Home Service Website
Implementation Guide

0Read this first: what this is, and what it is not

This is an implementation guide. It documents the structure we would build for you, in enough detail that your developer, your marketing lead, or your AI can evaluate it before anyone signs anything.

It is written for residential home service contractors generally — HVAC, plumbing, electrical, roofing, and the adjacent trades. The architecture does not change between them. What changes is the service list, the symptom vocabulary, the seasonal pattern and the licence that has to be verified, and section 1.1 sets out exactly how those map across four trades. Where a concrete example is needed below, it comes from our reference build, which happens to be an electrical contractor; that is an accident of which site we built most recently, not the scope of the method.

How to read it quickly

• Evaluating the method? Sections 1, 4 and 5 are the substance. The evidence model and the gate system are what make this different from a template.
• Scoping the build? Sections 2, 3 and 6 are the specification a developer works from.
• Deciding whether to engage? Section 8 is what we need from you, stated honestly and up front.

The full specification, roughly 120KB across two documents, is at llms-full.txt. A 4,000-character condensed version for AI review is at brief.txt.

0bWhat this architecture is built to earn

Three mechanisms. Described by what they do, because what they achieve is something your analytics will tell you and this document cannot.

Win the query that pays

When someone searches a service and a town together, one page on your site is built to answer it. The intersection page at /[service]/[city]/ is self-canonical, and it holds a topic key that no other page of the same type may reuse. Being precise about what that buys, because it is less than it sounds: the check is within a page type, so it stops a second city page claiming the same market, not a pillar and a symptom page overlapping in intent. And a self-canonical tag states a preference — Google still picks the canonical it judges best, using its own signals. What this actually does is reduce avoidable overlap and force every page to have a stated purpose. That is worth having, and it is the half of the problem an agency is responsible for, but it is not exclusive ownership of a query and we are not going to describe it as one.

Convert the visitor who lands

Someone searching for a contractor in their town is looking for evidence that you work there and have done this job before. The architecture places that evidence on the page they land on rather than on an about page two clicks away: a documented job in that city, the technician who performed it by name, the licence number in the footer, and a review count read live from your profile rather than typed into a template. The evidence model exists so that proof sits at the point of decision.

Sustain it after launch

Thin city pages published once and never touched are the standard failure mode of local SEO, and they lose to sites that keep adding real work. The gates make that failure harder to commit: a page cannot publish without documented work behind it, and a page with no inbound content links fails the build rather than sitting unlinked. Staleness is the third control and it is policy rather than code today — section 4 says so plainly, because nothing in this build is old enough to have triggered it. The system is designed so that you cannot quietly accumulate pages you are not sustaining.

Where these rules came from

This did not start as theory. Fortitude Creative has built and run websites for home service contractors for years — HVAC, plumbing and electrical, across several states rather than one metro — which mostly means watching the same failures recur: thin town pages published in a batch and never revisited, a projects section that stops updating three months after launch, credentials and claims that drift away from what the business can actually evidence. Every gate in section 5 is a response to one of those, turned into something a build can check rather than something a person is supposed to remember. That is the whole idea — not that we are more disciplined than anyone else, but that discipline you have to remember is the kind that lapses.

What you can observe right now

These are not descriptions of intent. Each one is checkable in the next minute:

• The Crowley electrical-repair page renders its layout and states on the page that zero documented jobs exist for it, rather than filling the space.
• /electrical-panel-upgrade/crowley/ on the same site returns 404. The evidence threshold has not been met, so the page was never generated.
• proof/gate-log-cost-rule.txt — the cost gate rejecting a page and halting the build at exit 1, with steps to reproduce.
• proof/gate-log-thin-content.txt — the content-floor gate doing the same, and the defect that producing it exposed.
• proof/rule-examples.txt — eight gates copied out of the generator source.

What is not yet observable: a page crossing the threshold in the other direction. No documented job has been captured for this build, so no URL has gone from 404 to live. When one does, this section will say so and name the evidence items behind it. As of publication no page has completed the 404-to-200 transition. The two gate logs and the blocked 404 are the inspection targets. Until that changes, the system is only demonstrated in the direction of refusal, which is the honest half of the claim.

0cThese are not SEO rules

Every gate in this document enforces something a reader needs before they will believe a claim: work documented in their town, a named person who did it, a page that is about one thing.

That is the whole justification. None of these rules exist because a search engine is thought to reward them, and section 5 says so explicitly where the two numbers people ask about are concerned. They exist because a homeowner deciding who to let into their house at nine at night is running a trust check, and the things that satisfy a careful reader turn out to be the same things that hold up to scrutiny generally. If a rule here only made sense as a ranking tactic, it would not be in the build.

1The architecture

Seven page tiers. Every page belongs to exactly one, and every tier has a defined job.

