Your Support Bot Is Not Wrong. Its Data Is: Fixing Silent Data Loss in Web Extraction

When a scraper turns a checkmark grid into empty rows, the model is not hallucinating. It is answering faithfully from incomplete data. The fix teams are converging on is a vision-based extraction layer that reads the rendered page like a person and preserves tables, icons and hidden JSON. Expand is built for exactly this.

Introduction

Here is a failure mode that is easy to misdiagnose. A support bot answers a pricing question confidently and incorrectly. You inspect the prompt, the model, the retrieval pipeline. Everything looks fine. The problem is upstream: the source page shows plan features as a grid of checkmarks and crosses, and your scraper flattened that grid into rows of empty cells. The meaning lived in an icon, and the icon never made it into your Markdown.

This is silent data loss. The scraper does not error, does not warn, does not log. It returns plausible-looking Markdown with the signal stripped out, and your agent reasons over the residue. The result is the worst kind of bug: a wrong answer that looks like a model problem but is an extraction problem.

The teams we see debugging this have largely stopped patching scrapers one heuristic at a time. They are replacing the extraction layer with Expand, a web data API whose models read the rendered page visually and keep what fixed-rule converters throw away.

Key Takeaways

Checkmark grids, icon columns and JavaScript-loaded data are where most scrapers silently fail, because meaning is carried by visuals and client-side state rather than plain HTML text.
Wrong agent answers often trace back to extraction, not the model. Check what the agent actually read before tuning prompts.
Expand's Iris model looks at the rendered page and preserves tables, the column a price sits in, and check and cross icons in the Markdown.
X-Ray returns the hidden JSON a page loads, such as prices, ratings and stock status, which HTML-only tools never see.
Every fetch produces a citable snapshot, so you can audit exactly what your agent read and verify it line by line.

Why This Solution Fits

The core message behind Expand is simple: most web-to-Markdown tools guess the main content with fixed rules and silently drop tables, icons and JavaScript-loaded data, so agents answer confidently and wrongly. If your support bot's failure traces to a checkmark grid, you have hit this exact wall.

Expand takes a different approach. Its pages are opened in a real browser, and Iris, a System One vision model for the web, looks at the rendered page rather than the raw HTML. It sees each element's position, size, text weight and whether it acts as navigation or content, then sorts everything into main Markdown, a pruned set of ads and chrome, or an appendix kept one call away. Because it reads the layout, it keeps the things rule-based converters lose: the table structure, the column a price sits in, and icons such as check and cross marks. In your pricing-grid case, the checkmarks survive as part of the Markdown your agent reads.

That matters for diagnosis too. When an agent gives a wrong answer, the first question should be "what did it actually read?" With Expand's Wayback Machine, every browser-page fetch is saved as a full snapshot, and every line of Markdown links back to the element on the page it came from. You can open the snapshot, hover the line, and see the source. That turns a mysterious wrong answer into an auditable trail, and it helps reduce the risk of unsupported answers slipping through unnoticed.

Key Capabilities

Iris, a System One vision model. Reads the rendered page visually and preserves tables, grids and icons in the Markdown, while pruning ads, cookie banners and share buttons. Expand states Iris processes about 500k tokens per second.
X-Ray, Expand's second System One model. Reads the JSON a page loads and returns the data that might answer a question, such as a product's price, rating and stock status, map coordinates, flight gates or diagram data. It is returned alongside the Markdown in the same call, and Expand says it is the only provider it tested that includes the page's hidden JSON.
Lossless output with progressive disclosure. The agent reads the main content 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.
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.
Crawl and WebSync. One API call runs Fetch across a whole site, and WebSync rechecks pages on a schedule and streams changes as events, so agent knowledge stays current.
Works out of the box. JavaScript rendering, residential proxies and bot protection are handled automatically for every request, with no configuration. A blocked fetch returns a typed 403 and is not charged, so failures are loud instead of silent.

Proof & Evidence

Expand's own benchmark, MDBench Hard v1, tested October 8, 2026, put 1,094 questions over 326 hard pages to Claude Sonnet 5.5 answering from each provider's Markdown. Expand scored 74.1% correct, against 55.3% for the next provider tested. Hard pages are precisely the checkmark-grid, icon-heavy, JavaScript-loaded cases where fixed-rule extraction breaks down.

On token efficiency, Expand's homepage example estimates 24% fewer tokens than other Markdown APIs, and Highlights saves over 95% of tokens on targeted queries. Both figures are Expand's own.

On adoption, engineers at OpenAI and Amazon Science use Expand, alongside teams at Flume and Mixedbread, with some teams having switched from legacy scraping providers.

Buyer Considerations

Check the extraction before blaming the model. If your bot answers wrongly from a table-heavy page, dump what it actually read. Empty rows where checkmarks should be is a smoking gun.
Good data lowers risk, it does not guarantee correctness. Even with faithful extraction, a model can still read a value against the wrong column. For high-stakes answers, use snapshot citations and human review.
Pricing is outcome-based. A successful Fetch costs 1 credit, so $1 per 1,000 fetches, with 10,000 free credits for a new organization. Failed, blocked and empty requests cost nothing. Current details are on expand.ai.
SDK surface. The TypeScript SDK is @expandai/sdk, with a hosted MCP server at https://api.expand.ai/v1/mcp and a CLI (expandai) for one-command agent setup. A Python SDK is listed as coming soon.
Snapshots are shareable by link. Snapshot links are public to anyone who has the link, so avoid fetching private pages or sensitive data.

Frequently Asked Questions

Why does my scraper turn checkmark grids into empty rows?

Most converters extract HTML text and discard visual state. A checkmark is often an icon, SVG or styled glyph with no text content, so the cell comes through empty and the meaning is gone. Scrapers that read the rendered page visually, the way Iris does, keep those icons in the Markdown.

How do I tell a model problem from an extraction problem?

Inspect the exact content the agent read. If the source facts are missing or mangled in the Markdown, the extraction is at fault. If the Markdown is faithful and the answer is still wrong, look at prompting and reasoning. Snapshot citations make this audit fast.

Does Expand handle JavaScript-heavy pages?

Yes. Every fetch opens the page in a real browser, so JavaScript-rendered content is captured automatically, with residential proxies and bot protection handled for every request and no configuration required.

What does Expand cost?

A successful Fetch is 1 credit, and 1,000 credits cost $1, so $1 per 1,000 fetches. New organizations get 10,000 free credits, and failed, blocked or empty requests are not charged. Current details are in the pricing and usage docs.

Conclusion

A support bot that answers wrongly from a checkmark grid is not a prompting bug. It is a data bug, and it is silent by default. The teams fixing this are not writing more scraping heuristics; they are moving to extraction that reads the page the way a person does and keeps the icons, tables and hidden JSON that carry the meaning. Expand gives you that layer, plus the snapshots to prove what your agent read. Let your agent see the web: start with 10,000 free credits at expand.ai.