Test, not a rollout

Five leads built to the new spec, and scored

v2 is untouched and still live. This is a parallel /v3/ build so the two can be compared before anything becomes a rule. Leads chosen to span the range: Willow & Oak is the fattest lead we have, KJP the thinnest.

The scores

Scored by me, which is worth stating plainly: I built these, so treat this as a structured self-assessment rather than an independent verdict. Owner card scored on a 500px phone frame, since that is where the QR lands.

Lead	Source material	Scorecard A — owner					A	Scorecard B — homeowner					B
		Looks dearer	Recognise themselves	Could not write it	Nothing wrong	Would send		Clear in 5s	Trust pre-scroll	Answers the query	Easy to contact	Learned something
Willow & Oak tree surgery, fat	85 reviews, 3 specific quotes, 5 real photos, own copy	5	5	5	4	5	24	5	5	4	4	5	23
Moe's Group landscaping, mid	3 quotes, real photos, 3 trades	5	4	5	4	4	22	4	4	4	4	5	21
Cheltenham Tree tree surgery, mid-thin	1 site testimonial, 1 usable photo, gallery refused	4	4	5	4	4	21	4	3	4	4	5	20
Greaterscapes landscaping, thin	199-word source site, 0 harvested quotes, 65 reviews unread	4	3	5	4	4	20	4	3	4	4	5	20
KJP Fencing fencing, thinnest	No usable photos (generated hero), 3 site testimonials, 4 services	4	3	5	3	3	18	4	3	4	4	5	20

What the spread says. The gap between fattest and thinnest is 6 points on the owner card and only 3 on the homeowner card. That is the guidance and FAQ layers doing their job: they are Tier 2 and Tier 3, so a thin lead gets exactly the same substance as a fat one, and Reader B barely notices the difference. Reader A does, because Reader A is looking for their own business in it, and that is Tier 1 which we cannot manufacture. Every score below 4 on the owner card is a Tier 1 shortage, not a template fault. That is the useful finding: better builds will not close it, better source data will.

Where the two cards disagree

KJP is the case the framework predicted: A 18, B 20. Useful page, but the owner sees a generated hero and only four services and does not fully recognise the firm. Per the spec that is the "never gets bought" quadrant, and the fix is harvesting their Google reviews and getting real photographs, not changing the template.

What the spec asked for, and what actually got built

Spec item	State	Evidence
Tier 4 variant pools, 5 each	done	process, areas, enquiry, guidance intro, mid-CTA. Selection deterministic and collision-free within a sub-cluster.
Neighbour test	passes	Moe's vs Greaterscapes went from 5/5 identical blocks to 0/5. Process headings now differ: "From first call to finished" vs "What happens after you get in touch".
Review header branching	done	Willow & Oak specific, Moe's/Cheltenham/KJP generic, Greaterscapes none. KJP correctly demoted after a fix.
Em-dash in title tag	fixed	All five clean. Was broken on every build.
Style toggle on customer page	fixed	Injected by JS only when `?style=` is present.
Duplicate hero, two H1s	fixed	One hero in the DOM, one H1 on all five.
Footer NAP trailing comma	fixed	Empty county no longer leaves "Stroud,".
FAQ block + FAQPage schema	done	6 questions per niche, 18 written. LocalBusiness + Service + FAQPage JSON-LD on every build.
Tier 3 region layer	partial	Cotswolds cluster only, 2 cards per niche, 1 rendered. Spec asked for 3 variants per cluster.
Durability test on trust bar	fixed	"10 years of tree work" became "Since 2016 in Nailsworth".
Mid-page CTA after reviews	done	Present on all five.
Areas: towns only, 4+ fallback	done	Counties filtered; KJP went from "Cheltenham, Gloucester, Gloucestershire" to 5 real towns.
Four-service layout variant	done	KJP renders 2x2 rather than a gapped 3-col grid.
Opening hours in NAP	not done	We do not hold opening hours for any lead. Would have to be invented, so it is omitted.
Blog: 2 national + 3 local	not done	Local article angles are written in `region.mjs` but not yet generated. Articles remain the 5 national ones.
Unique blog text per site	not done	Articles are still Tier 2 shared, so the duplicate-content problem persists on `/advice/`.
Article schema + internal linking	not done	Advice index still points at `index.html#services` rather than the specific service.
Reviews-to-services loop	not done	Willow & Oak's top review still names stump grinding, which is not in the six services.

Three things I got wrong while building this

The cohort was silently empty. I referenced region before it was declared, inside a try/catch that swallowed the ReferenceError. Neighbours kept identical copy and the build reported success. The catch now rethrows programming errors.
My neighbour test was wrong before the product was. It matched "How it works" in the nav link rather than the section heading, so it reported a failure that did not exist and then hid a real one.
The specificity detector matched the business name. "The quality of service offered by KJP Fencing was highly impressive" scored as specific because the word "Fencing" is in the company name. That is the textbook generic review, and the false positive would have restored exactly the header mismatch the branching exists to prevent. The firm's own name is now stripped before testing.

Open question for you

The spec says Tier 4 needs "five minimum each, never repeats within a postcode district". Five variants covers a sub-cluster of five firms. The Gloucestershire landscaping cohort already has nine names, and if we keep sourcing at this rate the pools will be exhausted within a wave or two. Options are more variants, tighter clusters, or accepting repetition beyond the immediate neighbours. I would take tighter clusters plus a warning when a cohort exceeds pool size, but it is your call.