LEARNINGS | BEST FIRST VISITS | 16 OCTOBER 2026
20 positive examples from 64 qualifying GET checks among 200 one-attempt visits. Human labels: 0/30. USD 0 paid inference/API/cloud charges.
These ten structural patterns are descriptive counts, not causal rules or a claim that every winner shares every trait. The full 200-app measurement context is given without naming unsuccessful apps. Missing captures stay missing.
The leaderboard gate (fast opening, no script errors, a working reading path and Back) is selection, not a newly discovered pattern. No model inferred customer intent or ranked visual taste.

P01 | A description is available in page metadata: 19/20 top-list apps; 181/199 captured first views in the 200-attempt sample. Positive examples: https://bigwords.page/; https://boostmyagents.com/. Search/share text exists; wording quality is not rated.
P02 | The first page has an H1 heading: 19/20 top-list apps; 177/199 captured first views in the 200-attempt sample. Positive examples: https://bigwords.page/; https://boostmyagents.com/. A heading is present; this does not prove the purpose is clear.
P03 | The H1 contains a short readable line (5-120 characters): 19/20 top-list apps; 175/199 captured first views in the 200-attempt sample. Positive examples: https://bigwords.page/; https://boostmyagents.com/. Length is measured, not comprehension.
P04 | A main content landmark is present: 17/20 top-list apps; 155/199 captured first views in the 200-attempt sample. Positive examples: https://bigwords.page/; https://boostmyagents.com/. Semantic structure is present; accessibility was not audited.
P05 | A navigation landmark is present: 14/20 top-list apps; 149/199 captured first views in the 200-attempt sample. Positive examples: https://bigwords.page/; https://boostmyagents.com/. Navigation structure exists; no claim that every link is good.
P06 | At least three named same-origin links are available: 17/20 top-list apps; 148/199 captured first views in the 200-attempt sample. Positive examples: https://bigwords.page/; https://boostmyagents.com/. This offers more paths; only one safe reading path was tried.
P07 | At least three links are visible in the initial viewport: 17/20 top-list apps; 154/199 captured first views in the 200-attempt sample. Positive examples: https://bigwords.page/; https://boostmyagents.com/. This is a visible affordance count, not proof all links should be prominent.
P08 | The initial viewport includes visible image/SVG/canvas/video: 16/20 top-list apps; 178/199 captured first views in the 200-attempt sample. Positive examples: https://boostmyagents.com/; https://yaystack.ch/. Visual content is present; visual polish and relevance are unlabelled.
P09 | The desktop viewport has no horizontal overflow: 20/20 top-list apps; 199/199 captured first views in the 200-attempt sample. Positive examples: https://bigwords.page/; https://boostmyagents.com/. 1440px only, not a phone-layout result.
P10 | No form is visible in the initial viewport: 16/20 top-list apps; 152/199 captured first views in the 200-attempt sample. Positive examples: https://bigwords.page/; https://boostmyagents.com/. No early form is observed; a form later in the flow may be appropriate.

Coverage limit: this gate favors apps with a readable link-and-Back journey. It can exclude useful one-screen tools and can include a clean help/privacy path without trying the product job. It does not rank full task completion, demand, visual taste or mobile layout.
What changed in the checker: a blocked non-GET background request no longer prevents independent GET navigation checks. POST, forms, sockets, unsafe-action GETs and private-network destinations are still blocked; robots rules still apply to allowed GETs. This is a local measurement adaptation, not a deployed product change.
Next research step: compare machine classifications with human labels on these same captured views; keep unjudgeable cases separate and report disagreement by cause. The earlier 50-report run happened at another time, so it is not a paired accuracy comparison with these screenshots. Do not tune on these labels and call the same set held-out accuracy. No old/new improvement number is claimed here.
The private raw evidence and blind-label packet are intentionally not included in this public source package.