Garage Door Science — garagedoorscience.com

Last updated: 2026-08-07

Orientation only. Deep implementation detail lives in that project's own Claude memory at ~/.claude/projects/-Users-sethshoultes-Local-Sites-garagedoorscience/memory/ (24+ entries). This file exists so a session working from any repo understands what GDS is and how it relates to the other properties.

What it is

An independent, AI-first educational and diagnostic site for homeowners. Brand-neutral by design. It monetizes by routing leads to partner service companies rather than by selling anything itself.

Three product layers:

Interactive Three.js science labs — nine labs, browsable, and doubling as the RAG corpus
AI chat — a Virtual Technician (diagnostic) and a Lab Trainer (educational), with bidirectional handoffs
ZIP-routed lead handoff to partners: A Plus Garage Doors and Utah Garage Doors

The positioning bet: every other garage door site is form-first. GDS makes the homepage a conversation, which suits users arriving mid-conversation from ChatGPT, Claude, or Perplexity.

Where it sits in the portfolio

GDS is now the primary consumer platform for A+, following a strategic pivot on 2026-04-21. It replaced aplus-garage-door-chatgpt-app, and it will front any future MCP or ChatGPT app surface. Partner data (ServiceArea) is owned here in GDS's own Prisma schema, seeded from the A+ admin's JSON, with DB-backed routing live.

It is also a content source for A+: blog posts, labs, and videos are linkable from A+ pillar content as external references, tracked in pillar-content/data/garage_door_science_links_library.json. 93 GDS links are placed across 66 A+ posts.

Stack and infrastructure

Next.js 16 App Router, TypeScript, Tailwind v4, React 19, Turbopack
Vercel (team seth-shoultes-projects, project garagedoorscience), domain on Cloudflare
Prisma + Neon Postgres with pgvector for RAG
Vercel AI Gateway routing to Anthropic models; OpenAI for embeddings, voice, vision, image gen
Repo: github.com/sethshoultes/garagedoorscience (private)
npm, not pnpm — pnpm 10.2 trips on workspace detection

What has shipped

v1.5 (2026-04-21) — the major sprint:

Inspection tool at /inspection — 24-point checklist, per-item photos, report view, lead capture with Turnstile and Upstash rate limiting, Resend email reports, partner notification on red-flag items, and per-item AI chat
Dual-persona chat — Virtual Technician and Lab Trainer with invokeTechnician / invokeTrainer handoffs, per-lab and per-item scoping, clickable inline citations, thumbs feedback persisted to a ChatFeedback table
About page with persona avatars and bios

Developer surface:

MCP endpoints — /mcp public (no auth, top-3 results) and /mcp-pro (bearer gds_live_, top-10), sharing a handler in lib/mcp/server.ts
REST shim at /api/v1/<tool> with auto-generated OpenAPI 3.1 and Scalar docs at /developers/api
Developer portal at /developers — self-serve magic-link signup, Prisma-backed API keys
Admin dashboard at /admin/leads and /admin/videos

Content and media:

Auto-content Phase 1 — draft → edit → fact-check → URL validate → gate → publish, working end to end, first auto-article live
Editorial team — four AI-assisted writer personas plus Seth, with bylines on every article
Remotion + ElevenLabs video pipeline for blog companion videos, with an admin review queue

Gotchas worth knowing before touching it

Remotion staticFile() resolves to remotion/public/, not the project root public/. Stale files there silently break renders. The Remotion sub-project is local-only and never ships to Vercel.
Verify rendered audio with ffmpeg silencedetect before shipping. Truncated or stale audio otherwise passes unnoticed.
Use printf, not echo, when setting Vercel env vars — echo appends a newline that breaks the value.
ServiceTitan scheduler URLs come in two formats. A+ uses the new one (hardcoded in the schedule button); Utah Garage Doors still uses legacy. A custom markdown renderer intercepts any ServiceTitan URL and opens the shared modal, which fixes legacy-URL 404s site-wide.
The Trainer persona over-triggers scope-redirect without a dynamic lab-context preamble in its system prompt.
Ollama Cloud launches with ollama launch --model kimi-k2.6:cloud, not the standard local Ollama command.

