You open GA4, the CRM, and last month’s vendor deck side by side. Three totals. Three stories. None of them tells you which channel actually sold the car.
That’s the everyday reality in most dealership back offices. GA4 says paid search drove the traffic. The CRM credits a third-party lead. The social vendor’s report claims the assist. The DMS, meanwhile, records a sold VIN with a trade, F&I product penetration, and a gross number — and no honest thread back to any of them.
So let me answer the question early, then walk through the measurement and where it breaks.
Attribution tells you where observed conversions appeared. It does not tell you what your advertising actually caused. Those are two different questions. You need both answers, and you need to stop treating one as the other.
The four layers, and why they never line up
Dig into most dealer attribution arguments and they’re really definition arguments in disguise. Two vendors count “leads” differently and the meeting devolves. So before anything else, I want to understand the architecture as four distinct measurement layers, each answering a different question.
1. Delivery. Impressions, reach, frequency, clicks. Did the ad run and did anyone see it. This is the media layer. It says nothing about intent and nothing about outcome.
2. Behavior. Sessions, vehicle detail page views, inventory searches, phone calls, chat sessions, form submits. What shoppers did on your properties. This is where GA4 lives, along with your call tracking and chat platforms.
3. Sales process. Qualified lead, contact attempt, appointment set, appointment shown, write-up. This is the CRM’s job — VinSolutions, DealerSocket, Elead, whichever you’re on. Every store has opinions about what “qualified” means, and every store’s opinions are slightly different.
4. Business outcomes. Sold VIN, front and back gross, trade acquired, F&I products, repair order down the line, repeat purchase in three years. This is the DMS. This is the only layer that pays the bills.
The technical work of attribution is connecting those four layers without letting definitions drift between vendors. That means persistent campaign parameters and click IDs riding through every website event. Source data captured on calls and chats, not guessed at. Consistent lead-source rules in the CRM enforced with training, not hope. The DMS as the source of truth for the sale and the gross. And lawful, deduplicated customer matching across the whole chain — same person, same household, one record.
Get that plumbing right and the four layers reconcile. Get it wrong and every meeting is a definitions fight.
Platform math is not dealership math
Every ad platform now wants you to optimize toward a deeper outcome — not a click, not a lead form, but something closer to the sale. That’s the right instinct. It’s also where the double-counting starts.
Google recommends enhanced conversions for leads as its offline-import method: you send hashed customer data and sale confirmations back to Google, and its models attribute the sale to a Google-observed touch. Meta permits offline CRM events through the Conversions API and does the same thing on its side. Both are working as designed. Both are also grading their own homework.
If you sum what Google claims and what Meta claims and what your video vendor claims and what your third-party lead provider claims, you will attribute the same sold VIN two or three times. It’s common to find dealer groups whose vendor-reported “influenced sales” exceed total sales for the month. That’s not a rounding issue. That’s a definitional one, and it will not fix itself.
The discipline is to pick one source of truth for the sale — the DMS — and let each platform’s report be an input, not a verdict.
Descriptive credit is not causal proof
Here is where most attribution conversations quietly go off the rails. There is a difference between describing which channels appeared on the path to a sale and proving which channels caused it. Attribution models describe. Only experiments prove.
One historical case makes this concrete. A 2019 Applied Marketing Analytics study looked at 300 dealerships, 420,000 consumers, roughly $72M in annual media, and 18 touchpoints per buyer on average. When they ran a Markov attribution model against that coalition’s paths, the reported returns by channel shifted materially versus last-click. In that particular dataset, social and video preroll came out around $2.50 per media dollar, and paid search came out around $0.89.
I want to be careful with that. It is one 2019 case, not a benchmark, and I would not carry those ratios into a 2025 budget meeting. The point is not the numbers. The point is that the model chose the answer, and the paths the model observed were not randomly assigned. Shoppers who saw preroll and shoppers who searched Google are different people making different journeys. Any model that ranks channels off observed paths is describing a correlation, not proving a cause.
The industry knows this, and the industry is unhappy about it. IAB’s 2026 survey of more than 400 senior US planning and analytics decision-makers found that 60 to 75 percent said current measurement approaches underperformed on rigor, timeliness, and trust. That is not a fringe complaint. That is most of the people paid to answer this question saying the answers they get are not good enough.
The rule I actually use
So here is the working rule, and it is the whole point of this post:
- Attribution — GA4, platform reports, multi-touch models, the CRM stitching — tells you where observed conversions appeared. It is descriptive. It runs continuously. It is useful for pacing, creative decisions, and week-to-week operations.
- Causal tests — holdouts, geo experiments, matched-market tests, incrementality studies — tell you what your advertising actually caused. They are the only honest way to know if a channel is doing work you would not have gotten anyway.
Use both. Run attribution to describe the flow. Run experiments to prove the lift. Any dealer group spending north of a million a year in working media that has never run a proper geo holdout on a major channel is, in my view, allocating on faith.
Where this lands for the store
Two bridges out of this, and they are both operational.
The first bridge is budget. You cannot allocate what you cannot attribute, and you cannot defend what you cannot prove caused a sale. When the OEM co-op cycle comes around, or when a vendor pitches a new channel, the question is not “what does your report say.” The question is “what would the DMS show if we turned this off in three matched markets for eight weeks.” If nobody at the table can answer that, you are not doing attribution. You are doing storytelling with numbers.
The second bridge is lead handling, and it’s the harder one to hear. A sold VIN you cannot trace back to a source is very often a lead your store lost somewhere between the click and the contact attempt. The attribution gap and the process gap are the same underlying problem. Fix the plumbing between website, call tracking, chat, CRM, and DMS, and two things happen at once: the reports start agreeing, and the BDC starts closing the leads it was quietly dropping.
That is the honest state of automotive marketing measurement right now. Four layers, three vendors grading their own work, one DMS that actually knows what happened, and a decision that lives with you.
If your rooftops are running six-figure monthly media without a clean line from click to sold VIN — and without a single geo test on the books this year — that is the work. Auditing the architecture, consolidating the vendor sprawl, enforcing lead-source discipline in the CRM, and standing up the experiments that tell you what your advertising is actually causing. It is not glamorous. It pays for itself in one quarter of reallocated budget, and it changes what the monthly meeting sounds like.
That’s the conversation I’d want to have with your group next — working alongside the business, not over the top of it.