Enoma
AI-generated business pages for local service businesses (landscaping, plumbing, HVAC, etc.) — one page built to be found by Google and AI assistants, not a generic ten-page site. Free 30-day trial, then $19.99/mo.
Production: enoma.io. Vercel serverless functions (api/*.js, ESM) + static HTML in public/ + Supabase (Postgres + Auth). Stripe for billing, Resend for email, OpenAI for page copy and outreach drafting, GA4 + Search Console for traffic.
Vercel Hobby plan caps this project at exactly 12 serverless functions per deployment. Before adding a new api/*.js file, check ls api/ — if already at 12, extend an existing endpoint or move pure logic into api/_lib/ (excluded from the cap) instead of adding a 13th file.
Admin tools
All gated to [email protected] only, via Supabase session:
/admin/admin-list.html — every business profile/admin/voice.html — voice-driven assistant for revenue/traffic/CRM questions, plus outreach-email drafting/admin/sales-queue.html — Sales Queue, see below
Sales Queue
/admin/sales-queue.html, backed by api/ga-metrics.js?action=sales_queue and ?action=update_lead_status.
Answers one question: who should I talk to today? It merges the two lead populations that actually carry contact info into a single scored, prioritized list:
- Raw Enoma leads —
contact_submissions rows where business_id IS NULL: people who filled the "get your website" form or asked for the concierge build. First-party interest in Enoma itself.
- Outbound prospects — the Outscraper pipeline (
prospects + outreach_messages): local businesses found via Google Maps search filtered to only_without_website, worked through draft → approve → send → reply/claim.
Scoring model (api/_lib/lead-scoring.js)
Every lead gets a 0–100 score and a tier (hot ≥70, warm ≥40, cold below), from real signals only:
- Raw leads: base score by source (a concierge request scores higher than a bare form start), +10 for a phone number on file, +10 if unread, −25 once replied to.
- Prospects: base score by pipeline stage (new → reviewed → drafted → approved → sent), overridden hard by real response signals — a reply scores 90, a claim scores 100. Skipped prospects and bounced/opted-out outreach are excluded from the queue entirely, not just down-ranked.
- Both: a recency multiplier decays the score over ~30 days but never to zero — an old hot lead still surfaces, just not first.
Unit tests: tests/lead-scoring.test.js (pure functions, no network/DB — run via npm test).
What's deliberately not in the queue
Self-serve wizard drop-off (funnel_events) is not turned into individual leads. Those rows carry only an anonymous anon_id — no name, email, or phone — until after a business page is already generated (see public/scripts/funnel-track.js), so there's no honest way to attribute early drop-off to a real, contactable person. It's shown instead as an aggregate "last 30 days" funnel-health panel on the same page (step counts only, not leads).
Actions
Both are JWT-gated (Supabase session bearer token, [email protected] only) and were added to the existing api/ga-metrics.js rather than a new file, to stay within the 12-function cap:
action=sales_queue (GET) — returns the scored queue, summary tiles, and funnel-health counts.action=update_lead_status (POST) — mark a raw lead replied/read, or advance a prospect's status (reviewed → drafted → approved, or skipped).
Sales strategy & prospecting status (2026-08-13)
Target segments: plumbing is the primary scale bet, landscaping stays as a warm/proof lane, junk removal was evaluated and ruled out.
Real data behind the call — web-presence rate by trade (unfiltered Outscraper pulls, Attleboro MA, ~35–50 businesses sampled per trade — one town, directionally useful, not a large-scale study):
| Trade | No website | Real website |
|---|
| Landscaping | 24% | 72% |
| Junk removal | 11% | 86% |
| Plumbing | 12% | 88% |
| HVAC | 16% | 84% |
| Electrical | 24% | 76% |
- Landscaping has the widest "no website" pool of anything tested and is the only vertical with a real paying customer/case study (Conway's Landscaping) — left running exactly as-is (phone-based outreach, existing Outscraper pipeline, no new tooling needed).
- Plumbing was picked as the new scale bet despite a lower "no website" rate than landscaping, because it's a state-licensed trade — meaning contact sourcing is structurally solvable via public licensing registries, which landscaping and junk removal have no equivalent of. Also: national search volume for plumbing sits in the hundreds of thousands/month (third-party published estimates, not independently verified against a paid keyword tool), demand is comparatively steady year-round (unlike HVAC's extreme seasonality), and a $19.99/mo ask is trivial against typical job value.
p.js's schema.org mapping (Plumber) and the AI copy-generation prompt (trust badges, offers_emergency) already support this trade correctly — zero engineering cost to add.
- Junk removal was ruled out: same structural sourcing gap as landscaping (no licensing-registry equivalent), but a more saturated, already web-savvy market (86% real websites, many with hundreds of Google reviews) — a harder "you're invisible online" pitch than any other trade tested.
Known gap, not yet fixed: Outscraper's leads_n_contacts/company_websites_finder enrichment does not appear to actually return email data — confirmed empirically across 250+ real raw listings spanning 5 trades (including businesses with real websites, where an email should almost always be findable), zero email-shaped fields returned anywhere. Every prospect in the pipeline today is phone-only as a result. Don't trust firstEmail()'s guessed field names in handleProspectPull (api/ga-metrics.js) as working until this is separately investigated — likely needs async=true + polling rather than the current synchronous call, but that's unconfirmed.
