BenchRank
Free to readMethod6 min read · 18 August 2026

How we read a landing page

Thirteen questions, 181 possible answers, one screenshot. The method behind every figure in the gallery — and the honest limits of what it can tell you.

By BenchRank

On this page

Every landing page in this catalogue is read the same way: from a full-length screenshot of the page as it actually shipped, against a fixed vocabulary, by a model that is shown the image and nothing else. No analytics, no traffic data, no interviews. What the page says about itself, and what it looks like saying it.

Thirteen questions

The reading is split into thirteen dimensions, in two families. Six of them are what the page says — the angle it leads with, the grammatical form of the headline, what fills the hero, the order of the sections, how it asks for the click, what it puts up as proof. Seven are how it looks — the interface claim in the hero, the imagery, the layout, the type, the colour, and what the whole thing adds up to.

The split matters more than it sounds. An earlier version had one twenty-value bag called design, which mixed a palette decision, a layout decision and a type decision into a single answer. That made the interesting question unanswerable: what does a page that leads with pain agitation actually look like? A bag cannot be cross-tabulated. Seven countable questions can.

181 answers

Each dimension has a closed list of labels — 20 angles, 16 ways of asking for the click, 25 kinds of proof, 20 pricing structures, and so on to 181 in total. Closed lists are the point. If the model can invent a label, every page gets a bespoke one and nothing can be counted.

It can propose a new label when nothing fits, but a proposed label does not enter the vocabulary or any published figure until a person accepts it. That is deliberately expensive, and it is why the queue behind this gallery is longer than the gallery.

What this cannot tell you

  • Whether any of it converts. We have no traffic data for anyone else’s site, and any claim that one angle outperforms another would be invented.
  • What a page did last quarter. Each reading is of one capture on one date; pages change and the reading goes stale.
  • Anything about the product. The reading is of the marketing, and a beautiful page in front of a bad product reads exactly the same as one in front of a good one.

Why every number here carries its denominator

The catalogue holds 1,233 published products. The number whose homepage has been read is far smaller, and it moves every week. So no figure in this section is written as a bare percentage — it always states what it is a share of, in the sentence rather than a footnote.

That is enforced in the code rather than left to good intentions: the component that draws these statistics takes a count and a total, and has no shape without both. It reads modest while the corpus is small, which is the correct thing for it to read.

A percentage without its sample size is a claim about the world. With it, it is a claim about what we looked at.

— The rule this section is built on