Relationship map

A Plus Garage Doors (aplusgaragedoor.com)  ── owns ──>  Garage Door Pros (garagedoorproservices.com)
        ^                                                        
        │ routes leads to                                        
Garage Door Science (garagedoorscience.com) ── also routes to ──> Utah Garage Doors (utahgaragedoors.net)
        │
        └── supplies external reference links to A+ pillar content

GDS is deliberately brand-neutral in public. It should not read as an A+ property to a visitor, even though it is A+'s primary consumer surface.

Early search performance (2026-08-10)

Google Search Console, trailing 28 days:


Clicks	52, up 643%
Impressions	10,500, up 223%
CTR	~0.5%
US share	94%

The site launched April 2026, so that is roughly three and a half months from scaffold to a real inflection. For scale, A+ recorded 379 clicks in July.

What is actually driving it

Page	Clicks	Prior
Best Garage Door Weather Seal: The Three That Actually Matter	10	0
LiftMaster vs Chamberlain: Are They the Same Company?	10	0
Garage Door Insulation for Summer Cooling: What It Costs and What It Saves	6	0
Best Garage Door Opener Under $500 (2026)	4	0
Homepage — Interactive Labs & AI Technician	3	+200%

Every top page was previously at zero. This is new content ranking, not old content improving.

The pattern worth noticing

All four leading pages are purchase-decision content: which weather seal, which brand, what insulation costs, best opener under $500. None are diagnostic.

That is a mild tension with the product bet. GDS is built around diagnosis — the Three.js labs, the Virtual Technician, the inspection tool. The search traffic is arriving on buying guides. Both can be true and useful, but if the labs are the differentiator and the buying guides are the traffic, the handoff between them is where the value gets captured or lost.

Video search is a real channel

5 of 52 clicks came from video search, roughly 10% of all traffic. The Remotion and ElevenLabs pipeline is producing assets that rank independently, not just page decoration.

The 90-day view, and the biggest single opportunity

Search Console, 2026-05-09 to 2026-08-07:


Clicks	72
Impressions	18,100
CTR	0.4%
Average position	14.9
Distinct queries	862

Impressions inflected around 2026-07-08, climbing from roughly 100/day to 500/day and holding. Worth establishing what changed then — new content, an indexing event, or the auto-content pipeline finding its stride.

Top queries by impressions tell the real story:

Query	Clicks	Impressions
best garage door opener under $500	0	2,754
garage door repair cost	0	344
garage door balance test	0	333
problems arising when you don't lubricate your garage door	0	171
how often should you lubricate your garage door	0	147
garage door repair	0	125
how often to lubricate garage door	0	101
garage door opener	1	16
garage door rubber seal bottom	1	2
estimated cost	1	1

best garage door opener under $500 has 2,754 impressions and zero clicks. That is the single highest-leverage number on this site. Google is showing the page nearly 3,000 times a quarter and nobody clicks it. At an average position of 14.9 the page is sitting at the bottom of results page two, visible and unreachable.

The lubrication cluster shows the same shape: four related queries, roughly 420 combined impressions, zero clicks. GDS has content Google considers relevant and no position to convert it.

This is not a content problem or an indexing problem. It is a position and presentation problem. The top ten queries alone account for about 4,000 impressions and three clicks.

The practical implication: moving existing pages up beats publishing new ones. A page already earning 2,754 impressions needs rank, title, and snippet work, not a replacement. That is a materially different plan from "write more articles," and cheaper.

The link profile explains the page-two ceiling

Search Console Links report, 2026-08-10.

External links: 99 total. Effectively two are independent.

Linking site	Links	Independent?
aplusgaragedoor.com	65	No — same group
sethshoultes.com	26	No — Seth's own site
garagedoorproservices.com	6	No — A+ owned
alibaba.com	1	Yes
claudepluginhub.com	1	Yes

97 of 99 backlinks come from properties Seth or A+ controls. Google discounts links between related properties, so the site's usable external authority is close to nothing. That is the most likely reason a site with 18,100 quarterly impressions cannot break past average position 14.9.

