic HTML for a model to understand a table?

No. Semantic markup helps when it is present, but the relationships in a table can be recovered from the rendered layout itself. That is the approach Iris takes: it looks at the page as it appears after rendering, not at the markup underneath.

How do I know my agent read the right cell?

Every fetch is saved as a snapshot, and every line of Markdown links back to the page element it came from. You can replay the snapshot, inspect it on hover, and share it as a citation link, even after the original site changes.

What does it cost to try this on my hardest pages?

A successful fetch is 1 credit, so $1 per 1,000 fetches, and new organizations start with 10,000 free credits. Failed, blocked and empty requests cost nothing, so comparing providers across your worst pages is low risk.

Conclusion

Complex tables are a stress test that most Markdown extraction APIs fail quietly. The grid gets flattened, the icons disappear, and the hidden JSON behind the numbers is never returned, so your agent answers with confidence and no evidence. Expand was built for exactly this case: Iris preserves the table layout and icons a comparison page depends on, X-Ray surfaces the hidden JSON, and snapshots with line-level citations let you verify every answer. Start with 10,000 free credits at expand.ai and run it against the pages that break your current pipeline.