The Visibility Stack: A Field Guide

Walk into the Monday manager meeting at almost any dealership and you will see the same scene. The CRM shows one lead count. The website vendor shows a bigger one. The call-tracking platform claims credit for half of both, and the GM is holding a DMS report that disagrees with all three. Fifteen minutes burn on whose number is right. The decision the meeting existed to make slides to next week.

That is not an AI problem. It is an architecture problem, and another dashboard will not fix it. What fixes it is a stack: five layers, built in a specific order, where each layer earns its keep by making the one below it usable.

I have been in this industry since 2005, starting as a technician in an independent shop, and I have built some version of this stack at every stop since: on the showroom floor, as a marketing director, as an agency owner, and now as an engineer building AI platforms. The order never changes. AI is the fifth layer. Most stores buy it first, and they pay twice for the privilege.

The stack at a glance

Here is the whole architecture on one screen. The rest of this guide walks it from the bottom up.

Two rules before we start. You build from the bottom. And every dollar spent on a higher layer while a lower one is broken is a dollar spent decorating a house with a cracked foundation.

Layer 1: The data layer

Five systems hold the raw truth of a store. Everything else is a view, a reskin, or a sales pitch.

• The CRM holds conversations: every lead, call note, appointment, and unsold follow-up.
• The DMS holds money: the deal, the gross, the trade, the F&I products, the RO history.
• GA4 holds behavior: which pages, which vehicles, which paths ended in a form or a call.
• Call tracking holds the phone: source numbers, recordings, durations, outcomes.
• Inventory holds the product: age, cost, price changes, photos, syndication status.

The data layer is healthy when you can answer one question without leaving it: show me this customer, this car, and this conversation, end to end.

The floor-level test

Pick one delivered unit from last weekend. Now trace it backward. Can you connect the DMS deal record to the CRM customer, the CRM customer to the call recording or form fill that started it, and that form fill to the page and the vehicle the customer was actually looking at?

If that walk takes more than ten minutes, or dies at a broken handoff between systems, you have found your first project. Not an AI project. A plumbing project.

Most stores fail the test at the same two joints: the CRM-to-DMS match, where two systems hold two different customer records with no shared key, and the website-to-CRM handoff, where the lead arrives but the source data does not survive the trip.

Layer 2: The definitions layer

This is the layer almost nobody builds, and it is the cheapest one in the stack. It costs one hard meeting and a page of writing.

Every tool in your store counts differently because every vendor benefits from counting generously. A chat widget counts a hello as a lead. The website platform counts a newsletter signup. The CRM counts whatever got typed into it, including the duplicate, the wrong number, and the service customer who wandered into the sales pipeline. None of these tools is lying. They are all answering a question you never defined.

One definition of a lead

Write a sentence your GM, your BDC manager, and your accountant would all sign. Give it edges: a uniquely identified consumer, not already in an active deal, who gave you working contact information and expressed interest in a vehicle transaction within a window you pick. Pick your own edges, but pick them once. Then make every report in the building use that sentence. Duplicates merge. Dead numbers do not count. Service-to-sales handoffs get their own bucket so you can see them instead of arguing about them.

One definition of a sale

Same exercise, harder edges. Does a sale count at contract, at funding, or at delivery? Which store gets credit on a dealer trade? What happens to the unwind? Decide once, write it down, and make the CRM, the DMS reporting, and the pay plans agree with the answer. If the definition of a sale changes depending on which report you open, your store does not have reporting. It has opinions.

Data contracts, in plain English

A data contract is not enterprise jargon. It is one page per integration that says: this system sends these fields, in this format, on this schedule, and here is who gets called when it breaks. When a vendor wants into your stack, the contract is the price of admission. Ask the simple version on the demo call: will you push the lead source into this field, every time, verbatim? The answer to that question tells you more than the rest of the demo combined.

Layer 3: The attribution layer

Attribution is where I learned the whole game, and I learned it because I had no other option.

In 2009 I was selling cars at Lazare Auto Group. By the time I was promoted to marketing director, one question ran my life, and it is the oldest question in the business: which of these vendors actually produces buyers? Every vendor had a report. Every report reached the same conclusion: we are the reason. They could not all be right, and the ad budget was real money.

