The marketing system, reverse engineered from a funnel that already ran.

00 / Where the evidence comes from

Read this first or the numbers will mislead you.

The funnel data behind this document is not Larchmont's. It is Elite Northwest Flooring's: 8 delivery days, $252.42 of Meta spend, 10 leads, 7 booked, 3 estimates run, 3 quotes sent, $0 closed. Snohomish County, homeowner buyer, Facebook lead form.

It transfers because the shape transfers. Same county, same buyer age, same channel, same job size band, same problem of a two person operation trying to answer a phone while standing on a job. It does not transfer as a forecast. Nobody can tell you Larchmont's cost per lead from someone else's account, and this document never tries.

What Larchmont already has that the reference account did not

• A first party beacon that captures UTMs, platform click ids, a persistent visitor id, and fires on phone taps and text taps, because for a contractor the call is the lead.
• A lead store and dashboard at larchmont-analytics.vercel.app with source, campaign, and journey stitching.
• A nightly SEO bot that proposes, waits for a green light, and ships prebuilt changes.
• A Google Business Profile, a 153 page programmatic SEO property, and a GMB post library.
• A clean legal base. Fake proof was purged, the real registration is published, TCPA consent sits under every form.

The gap is not traffic and it is not tracking. The gap is that nothing catches the lead in the first sixty seconds and nothing carries it to a signature. That is what this builds.

01 / The finding that pays for the whole system

Seven of eight bookings happened between 1.3 and 5.2 minutes.

Not hours. Minutes. The reason is boring and mechanical: the calendar sat on the last step of the form, so the person booked themselves while they were still holding the phone.

The one lead who took 883 minutes to book is also the only booked lead who never got an estimate run. One data point, so treat it as a hint rather than a law, but it points the same direction as everything else.

This is the cheapest finding in the document and the most expensive to ignore. It costs no ad budget. It is a form layout and two automations. Every other build below is downstream of it.

02 / Do not build these

The expensive things that did not work.

This section exists because every one of these will be sold to Larchmont by someone, and four of them have hard evidence against them from the reference account itself.

03 / Evidence to build

Every build traced back to the number that caused it.

04 / The pipeline

Six stages, and the two where money actually dies.

Stage 3 and stage 5 are the whole game. Booking is close to solved by speed. Getting the booked person to actually be home, and getting the quoted person to actually sign, is where every dollar of the reference account died. WF-03 and WF-05 exist for exactly those two gaps and nothing else.

05 / GoHighLevel

The workflows, written so they can be built without a second conversation.

Twelve workflows. Each one names its trigger, its exact waits, and its exit. Every SMS line below has been run through the Larchmont copy gate.

06 / Measurement

Stop optimizing toward form fills.

The reference account fired its pixel on submit and nothing else. 153 link clicks produced 10 landing page views, and landing page views equalled leads exactly. That is not a funnel, it is a single event, and Meta optimized against it for eight days.

The event map

Wiring it to what already exists

Larchmont's beacon already produces the harder half: persistent visitor id, first touch source, click ids, per campaign UTM links, and phone and text tap events. The missing links are three:

1. The site's /api/lead posts to GHL by webhook as well as to the existing store, carrying the beacon's _lv_vid, utm fields and click ids into GHL custom fields.
2. GHL writes the appointment and the closed value back to the same store, so the dashboard's source to lead panel becomes source to revenue.
3. WF-12 sends the offline Purchase to Meta CAPI with the click id captured at first touch. Without the click id the conversion cannot be matched and the whole loop is decorative.

07 / The AI layer

Six builds worth making, in the order they pay.

Every one of these processes data that already exists or that the system above starts producing. None of them talk to a customer unsupervised, because the one that did produced zero bookings out of 167 messages.

1. Call intelligence. Build first.

Every inbound call gets recorded and transcribed, then mined for the objection, the service, the city, and the words the caller actually used. For a contractor the call is the lead, which means the call recording is the only unfiltered voice of customer data in the business. It writes the ad copy, it writes the FAQ, and it tells you which objection is killing stage 5. Feed the top objections into the quote ladder every month.

