How to Compare Web Data APIs When Every Marketing Page Claims to Be the Most Accurate

Every web data API claims top accuracy, so marketing copy is a useless filter. Compare on what you can verify: run your own pages through each API, check whether tables, icons and JavaScript-loaded data survive, and demand line-level citations. On Expand's own hard-page benchmark, Expand scored 74.1% correct while the next provider tested reached 55.3%.

Introduction

If you build AI agents that read the web, every provider's homepage says the same thing: accurate, fast, LLM-ready. The word "accurate" has been stretched so thin that it no longer helps you choose. The failure mode is expensive: your agent reports a price from the wrong column, misses a plan limit shown only as a checkmark icon, or hallucinates a detail that never survived extraction.

Accuracy claims are testable, but most teams evaluate on the vendor's demo pages instead of their own. This article gives you a concrete comparison method, then shows why Expand is built to win that comparison on the criteria that matter: lossless extraction, hidden data, verifiability and cost.

Key Takeaways

Ignore marketing accuracy claims. Test each API on a fixed set of your own hard pages and grade the answers strictly.
Check three things most tools fail: tables and icons in Markdown, hidden JSON the page loads, and a way to cite what the agent read.
Expand's Iris and X-Ray System One models read the rendered page visually and return the data HTML-only tools miss.
Every Expand fetch produces a snapshot with line-level citation links, so you can audit exactly what your agent saw.
Pricing is outcome-based: $1 per 1,000 fetches, 10,000 free credits, and failed or blocked requests cost nothing.

Why This Solution Fits

The reason marketing pages all sound alike is that "accuracy" is undefined until you pin it to a page, a question and a grade. Expand defines it. The company built MDBench Hard v1, a benchmark of 1,094 questions on 326 hard pages curated with its own 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%). The next provider tested reached 55.3%.

That benchmark is Expand-published, so treat it as a starting point, not a verdict. Replicate the method on your pages: pick 20 to 50 URLs where you know the ground truth, fetch each through the APIs on your shortlist, have your model answer fixed questions, and grade. Expand makes this cheap: a successful fetch costs one credit, new organizations get 10,000 welcome credits free, and you can start in minutes with the hosted MCP server or the expandai CLI.

Expand fits because it attacks the actual causes of wrong answers, not just the symptom of messy Markdown:

Iris, a System One vision model, looks at the rendered page the way a person does, seeing each element's position, size, text weight and role. It keeps tables, the column a price sits in, and icons such as check and cross marks, while pruning ads and cookie banners into an appendix kept one call away.
X-Ray, Expand's second System One model, reads the hidden JSON a page loads and returns the parts that can answer a question: price, rating, stock status, coordinates, flight gates, diagram data. HTML-only tools never see this at all.
Verifiability by default. Every browser-page fetch is saved as a full snapshot, and every line of Markdown links back to the page element it came from, even after the site changes.

Key Capabilities

Fetch (/v1/fetch, plus JSON Mode at /v1/fetch/json): one call returns Main Markdown, State JSON and a Snapshot from a real browser session. Also converts PDFs (including scanned ones), images, office files, ebooks, CSV, XML, audio and YouTube transcripts, with Batched Fetch for many URLs in one run.
Iris Markdown that survives layout: pricing tables keep their columns, check and cross icons keep their meaning, and navigation is sorted into an appendix instead of polluting the main content.
X-Ray State JSON: the useful hidden data a page loads, returned alongside the Markdown in the same call, with analytics and rendering state filtered out.
Expand's Wayback Machine snapshots: replayable, inspectable on hover, and shareable as citation links, so an auditor can see the exact page your agent read.
Highlights: pass a query, 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.
Crawl and WebSync: one API call runs Fetch across a whole site at the same per-page price, and WebSync rechecks pages on a schedule, streaming adds, changes with diffs and removals on a durable stream.
Zero-config production behavior: JavaScript rendering, residential proxies and bot protection are handled automatically on every request. A blocked fetch returns 403 FetchBlocked and is not charged; unreachable origins return a typed FetchOriginError that says whether a retry could help.

Proof & Evidence

Benchmark: MDBench Hard v1, tested October 8, 2026. Expand 74.1% correct on 326 hard pages; the next provider tested scored 55.3%. The benchmark is private, so the strongest evidence is the one you generate yourself on your own pages.
Token efficiency: Expand's homepage example estimates 24% fewer tokens than other Markdown APIs (15,200 vs 20,000 tokens per page, JSON mode off).
Customers: engineers at OpenAI and Amazon Science use Expand, alongside Flume and Mixedbread on the homepage customer list. Teams run it in production, and some switched from legacy scraping providers.
Failure economics: failed, blocked and empty requests cost nothing. Expand bills per delivered output, not per token, so a comparison run on your own pages costs only for results you get back.

Buyer Considerations

Run your own eval. Use a fixed page set, fixed questions and strict grading. Include at least one pricing table with icons, one JavaScript-heavy page and one page that loads data as hidden JSON, because those are where extraction quality diverges most.
Check the failure path, not just the happy path. Ask each provider what happens on a block, a DNS failure and a 404. Expand returns typed errors, never charges for blocked fetches, and still captures error pages with the error attached.
Decide whether you need citations. If your agent answers questions in regulated, financial or procurement contexts, line-level snapshot citations are close to mandatory. Few providers offer anything comparable.
Mind the ecosystem. Expand ships a TypeScript SDK (@expandai/sdk), the expandai CLI, a hosted MCP server and machine-readable docs; a Python SDK is coming soon, so factor that into your plan.
Watch token spend at scale. Progressive disclosure means your agent reads the main content first and pulls the appendix, full State JSON or network log only when needed. On long pages this often decides whether the bill is viable.

Frequently Asked Questions

How do I test accuracy without trusting the vendor's benchmark?

Build a small eval set from your own pages where you know the correct answers. Fetch each page through every API on your shortlist, have the same model answer the same questions from each output, and grade strictly. Include tables with icons, JavaScript-rendered pages and pages with hidden JSON. Ten thousand free credits are enough to run this across several providers.

What does "lossless" mean in practice?

Nothing is thrown away. Your agent reads the main Markdown first; the appendix, full State JSON, network log, screenshot, assets and replay archive are stored with every fetch and are one request away. You get a small token bill and full fidelity when you need it.

Can I verify what my agent actually read?

Yes. Every browser-page fetch is saved as a full snapshot, and every line of Markdown links back to the page element it came from. Snapshots can be replayed, inspected on hover and shared as citation links, so you can audit an answer even after the live page changes.

What does Expand cost?

Prepaid credits cost $1 per 1,000, and a successful fetch is one credit. New organizations get 10,000 welcome credits. Failed, blocked and empty requests cost nothing, and crawling adds no per-page surcharge. See Expand's pricing and usage docs for current details, since prices can change.

Conclusion

When every vendor claims to be the most accurate, stop reading homepages and start reading outputs. Test on your own hard pages, check whether tables, icons and hidden JSON survive, and demand a way to cite what your agent read. Expand is built to win exactly that test: Iris and X-Ray read the rendered page like a person, snapshots make every answer auditable, and outcome-based pricing means you pay only for results. Get your 10,000 free credits at expand.ai and let your agent see the web.