So we built attribution ourselves with the tools of the era: early Google Analytics plus cookies, wired so the source of every lead was recorded by us, on our side, before any vendor touched the data. Vendor-independent by design. Once we could see which sources actually produced, we cut what did not work, fed what did, and the lean team I ran 4X'd lead volume.

The tooling has changed since 2011. The principle has not moved an inch.

Vendor-independent or it does not count

The rule: no vendor grades its own homework. Attribution data must land in a system the vendor cannot edit, in a format you defined at layer 2, keyed to records you own at layer 1. Today that usually means GA4 with disciplined tagging, call tracking you control, and source fields that arrive in the CRM untouched.

The test is one question: if a vendor disappeared tomorrow, would the history of what they produced disappear with them? If yes, you do not have attribution. You have testimony.

I applied the same discipline at Lia Auto Group when I brought SEO and SEM in-house across 18 dealership sites. Total traffic rose 34%, with organic accounting for 84% of it, and nobody could argue with those numbers because no vendor owned the measurement.

Layer 4: The decision layer

Everything below this layer exists so this layer can be small. The decision layer is not a BI suite. It is one number and one page.

The Monday number

Every store should have one number that opens the Monday meeting, chosen because it predicts the month while there is still time to change the month. Pick the one that fits how your store actually makes money. The criteria are strict: it must compute straight from layers 1 through 3 with no manual heroics, everyone must agree on its definition, and a bad reading must point at an action, not a debate.

The one-page scorecard

Under the Monday number, one page. Leads by source against the same period last month, using your definitions. Set, show, and sold rates. Aged inventory and what is being done about it. Front and back gross. Appointments on the board for the week ahead. If your scorecard needs a second page, you are hiding from a decision somewhere on the first one.

When my agency took a Honda dealer from #182 in the nation into the Top 25, from 150 units a month to 500+, with three consecutive record months above $1.2M in variable operations, the unglamorous core of that run was exactly this: one set of numbers everyone trusted, reviewed on a cadence nobody skipped. The proprietary lead-capture and conversion tech we built mattered. The shared numbers are what made it compound.

Layer 5: The AI layer, last on purpose

Now, and only now, AI earns its seat.

An agent answering after-hours leads needs layer 2 to know what a lead is and layer 1 to write the conversation somewhere durable. A forecast trained on vendor-reported attribution learns fiction. A pricing recommendation built on a dirty inventory feed recommends nonsense with confidence. AI on top of a working stack compounds. AI on top of a broken stack automates the noise and scales it.

Over 20 years I have been credited with helping generate a reported $2.4B in dealer profit, and the honest summary of how is boring: visibility first, automation second, in that order, every time. The order is the strategy.

One more reason the order matters: the layers below determine what your AI can even see. A model is a reasoning engine, not a forensic accountant. Give it five systems that disagree and it will reason beautifully from the wrong premises. Give it one set of numbers built on your definitions and your attribution, and the same model becomes the sharpest analyst in the building: one that reads every record, never sleeps, and never argues about whose number is right, because there is only one.

The visibility audit: cut, keep, connect

Run this in a week. One person owns it, the GM reads the result.

Cut

• Any dashboard nobody has opened in 30 days. Check the login history, not the meeting claims.
• Any tool whose numbers cannot be reconciled to your lead and sale definitions.
• Any vendor report about that vendor's own performance, as a source of record. Keep it as a claim to verify, nothing more.
• Any point tool duplicating a capability you already pay for somewhere else.

Keep

• The five core systems: CRM, DMS, GA4, call tracking, inventory.
• Any tool that exports complete, machine-readable data on demand without an extra fee.
• Any tool with a working, documented integration into your CRM that preserves source fields verbatim.

Connect

• CRM to DMS on a shared customer key, so a sold unit traces to its first touch.
• Website forms and call tracking into the CRM with the original source preserved, not rewritten.
• All five systems into one place you control. A warehouse is ideal. A disciplined weekly export into one spreadsheet beats ten beautiful dashboards that disagree.
• One written data contract per integration, with a named owner.

Start where the money is

You do not need a seven-figure data project to begin. You need one delivered unit and ten minutes: run the floor-level trace and let the first broken handoff set your priority list. The stack gets built one layer at a time, and every layer pays for itself before the next one starts.

If you want a second set of eyes on your stack, or you want the stack built, that is the work I do. See the work for what I have built and pricing for how an engagement starts.