117 claims on 20 YC landing pages. 76% of them point nowhere.

26 September 2026

A landing page is a stack of claims. The problem is real, these people have it, we solve it faster, customers see a 40% lift. Some of those sentences come from somewhere: a study, a customer who agreed to be named, a public dataset. Most don't, and on the page the two kinds look exactly the same.

I wanted to know how often each kind shows up on pages written by people who are very good at this. So I took 20 homepages from the Y Combinator Summer 2026 batch and counted.

How the sample was picked

The batch has 231 active companies with a website in the yc-oss dataset. I drew 26 of them at random with a fixed seed, fetched each homepage in a real browser, and kept the visible text with navigation and footers stripped. Six pages didn't make it: three timed out, one rendered no text, and two had fewer than 50 words, which isn't enough to hold claims. That left exactly 20. I didn't pick or swap any of them.

A model then read each page and quoted every sentence that asserts something about the world: a problem, who has it, a result, a number, a claim to be first or best. It skipped plain feature descriptions like “exports to CSV”. For each quote it noted what the page itself points to as the source, if anything.

Then flawline xray checked every quote against the saved page, word for word. A paraphrased quote gets rejected, and a source only counts if the page actually carries it. All 117 quotes matched their page, so none were rejected.

What came back

Broken down by what the sentence is about:

Result claims do best, which makes sense: that's where testimonials and case studies go. The promise of what you get does worst, and that's the sentence most visitors act on.

The pages that point

The interesting part isn't the 76%. It's what the other 24% have in common.

Ekho Labs sells freight disruption forecasts. Next to “25 days later, seven forecasts confirmed” it links to a forecast-versus-outcome page, and the brief it quotes lists its inputs: EIA, Lloyd’s List, carrier filings, AIS. A forecasting company that shows its record is the strongest thing in the sample.

Mass Magnetics opens with a countdown to a supply restriction and quotes the statute behind it (10 U.S.C. § 4872). Its radioactive-waste figures are credited to Benchmark Mineral Intelligence, by name, under the chart.

Robocurve says robotics capabilities are moving fast and puts METR’s time-horizon chart under the sentence, source line included.

CarSignal compares itself with three competitors, plan by plan, and adds: “Competitor pricing and plan details are from their public pricing pages as of July 2026.” Its approval-rate stat is labelled “Live platform data, August 2026.” One date turns a boast into something a reader can check.

Most of these backed claims are about the world outside the company: a law, industry data, a competitor’s public price, a forecast anyone can compare with what happened. A source already exists for those, so pointing at it costs one line.

The pages that don't

The unsourced claims cluster around the company’s own results. I'm leaving these unnamed. The batch is weeks old, these pages were written in a hurry, and the point isn't any one team.

One page puts “95% forecast accuracy” in its hero stats, with no word on what was forecast, against what, or how it was measured. Put that next to the freight company above. Both sell predictions. One shows its record, the other shows a number.

Another has seven claims with numbers, and six of them have no source: an auto-triage rate, a response time, days of faster collections. The one it does back, a named customer story, is the one a buyer is most likely to click.

A third names a customer that “increased sales by over 50% in six months”, then right under it says “a San Diego restaurant almost doubled its revenue in the first month”, unnamed. Same page, same format, and only one of the two can be checked.

None of this means those numbers are false. flawline doesn't judge truth, and I didn't either. It means the page gives a reader no way to tell a measured number from a hopeful one, so a sceptical buyer discounts both.

The cheap fix

On most of these pages, the gap between a claim that points and one that doesn't is a few words: a date, a sample size, a customer's name, “internal testing”. One page already does it, putting “Pass@1 (n=5), select from the DataAgentBench leaderboard as of July 16, 2026” under its benchmark chart. Almost nobody reads that line, and it still makes the chart easier to believe.

To run the same check on your own page:

npx flawline xray landing.md

It prints what an agent should return, then checks the agent's quotes against your file, so the agent can miss a claim but can't invent one. In Claude Code it's /flawline-xray <file-or-url>.

Limits of this count

Twenty pages is a small sample, and one reader decided what counts as a claim. A different reader would find a slightly different 117. Pages from one YC batch, weeks after launch, aren't landing pages in general. I counted a named customer or a named person as a source, and an anonymous “HSEQ Manager, Fitout Contractor” as none. A few animated counters rendered as zero when captured, and I left those out rather than guess their values.

The sample script, the saved page text, every quoted claim and every result are in github.com/bilhokista/yc-xray, so you can check my count against the pages.

npx flawline xray landing.md · github.com/bilhokista/flawline