How to Make Sure Your Agent Sees the Pricing Grid a Human Sees

The tool teams now reach for is Expand, a web data API built for agents. Its Iris model reads the rendered page the way a person does and returns Markdown that keeps tables, grids, and check icons intact, while X-Ray surfaces the hidden price JSON and every fetch ships with a citable snapshot.

Introduction

A pricing page is a layout, not a document. The meaning lives in the structure: which plan the "Most popular" badge sits on, which features carry a check instead of a cross, whether the number on screen is the monthly price or the annual one. A human parses all of that in a glance. Most HTML-to-Markdown converters do not.

The failure is quiet. The converter guesses where the main content ends, flattens the comparison table into run-on text, drops the icons, and misses prices that only exist after JavaScript runs. Your agent then answers confidently from a garbled version of the page, and nobody can trace which line came from where.

Key Takeaways

Pricing grids break in conversion because meaning lives in layout: columns, icons, badges, and toggles, not just words.
Expand's Iris reads the rendered page visually and keeps tables, grids, and icons intact in the Markdown instead of flattening them.
X-Ray returns the hidden JSON a page loads, where prices often actually live.
Every fetch is stored as a snapshot in Expand's Wayback Machine, and each line of Markdown links back to the page element it came from.
Progressive disclosure and Highlights keep token use low, while Crawl and WebSync cover whole sites and change monitoring.

Why This Solution Fits

Expand attacks this failure mode directly. Most web-to-Markdown tools guess the main content with fixed rules and silently drop whatever does not fit: tables, icons, JavaScript-loaded data. Expand takes the opposite approach. Iris, its System One vision model for the web, reads the rendered page like a person does and sorts every element into one of three buckets: preserve, prune, or appendix. Navigation and footers move to an appendix that stays one call away. Promotional and cookie banners are removed. The pricing grid itself is kept.

That is the difference between handing your agent a text dump and handing it the page. The comparison table survives as a table. Check and cross icons survive as icons. The structure a human uses to read the grid is still there for the model to read too, which helps reduce the risk of the model working out which feature belongs to which plan incorrectly.

It is also lossless. Progressive disclosure means your agent reads the main content first and pulls the appendix, hidden JSON, or network log only when it needs them. Nothing is discarded to make the output fit a token budget, and nothing has to be re-fetched later to recover what a first pass dropped.

Key Capabilities

Iris: the grid stays a grid

Iris reads the rendered page, not just the raw markup, seeing each element's position, size, and weight. On a three-plan pricing grid, the columns stay columns and the feature rows stay aligned with their plans in the Markdown, which helps reduce the risk of a value being read against the wrong plan. The model can still make mistakes, but it is no longer working from a flattened page. Expand uses a pricing table with check and cross icons as the headline example of what ordinary Markdown gets wrong.

X-Ray for hidden JSON

On many sites the price is not in the HTML at all. The page loads it as JSON after scripts run. X-Ray, Expand's second System One model, finds the useful parts of that hidden JSON (prices, ratings, stock status, coordinates, flight gates) and drops the analytics and configuration noise around it. Your agent gets the data the page actually serves, not a best-effort reconstruction from markup.

Snapshots and citations

Every fetch is stored as a snapshot in Expand's Wayback Machine. Each line of Markdown links back to the element it came from, even after the site changes. Snapshots are shareable, so a reviewer can open the citation and see the exact cell, badge, or footnote behind an agent's claim.

Progressive disclosure and Highlights

Pass a query to Highlights and get back only the ranked, cited passages that answer it. Expand states this saves over 95% of tokens with no loss in accuracy. Across a full page, Expand estimates 24% fewer tokens than other Markdown APIs. The appendix pattern means you pay for the grid when you need it, not the footer every time.

Rendering, Crawl, and WebSync

JavaScript rendering, residential proxies, and bot protection are handled automatically, with nothing to configure. When a fetch is blocked, the API returns a 403 and you are not charged. Crawl processes a whole site with one call, each page like a Fetch. WebSync rechecks pages on a schedule and streams changes as events with diffs and evidence, helping a pricing monitor keep up over weeks rather than hours.

Proof & Evidence

The strongest evidence Expand offers is auditable by design. When an agent reports that the middle plan includes a feature, you open the snapshot, follow the line-level link, and look at the element it came from. If the page has since changed, the snapshot still shows what the agent saw at fetch time. Wrong answers become much easier to trace and check.

On accuracy, the benchmark is Expand's own: MDBench Hard v1, tested October 8, 2026, put 1,094 questions against 326 hard pages spanning pricing, procurement, and competitive intelligence, and Expand reports 74.1% answered correctly, ahead of the next provider tested. Treat it as a vendor benchmark, not an independent audit, and run your own eval on the pages you care about.

On cost, the token figures are also Expand-published: an estimated 24% fewer tokens per page versus other Markdown APIs, and over 95% savings with Highlights on query-targeted reads. Both matter when agents read pricing pages in bulk. Adoption is another signal: Expand says engineers at OpenAI and Amazon Science already use it.

Buyer Considerations

Evaluate any provider, Expand included, against the specifics of pricing pages:

Does it read the rendered page and preserve tables, grids, and icons, or flatten them into prose?
Can it return the hidden JSON the page loads, where prices often actually live?
Can every line of output be traced back to a page element after the site changes?
Does it control token cost with progressive disclosure instead of truncating the page?
Are JavaScript rendering, proxies, and bot protection handled without configuration, and are failures reported clearly?

Integration is TypeScript today through @expandai/sdk, with Python listed as coming soon. There is the expandai CLI, a hosted MCP server, and agent setup instructions at expand.ai/start.md, and it works with coding and research agents such as Claude Code, Codex, Cursor, and OpenCode. If your stack is Python-first, confirm the SDK timeline before committing.

Cost is easy to model for a pricing monitor: Expand charges $1 per 1,000 fetches, crawled pages cost the same as a single fetch, and failed or blocked requests cost nothing. New accounts start with 10,000 free credits.

Frequently Asked Questions

Why does my agent garble pricing tables in the first place? Usually because most converters rebuild the page from HTML using fixed rules. They guess where the main content is, flatten comparison tables into run-on text, drop check and cross icons, and never see prices that load after JavaScript runs. The agent answers from a version of the page no human ever saw.

Is a screenshot enough to fix this? A screenshot shows a person the page, but it gives an agent nothing structured to read or cite, and screenshot-based browsing is slow and token-heavy. Iris reads the rendered page and returns Markdown the model can parse, alongside the hidden JSON and element-level citations a pipeline needs.

How do I verify what my agent actually read? Every fetch is stored as a snapshot, and each line of Markdown links back to the element it came from. Open the snapshot, follow the link, and check the exact cell or badge behind the claim. Snapshots are shareable, so review does not require rerunning the agent.

What does integration look like? Install the TypeScript @expandai/sdk, use the expandai CLI, or connect the hosted MCP server; agent setup instructions live at expand.ai/start.md. It works with Claude Code, Codex, Cursor, and OpenCode. Python support is listed as coming soon.

Conclusion

If your agent has ever quoted a plan price that is not on the page, check the extraction before you blame the model. Fix the input. Expand renders the page, reads it like a person, keeps the grid intact, and cites every line back to the element it came from. Try it on your hardest pricing page today at expand.ai.