Section design examples

6,918 sections captured and categorized, from 2,587 funnels x-rayed.

Section

Showing 24 of 6918 sections

memdeklaro.org — a real FAQ section
memdeklaro.org · FAQ
promptfix.online — a real Pricing section
promptfix.online · Pricing
loweef.io — a real FAQ section
loweef.io · FAQ
applybtn.app — a real FAQ section
applybtn.app · FAQ
reassign.ai — a real Pricing section
reassign.ai · Pricing
echorelay.dev — a real Hero section
echorelay.dev · Hero
sakara.com — a real FAQ section
sakara.com · FAQ
justsimple.tools — a real Hero section
justsimple.tools · Hero
waitlistos.com — a real Pricing section
waitlistos.com · Pricing
redditoro.com — a real Pricing section
redditoro.com · Pricing
findvyb.com — a real FAQ section
findvyb.com · FAQ
longcatworld.com — a real FAQ section
longcatworld.com · FAQ
top10vpn.com — a real Logo Wall section
top10vpn.com · Logo Wall
topaiwritertools.com — a real Pricing section
topaiwritertools.com · Pricing
tryunlockdm.com — a real Features section
tryunlockdm.com · Features
stealthgpt.ai — a real Logo Wall section
stealthgpt.ai · Logo Wall
tarotai.chat — a real Hero section
tarotai.chat · Hero
aipalmreading.net — a real Hero section
aipalmreading.net · Hero
maxaeo.cn — a real FAQ section
maxaeo.cn · FAQ
adsturbo.ai — a real FAQ section
adsturbo.ai · FAQ
blockposter.pro — a real Pricing section
blockposter.pro · Pricing
vehla.app — a real Features section
vehla.app · Features
apexsuiteai.com — a real Hero section
apexsuiteai.com · Hero
posforshops.com — a real FAQ section
posforshops.com · FAQ
2587 funnels x-rayed3919 pages measured10 section types

X-ray your own funnel · free, no signup required

Method: every crawled page is segmented into full-width bands and each band classified into a section archetype by deterministic DOM signals: no LLM, no opinion. Prevalence counts each archetype once per page, across pages crawled all the way to the footer (partial crawls are held back, not guessed). "Typically sits" = the average of the section's position within the page body (0 = top, 1 = bottom). Scoped to the pages we've crawled, not "the web"; auto-updating as the corpus grows.