HOMEPAGE
  │
  ├── TIER 1 · SERVICE PILLARS  — city-agnostic authority
  │     /[service]/                      e.g. /ac-repair/, /drain-cleaning/
  │
  ├── TIER 2 · SERVICE × CITY  — the money pages, evidence-gated
  │     /[service]/[city]/               e.g. /ac-repair/mesa/
  │
  ├── TIER 2b · SERVICE AREAS  — local entity + project feed
  │     /service-areas/[city]/
  │
  ├── TIER 2c · CONVERSION SURFACES
  │     /[service]/estimate/   /specials/   /brands/[brand]/[model]/
  │
  ├── TIER 3 · SYMPTOM PAGES  — urgent intent, top of funnel
  │     /[symptom]/                      e.g. /ac-blowing-warm-air/, /no-hot-water/
  │
  ├── TIER 3b · COMPARISON & DECISION
  │     /[option-a]-vs-[option-b]/
  │
  ├── TIER 3c · EDITORIAL  — optional; skip if nobody can sustain it
  │     /resources/[slug]/
  │
  └── TIER 4 · PEOPLE & PROOF
        /team/[technician]/   /projects/[job]/   /reviews/   /about/   /contact/

The decision that creates the stalemate

The query that pays is [service] [city] — "panel upgrade Crowley", "AC repair Phoenix". The common pattern is a service page and a city page published separately, with no decision made about which of them should rank for the combined phrase. The two compete for one query and neither resolves it.

We resolve it explicitly: a dedicated page at /[service]/[city]/, self-canonical, never canonicalised back to the pillar. The service-area page becomes an index and carries the local business entity; it does not compete.

1.1  The same taxonomy across four trades

Every tier below is the same structure with different vocabulary. This is the table to read if you want to know whether this applies to your trade.

2Page specifications

What each page type must contain before it is allowed to publish.

These specifications, built and live

The reference implementation below is an electrical contractor. Read it as evidence that the gates and templates run, not as the boundary of the method — the taxonomy in 1.1 is what changes for your trade, and it is the only thing that changes.

The wireframe implementing the table above is at allspark-seo-build.vercel.app, built for a licensed Texas electrical contractor. It is served with noindex, so it will not appear in search; it is reachable directly and by crawlers that ignore the directive. 45 pages. Every status below is checkable right now.

The two states are different on purpose, and both are shown. A gated page renders so the layout can be reviewed and states plainly what is missing. A page with no evidence and no review purpose is simply not generated, and 404s.

3Component inventory

UNIVERSAL on every site   CONDITIONAL when the business qualifies   VERTICAL trade-specific

4The evidence model

This is the part that makes the architecture different from a template, and the part that asks something of you.

Pages do not publish because someone wants them to. They publish when evidence exists:

What counts as a documented job

• Two before and two after photographs, original files, carrying their location metadata
• The address or nearest cross street, and the date
• The technician's name, spelled consistently
• A work order reference
• The named technician's confirmation that the write-up is accurate

A claim you can check in ten seconds

Open https://allspark-seo-build.vercel.app/electrical-panel-upgrade/crowley/. It returns 404. That is not a broken link, it is the rule holding: no documented panel-upgrade job in Crowley exists, so the page was never generated. When a job is documented, that URL starts resolving. Until then it does not.

Now open /electrical-repair/crowley/ on the same site. It returns 200 — and tells you, on the page, that there are zero documented jobs on file and that this is a layout for review rather than a real page. Two different treatments, because a wireframe has to show the layout somewhere while a live site must not ship an empty page. We would rather you saw both than only the flattering one.

Evidence goes stale

A documented job does not become fictitious at 19 months, and an absence of newer photographs does not prove the contractor stopped serving the town. So age triggers scrutiny rather than deciding the outcome: when the newest documented job in a market passes 18 months, the page is flagged for review, and a person answers three questions — does the business still serve this area, is the page still accurate, and does the evidence need refreshing. The page is re-gated only if those answers say so. What we will not do is let one job from years ago prop a city page up forever with nobody ever looking at it again.

5The quality gates

These are not guidelines in a document. They run on every build and they stop it.

Build fails

Two numbers people reasonably ask about

Why 350 words. It is not a ranking formula and we are not claiming Google counts words. It is the point below which a page on a service in a town has generally stopped containing diagnostic reasoning, scope, and anything locally specific, and has become duplicate-adjacent — offering nothing a reader could not get from five other sites. The gate is aimed at the pattern of thinness, not at an algorithm.

Why one, three and six documented jobs. These are not thresholds Google publishes and we did not reverse-engineer them from anything. They are our own requirement that you demonstrate you do this work in this place before we publish a page saying you do. One job is the least that makes an intersection page defensible; three is the least that makes a city page more than an assertion; six is what we want behind a page whose content target runs to 1,500 words, because a target that size without job history behind it gets met with padding. The thresholds exist to make the claim defensible, not to game a ranking factor.

Build warns

Title over 60 characters · under the tier word target · over 90 internal links · over 2,500 DOM nodes · page over 150KB · one or two inbound content links · near-duplicate topics at 0.8 similarity · llms.txt missing or stale · a face detected in a project photo without a consent flag.

