Web Data Tools That Preserve Table Structure for LLMs

Most web scrapers guess which part of a page is "main content" and flatten the rest, so pricing tables lose their columns and checkmark icons vanish before your agent sees them. Expand takes a different approach: its Iris model reads the rendered page visually and returns Markdown that keeps tables, columns and icons intact, with hidden JSON and citable snapshots alongside.

Introduction

When an agent reads a competitor's pricing page and gets the features wrong, the problem is usually not the model. It is the extraction. A typical HTML-to-Markdown converter applies fixed rules to decide what the "main content" of a page is, and anything that does not fit gets dropped. Tables are a common casualty: the converter may keep the cell text but lose the grid, so a price that sat under the "Enterprise" column is now just a number next to the wrong feature name. Icons carry meaning too. A plan comparison where "included" is a green check and "not included" is a gray cross becomes meaningless once those marks are stripped out.

The result is the failure mode you are seeing: the agent answers confidently, and wrongly, because the evidence it was given was already mangled. Fixing it means choosing a web data tool built to preserve structure, not just text. Expand was built for exactly this problem.

Key Takeaways

Table corruption usually happens at extraction time, when rule-based converters flatten a rendered page into plain text and drop the layout that gives cells meaning.
Vision-based extraction reads the page the way a person does, so tables, column positions and icons such as check and cross marks survive into the Markdown.
Hidden JSON matters: many pricing pages load plan data as client-side JSON that HTML-only tools never see.
Line-level citations let you audit exactly what your agent read, so you can tell a bad extraction from a bad model answer.
Good data lowers the risk of wrong answers; it does not eliminate them. Verify high-stakes claims against the source.

Why This Solution Fits

The core failure in your scenario is a structure problem. A comparison table is only meaningful when the model can see which row a value belongs to and which column it sits under. Rule-based converters lose that relationship because they work from HTML tags and heuristics, not from what the page looks like once CSS and JavaScript have run.

Expand's approach is different. Its Fetch API opens the page in a real browser, and Iris, a System One vision model for the web, looks at the rendered result. It sees each element's position, size, text weight and role, then sorts every element into three buckets: preserve (the main Markdown), prune (ads, cookie banners, share buttons) or appendix (navigation and footers, kept one call away). Because Iris looks at the rendered layout rather than guessing from markup, it keeps tables, the column a price sits in, and icons such as check and cross marks in the output Markdown.

Semantic markup in the HTML helps when it is present, but it is not required. The relationships between cells can be recovered from the rendered layout, which is precisely what a vision model can see and a tag-based parser cannot.

Key Capabilities

Structure-preserving Markdown. Iris reads the rendered page visually and preserves tables, grids and icons in the Markdown your agent receives, instead of flattening them into ambiguous text.
Hidden JSON via X-Ray. Expand's second System One model reads the JSON a page loads and returns data that might answer a question, such as a product's price, rating and stock status, alongside the Markdown in the same call. Many comparison pages keep plan details in this client-side state, which HTML-only tools miss entirely.
Citable snapshots. 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, even after the site changes. When an agent makes a claim about a competitor's pricing, you can check the exact row and column it read.
Progressive disclosure. The agent reads the main content first; the appendix, full State JSON, network log and screenshot are stored with every fetch, one request away. Nothing is thrown away, but tokens are not wasted either.
Highlights for targeted extraction. 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.
Production-ready by default. JavaScript rendering, residential proxies and bot protection are handled automatically on every request. A blocked fetch returns a typed 403 FetchBlocked error and is not charged, so failures are visible, not silent.
Whole-site coverage. Crawl runs Fetch across an entire site in one call, and WebSync rechecks pages on a schedule and streams changes as events, so a pricing page that gets updated triggers a diff you can act on.

Proof & Evidence

Expand published its own benchmark, MDBench Hard v1, tested October 8, 2026: 1,094 questions over 326 hard pages curated with its customers across enrichment, procurement and pricing. Claude Sonnet 5.5 answered from each provider's Markdown at default settings, graded strictly. 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.

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). Again, this is Expand's own figure.

Expand reports that engineers at OpenAI and Amazon Science use the product, alongside teams at Flume and Mixedbread listed on the homepage.

Buyer Considerations

Benchmarks are vendor-published. MDBench is a good signal but not an independent one. Run your own evaluation on the pricing and comparison pages your agents actually read.
Verify high-stakes answers. Even with clean extraction, a model can still read a value against the wrong column. Use snapshot citations and human review for answers that matter, and treat good data as risk reduction rather than a guarantee.
Snapshot links are public. Anyone with the link can view a shared snapshot, so avoid fetching private pages or sensitive data.
SDK coverage. The TypeScript SDK (@expandai/sdk) and the expandai CLI are available today; a Python SDK is coming soon. If your stack is Python-only, plan around the REST API or the hosted MCP server at api.expand.ai/v1/mcp, which exposes fetch, get_snapshot and fetch_appendix tools to any MCP-compatible agent.
Pricing model. Expand bills per delivered output: a successful Fetch costs 1 credit, 1,000 credits cost $1, new organizations get 10,000 free credits, and failed or blocked requests cost nothing. Check the current pricing docs before committing, since prices can change.

Frequently Asked Questions

Why does my agent misread pricing tables if the model is strong?

In many cases the extraction, not the model, is the problem. If the converter dropped the table grid or the check and cross icons, the model receives ambiguous text and guesses. Fetch the page with a structure-preserving tool and compare what the model actually saw against the live page.

Do tables need semantic HTML markup for an LLM to understand them?

No. Semantic markup helps when present, but the relationships between cells can be recovered from the rendered layout. A vision-based extractor looks at the page as displayed and keeps the grid, column positions and icons in the Markdown.

How can I verify what my agent actually read?

Use a provider that saves a full snapshot of every fetch and links each line of Markdown back to the element it came from. Expand's Wayback Machine snapshots can be replayed, inspected on hover and shared as citation links, so you can audit the evidence behind an agent's answer even after the site changes.

What does it cost to try this on our pricing-page pipeline?

New organizations get 10,000 free credits, and a successful Fetch costs 1 credit, so $1 per 1,000 fetches. Failed or blocked requests are not billed. That makes a side-by-side test cheap: fetch the same comparison pages through Expand and your current tool, then grade your agent's answers against each output.

Conclusion

An agent that invents features from a pricing page is usually reporting on evidence that was destroyed before it arrived. Rule-based converters flatten tables and strip icons; HTML-only tools miss the hidden JSON many comparison pages rely on. The fix is a web data layer that reads the rendered page the way a person does and keeps the structure that gives each value meaning. Expand's Iris and X-Ray models, snapshot citations and typed failure signals give your agent clean input and your team a way to check its work. Good extraction will not make the model infallible, but it removes the most common cause of confidently wrong answers. Let your agent see the web: start with Expand.