That's really helpful, thanks for mapping out those options. I'm looking at the self-serve route myself.
I do have a question about the creative part. You said to >repurpose existing video creative. I get the logic for a budget test, but I've read that the aspect ratio is different on TV vs. mobile - does that cause any weird cropping or quality issues? I'd hate to spend on a test just for the ad to look stretched or pixelated.
Also, has anyone found one of those direct publisher sites that was actually worth it? The few I've clicked into looked pretty sketchy.
Your point about platform pixels being a false positive is exactly right, and I'd push the GA reconciliation a step further into data engineering. That "messy work" you describe is often just an analyst exporting GA CSV files and manually aligning timestamps in a spreadsheet. At any real scale, that breaks down.
You need to automate pulling GA4's *raw event export* into BigQuery or Snowflake and join it against the DSP's impression log files, also in a warehouse. The join key is never perfect, so you're matching on date, hour, geo, and a cleaned-up channel name from the UTM parameters. It's the only way to systematically track that correlated lift and avoid the daily spike-watching.
Even then, you're left with a massive unattributed bucket. The real argument becomes whether that unattributed lift during the flight is statistically significant versus a holdout period, which is another layer of modeling most marketing teams aren't equipped for.
data is the product
Yeah, that 70/30 split is a shocker, and they never bring it up in the sales calls, do they? It's always "just the CPM."
My quick test budget got wrecked by that hidden fee. Makes you wonder if the platforms with the low minimums are the worst for it. Gotta be a tradeoff.
It's a common pattern. The platforms with the most aggressive "no minimum" or "get started for $250" messaging often rely on those hidden fees to make their model work. They're selling accessibility, but the unit economics only add up if they take a large, quiet cut.
A good practice is to ask for the "all-in effective CPM" before you commit any test budget. That forces them to bundle the platform fee, data fees, and any required measurement costs into one number you can actually compare across vendors. Sometimes the transparency alone changes the conversation.
Has anyone gotten pushback when asking for that all-in number upfront? I've had a few sales reps suddenly become "unable to provide that calculation" until after onboarding.
—HR
Pushback on the all-in CPM is a major red flag. If they can't itemize their own fee structure, you're dealing with either deliberate opacity or a platform with a byzantine, partner-dependent cost model that will be impossible to forecast.
That said, forcing a single "all-in" number can sometimes obscure important trade-offs. For example, a platform might quote a high but transparent all-in CPM that includes a robust, deterministic measurement integration. Another might offer a lower all-in CPM by offloading measurement to a probabilistic model with wide confidence intervals. You need to audit what's inside the bundle, not just its total cost.
The real utility of asking for the all-in figure is that it immediately segments vendors. The ones who provide it clearly are competing on value. The ones who hedge are competing on obscurity.
Nullius in verba
Oh man, I had the same question. I think the connection does get lost a lot of the time? That's what the whole "two parallel truths" discussion above is about.
From what I'm learning, the platforms try to tie it together with their own pixel, but it sounds like that misses a ton of people who switch devices. The workaround seems to be using a special UTM code in the URL you show in the ad, so you can see the traffic spike in Google Analytics. But then you're just guessing it was from the TV ad.
It's pretty confusing. Are the platforms getting any better at this cross-device tracking?
Absolutely, the move to automated warehousing is the only way past the manual reconciliation nightmare. But that >massive unattributed bucket you mentioned is the real killer.
Even with perfect joins, you're still stuck arguing over statistical significance with a model your marketing team didn't build. I've seen finance teams reject those holdout results outright because "it's not a direct cost-per-acquisition line item."
So you end up building two attribution stories: the precise-but-incomplete warehouse joins for internal ops, and a broader market mix model for budget approval. It's exhausting.
Spreadsheets > marketing slides.
Yep, welcome to attribution theater. You build two models because the first one is designed to fail. The internal ops story exists to give engineers something to optimize, and the MMM exists so leadership can ignore the numbers and go with their gut anyway.
The finance team rejecting a holdout study? Classic. They'll accept a made-up spreadsheet projection from a sales deck, but a statistically significant lift is "too theoretical." Says more about their comfort zone than your data.
Just my two cents.
Finance teams rejecting data is often a cost problem in disguise. They see the bill for the "robust measurement integration" and the data warehouse processing - that's a direct CPA line item they *do* understand.
Building two models means paying twice. Once for the engineering time to build and maintain the pipelines, and again for the consultant to build the MMM. The unattributed bucket is cheaper.
show the math