2. Creative defect gate. Build second.

Not a performance predictor. A mechanical check that runs before anything publishes, because predicting failure is far easier than predicting success. It fails a piece for: no written final line, first cut later than one second, wrong caption font, missing reg number on a paid piece, a claim that is not on the approved facts list, an em dash, or a banned word. It is the existing lint gate pointed at video and static instead of only at print.

3. Competitor ad monitor. Build third.

A weekly sweep of the Meta Ad Library across the service radius. It records days on air, variant count, offer, and relaunch behaviour. Those three are the only honest proxies that exist, because Meta publishes no cost per lead, spend, impressions or click through rate for commercial advertisers. Any tool claiming a competitor's cost per lead invented it. What the sweep is actually looking for is the lane nobody is running, which is how the open refinish lane got found in the reference market.

4. Quote autopsy. Build fourth.

Once thirty quotes exist with loss reasons attached, cluster them by service, job size, city, source, and reason. The output is a single sentence each month: this is the job we keep losing and this is why. It cannot run before the data exists, so it is fourth by necessity rather than by importance.

5. AI discovery monitor. Build fifth.

Assisted local discovery moved from roughly 6 percent to roughly 45 percent inside a year. That plausibly outweighs every posting tactic in this document. The monitor asks the major assistants the twenty questions a Snohomish County homeowner would actually ask, records whether Larchmont appears, and logs what got cited instead. Larchmont's AEO pages already exist for this; the monitor tells you whether they are working.

6. Nurture drafter. Build last.

Drafts the WF-10 variants so the every 45 days text is not identical for two years. A human approves the batch monthly. The model never sends, never replies, never holds a thread.

08 / The commercial lane

The branch with the biggest jobs and no marketing attached.

In the reference account every large quote on file came from a general contractor or a property manager, and not one of them came from an ad. The largest was $200,500 against an average residential quote of $12,517. Nothing in the marketing stack targeted that buyer.

Larchmont is a general contractor rather than a flooring sub, so the mirror image applies. The repeat volume buyer is the property manager, the insurance restoration referrer, and the realtor with turns. Larchmont already has a partnership SOP built for exactly this and it is not wired to anything.

Paid social has no role in any row. This lane runs on a named list, a calendar, and WF-08. It is slower to start and it does not decay when the ad budget stops.

09 / Profile and reviews

Fix the category before anything else in this section.

Larchmont's own GBP audit found the primary category set to Remodeler, which made the profile invisible for the terms that actually carry intent. Category is the one field that moves the most, and it costs nothing to change.

1. Set the primary category to the service that carries the most revenue, then add secondary categories for the rest. One primary, several secondary, reviewed quarterly.
2. Service areas by city, matching the site's city pages exactly. The place name law in the copy doctrine governs which names are allowed.
3. Reviews via WF-06, ungated, one ask plus one reminder. Larchmont currently has no verified review count, so no review number gets published anywhere until it exists.
4. Posts continue from the existing library. No phone number and no URL in post text.
5. Photos: real job photos only. The site photos are stock and disclosed as representative, which is fine for the site and not fine for the profile.

10 / Paid

Small, boring, and pointed at the feed.

What a test costs before you promise one

At realistic variance for a small account, detecting a 2x effect takes roughly 46 posts, or about 28 with topic matched pairs. A 1.5x effect takes roughly 138. A 1.25x effect is not reachable.

11 / Build order

What ships in what week, and what is already done.

12 / The numbers you have to supply

Five inputs. Without them this is a machine with no gauge.

Everything above is structural and holds regardless of the figures. What it cannot do is tell you whether to scale or stop. That takes five numbers that only Larchmont has, and none of them are in this document because inventing them would be worse than leaving them blank.

Once the first two exist the governing question stops being cost per lead and becomes a single sentence: how many jobs does one month of spend have to close to pay for itself. In the reference account that answer was one job in seven, and it had nothing to do with the $25.24 cost per lead everyone was staring at.