MA plumbing license bulk data — investigated, not viable today: Massachusetts's plumbing/gas-fitting license system (licensing.reg.state.ma.us / ePlace Portal) is a one-license-at-a-time verification lookup, not a browsable or bulk-downloadable registry. A "Professional Licensing API" exists but requires vendor/municipality-level API access, not self-serve signup. This is a real potential lever (worth a formal public-records request someday), just not something to re-attempt the same way expecting a different result.
Fixed this session: handleProspectPull was reading p.site for a business's website, but Outscraper's actual field is p.website — confirmed against a real raw API pull. prospects.website/found_website had been NULL for every prospect ever pulled until this was fixed; invisible under the default only_without_website filter since Outscraper does that filtering server-side.
Open items / backlog (2026-08-13)
Not yet started, none of this exists in the codebase — confirmed via a repo-wide grep for UTM/attribution fields before writing this.
Visitor attribution tracking — the underlying question was "can we know who's visiting via organic search vs. direct vs. an AI assistant, and follow up on their interest." Current state: page_events/funnel_events capture the referring domain, but only per-request, not persisted — so if someone lands via Google then clicks around for five minutes, every later event's "referrer" is just enoma.io itself, not Google. No UTM params, no landing page, no first-touch source captured anywhere.
Real distinction worth keeping straight: an anonymous visitor can't be identified before they submit a form — that's not a tooling gap, it's how privacy-respecting tracking works. (B2B "visitor deanonymization" tools like Clearbit Reveal / RB2B exist and do reverse-IP-lookup tricks, but they're built for corporate-network traffic and would perform poorly against Enoma's actual visitors — small-business owners on home wifi/mobile.) The real, buildable value is at the moment someone does submit a form — that's when attribution can actually attach to a real lead:
- Client-side: capture
document.referrer + parsed UTM params + landing page once per anon_id, persist alongside it in localStorage (extends public/scripts/funnel-track.js).
- Server-side: attach that first-touch bundle to the
contact_submissions row when a real lead-capture happens (send-contact.js) — needs new columns (first_touch_source, utm_campaign, landing_page).
- Surface it in the Sales Queue — a natural column/filter addition to the existing raw-lead rows, not a new page.
AI-referral traffic specifically: referrer behavior varies by AI product — some pass chatgpt.com/perplexity.ai/claude.ai cleanly, others strip it. GA4 (already connected) has started bucketing some of this natively — check what's actually showing up there before building custom detection. If needed, a cheap addition is matching the referrer against a known-AI-domain list rather than relying on GA4's generic "referral" bucket.
Conversion chatbot — raised again 2026-08-13. Explicitly out of scope for the sales-queue work this session — see the "Sales strategy" section above; the original scoping decision was distribution tooling only, no new product surface (chat widget, SEO pages, referral loops all deliberately skipped). Not decided whether to revisit; flagged, not started.
"Claude Dispatch" — mentioned 2026-08-13, unclear what specific feature/product this refers to. Needs clarification before any work starts — don't guess and build the wrong thing.
Security note (2026-08-12)
outreach_messages and suppression_list (the permanent do-not-contact list) had Row Level Security disabled — publicly readable and writable via the anon key already embedded in the admin pages' page source. Fixed by enabling RLS with no public policies added, matching the existing prospects/funnel_events pattern (service-role-only access — nothing legitimate touches these tables from a browser). Run Supabase's get_advisors(type: "security") after any new table or migration to catch this class of issue early.
Preview pages, custom domains, monthly results (2026-09-30)
All server actions below are on api/ga-metrics.js (function cap) and need the x-admin-secret header.
Automatic preview pages — POST /api/ga-metrics?action=generate_preview&prospect_id=<uuid> (&force=1 skips the qualification gate). Builds an unclaimed enoma.io/<slug> page from ONLY the prospect's own website text + Google Business Profile (api/_lib/preview-generator.js: facts-only prompt, then validation strips numbers not in the source, rating talk, puffery, and badges/bullets with no grounding in the site text). Qualification: 4.5★+ with 5+ reviews, a phone number, and — if they have a website — 2+ failed objective site checks (api/_lib/site-checks.js). The daily crawl now runs this automatically for weak-site prospects before drafting, so the outreach email links their page. Previews stay visitable for 90 days (trial subscription row), noindex until claimed.
Outreach drafts — generateDraftCopy only states verified facts: the preview link and the objective site problems. It no longer claims a prospect is invisible on Google or cites other customers' results.
AI search — api/p.js now renders About / Services / Areas / Why / FAQs / Testimonials into the HTML (api/_lib/page-ssr.js) so crawlers that don't run JavaScript see the content, and emits FAQPage JSON-LD. The client script still re-renders the same sections; keep renderSections in sync with loadProfile() in public/profile.html.
Custom domains — set small_business_profiles.custom_domain (lowercase, no scheme, no www, e.g. grilloplumbing.com), add the domain (and www.) to the Vercel project, and have the customer point DNS at Vercel. middleware.js sends that host's / to api/p, which looks the business up by custom_domain; the canonical URL becomes the custom domain and the page drops out of enoma.io's sitemap.
Monthly results email — action=results_email runs on the 1st (cron). Every claimed, active customer with an email gets last month's visits, taps to call and quote requests. &dry_run=1 returns the emails without sending; &[email protected] sends them all to you to proofread. Idempotent per month (results_email_last_sent_at); results_email_opt_out turns it off for a customer.