6The schema model

Not a list of types. One graph, joined by @id, so the site describes a company, its people, and its work rather than emitting disconnected markup.

Organization /#org
  ├─ sameAs ─────→ business profile, BBB, socials
  ├─ employs ────→ Person /team/[name]/#person
  └─ hasPart ────→ LocalBusiness /service-areas/[city]/#localbusiness
                      ├─ areaServed ──→ City
                      ├─ sameAs ──────→ profile ID for that location
                      └─ makesOffer ──→ Service /[service]/#service

Person
  ├─ worksFor ──────→ /#org
  ├─ hasCredential ─→ licence, certifications
  └─ linked from each Project via author

Project (typed Article)
  ├─ about ─────────→ Service
  ├─ author ────────→ Person
  └─ contentLocation → LocalBusiness

Emission rule: every page emits its slice of this graph, and every @id it references must resolve to a node emitted somewhere on the site. The build enforces it.

7Content rules

Cost

Cost intent is wanted — pages should target "cost to replace a [system] in [city]" — a furnace, a water heater, a panel, a roof and rank for it. Cost numbers never appear. Pages answer with what drives the number: labour complexity, access, equipment age, emergency versus scheduled, repair versus replacement, then "we provide a detailed estimate after assessing the job." Enforced by a build gate, not a style note.

What a page must do instead. Naming the variables is not optional, it is the substitute. A cost page that says “it depends, call us” is worthless; one that explains that a panel replacement turns on the service size the house actually needs, meter and mast condition, whether the run has to be re-routed, permit and inspection requirements in that jurisdiction, and whether it is scheduled or an emergency, has answered the real question and given the reader a way to think about the quote when it arrives.

Expertise

We write what is true everywhere: how a system behaves, what a symptom indicates, what is safe for a homeowner to check. We stop at the jurisdiction line. Code editions, local enforcement policy and regional housing-stock patterns are marked on the page as questions for the licence holder, and they stay marked until answered.

A confidently wrong code citation is worse than a visible gap. A tradesperson reading it knows instantly that nobody competent wrote the page, and so does a well-prompted AI.

Attribution

Where we draft a project write-up from a briefing, the page states the work was performed by a named technician. It never claims that technician wrote or approved the write-up unless they did.

Banned language

"Leading provider of", "best in class", "100% satisfaction guaranteed", "your trusted partner". Not because they are ugly, but because they are evidence of absence: a business with a named technician, a dated job and a verified licence does not need to claim it is trusted.

8What we need from you

Stated up front, because the architecture does not work without it.

The routine loop, per documented job

About five minutes of your people’s time per documented job, across three touches: roughly two minutes for the technician to photograph and note the work order, two for the office intake, and about a minute later for the technician approval that the attribution rule requires. Earlier drafts of this table left the approval step out of the total, which undercounted it. If your crews already photograph work for warranty, insurance or callback protection, the marginal cost is the work order number and nothing else. If they do not, this is a genuine process change, and a small one, but the evidence model does not function without it. That is the honest trade and it is why it appears twice in this document.

The reference build runs the generator locally and deploys to a static host; photographs arrive through a shared drive folder, because chat apps strip the location metadata out of images and a drive does not. Per client we set up whatever intake form the office will actually use. The tool names matter less than the fact that the gate runs before the deploy, not after.

Where to look next

• The full specification — sections A to R, the complete rule set
• /proof/ — build logs, gate source, checksums
• brief.txt — 4,000 characters, sized to fit an AI context window whole
• llms-full.txt — both documents, complete, one fetch
• The off-site protocol — because on-site work is necessary and not sufficient

9How we measure whether a site is actually built well

These are the structural measures we hold our own work to. They fall into three groups, and the groups matter: they are not equally verifiable, and an earlier version of this section claimed they were.

Countable from the published pages, by anyone

No analytics access, no cooperation from the site owner, no judgement call. A crawler or a capable assistant can produce these numbers from public HTML.

Requires a browser measurement, not a crawl

Requires records, and cannot be verified from outside

This is the group we overstated, and it is worth being exact about because the whole evidence model rests on it. A page can show two photographs and a technician's name. That establishes that the page claims a documented job. It does not establish that the job happened, that the photographs are of it, that the work order exists, or that the named technician read and approved the write-up.

So “evidence coverage” is two different measures wearing one name. From outside, you can count how many pages present documented work. Whether that documentation is sound is an audit, and the right person to run it is the contractor whose jobs they are. We would rather say that than let a countable proxy stand in for a verified one.

10If you want to hand this to an AI, here are the questions worth asking

You are going to paste this into an assistant. Everyone does. So here are the questions that actually test it.

We are deliberately not supplying prompts tuned to produce a favourable answer. An evaluation you can only pass because we wrote the exam is worth nothing, and a competent model will notice the framing and mark us down for it. Three of the five below are ones we would genuinely prefer you skipped.