There is a second, sharper problem. GDS is positioned as an independent, brand-neutral educational resource that "should not read as an A+ property to a visitor." Its link graph says the opposite: 65 links from one garage door company and 6 more from that company's subsidiary. Whatever the public copy says, the backlink profile describes an A+ property. Google sees the graph.

This is the same gap Corey identified for A+ in the 2026-08-07 meeting, where he argued local links now outrank citations and named chambers, Rotary clubs, charities, adjacent businesses, and suppliers. GDS needs the equivalent for an informational site: home-improvement publications, trade associations, tool and materials manufacturers, and university or extension resources. Genuinely third-party links are the constraint on this site, not content.

Internal links: 225 total, pointed away from what earns traffic.

Page	Internal links	Quarterly impressions
Homepage	91	—
`/labs`	88	—
`/labs/science-of-garage-doors`	29	—
`/labs/energy-efficiency-lab`	13	—
`/blog/sectional-vs-rollup-garage-door-commercial`	3	—
`/blog/garage-door-repair-cost-common-fixes`	1	344
`/blog/best-garage-door-opener-under-500`	not in top 6	2,754

The internal link graph funnels almost everything into /labs, while the pages actually earning impressions get one to three links each. The single highest-impression page on the site does not appear in the top six at all.

Internal linking is a ranking signal fully under our control, and it is currently being spent on the product bet rather than on the pages Google is already showing. That is a cheap, immediate fix: link the buying guides from the homepage and from relevant labs, and give the opener guide real internal support.

Taken together, the two findings explain the ceiling. Content is ranking well enough to earn 18,100 impressions. It has no external authority and no internal support, so it sits at position 14.9 and converts 0.4%. Neither problem is solved by writing more articles.

Note on overlap with A+ pricing content

GDS ranks for "garage door repair cost" (4,400/mo) via /blog/garage-door-repair-cost-common-fixes. A+ article 71 targets the adjacent "garage door cost" cluster — buying a door rather than repairing one. Different intent, so not direct cannibalization.

Worth watching rather than acting on. GDS ranking for informational terms and routing leads to A+ is the designed behavior, unlike the Toni's and Jim Dyer overlaps, which are two service companies chasing identical local terms. The distinction matters: GDS is supposed to intercept the question and hand off. Toni's and Jim Dyer are supposed to be the answer.

Related: PROFILE · garagedoorscience · publish-a-post · garagedoorscience-remotion-pipeline

Google Search data points

2026-08-10: 50 clicks / 28 days (first Search Console read, recorded in FINDINGS).
2026-08-25: 60 clicks / 28 days (Google "Search Impact" milestone email). +20% in two weeks; still page-two ceiling territory. On Bing the same period shows GDS-fed explainers earning the most Copilot citations of any A+ property (bing-baseline-2026-08).

Google AI features (AI Overviews / AI Mode) — GSC export 2026-08-28

data/gsc-gds-ai-2026-08-28/ (Search type Web, last 3 months, 2026-05-27 → 2026-08-26). 3,508 impressions in generative AI features, first half 141 → second half 3,367 (+2288%). By month: 2026-05 2, 2026-06 86, 2026-07 847, 2026-08 2,573. Peak day 2026-08-17 (150). Devices: mobile 1923 / desktop 1500 / tablet 85; US 87%. 41 pages appeared; the top ten are all cost/comparison explainers:

Page	AI impressions
/blog/garage-door-repair-cost-common-fixes	718
/blog/garage-door-repair-cost-common-fixes-2026	630
/blog/liftmaster-vs-chamberlain	562
/blog/sectional-vs-rollup-garage-door-commercial	492
/blog/how-often-to-lubricate-garage-door	238
/blog/garage-door-insulation-summer-cooling-cost	155
/blog/best-garage-door-opener-under-500	139
/blog/insulate-garage-door-summer-cooling	111
/blog/best-garage-door-weather-seal	56
/blog/garage-door-safety-sensors	54

Read with the Bing baseline (bing-baseline-2026-08): on both engines, GDS's AI visibility comes from cost and brand-comparison articles, not the labs. Note the near-duplicate repair-cost-common-fixes and …-2026 pages both rank — consolidate or differentiate.