Attribution and Measurement: Build the System, Accept the Uncertainty

Bottom Line:

• Every attribution model is wrong. Last-click, multi-touch, incrementality - all of them. The goal is useful, not perfect.
• Platform-reported conversions are inflated by 20-40% as a baseline.[^1] Calibrate with geo-holdout tests on your top 2-3 channels every quarter.
• The real competitive advantage is measurement infrastructure - event tracking, identity resolution, first-party data. Teams that invest here compound. Teams that don't reset every cycle.

Most teams have an attribution problem. They just misdiagnose it.

They think the problem is picking the right model. Last-click vs. data-driven. Multi-touch vs. incrementality. They spend weeks in analytics debates trying to find the one view that tells the true story.

The true story doesn't exist. Accept that early and you'll make better decisions faster.

Every model is wrong. Some are useful.

Last-click gives all the credit to the final touchpoint before conversion. Search and retargeting look like heroes. Social, content, and brand channels look useless. Your media mix shifts toward bottom-funnel. Top-funnel atrophies. Eventually the pipeline thins and you can't explain why.

First-click has the opposite problem. It over-credits awareness channels and ignores the work that closes the deal.

Multi-touch models distribute credit across the journey. They feel fairer. But they're built on assumptions about how much each touchpoint contributed - and those assumptions are yours, not your customer's.

Incrementality testing is the closest thing to ground truth. Geo-holdouts, lift studies, conversion lift experiments. You measure what actually changes when a channel runs versus when it doesn't. This is directionally honest. It's also expensive to run continuously across every channel.

The conclusion is not to pick the best model and trust it. The conclusion is to use multiple models, understand what each one over- and under-counts, and triangulate. This is especially true when you're trying to understand how CAC economics actually work across different stages of growth.

Platform-reported numbers are not your numbers

Every ad platform reports conversions in its own favor. Facebook counts a view-through as a conversion. Google counts an assisted click. Both platforms count the same conversion. You see 180 conversions in your dashboards. Your CRM recorded 100 sales.

This is not fraud. It's how attribution works when each platform only sees its own data. But if you're making budget decisions off platform-reported ROAS, you're optimizing toward numbers that don't reflect reality.

The baseline adjustment: discount platform-reported conversions 20-40% as a starting point.[^1] Not because the platforms are lying, but because overlapping attribution windows inflate reported numbers structurally.

Run a geo-holdout test to calibrate. Split your geography into exposed and held-out regions. Pause a channel in the holdout. Measure the difference in actual conversions - orders in your backend, not impressions or platform events. That delta is your real incrementality. It's usually lower than the platform claims. Sometimes significantly lower.

Run this quarterly on your top 2-3 channels. When two measurement approaches disagree by more than 30%, that's not a data problem. That's a signal to run an incrementality test before you touch the budget. The channel you're measuring may also behave differently depending on which growth game you're playing.

The infrastructure question matters more than the model question

Here's what most attribution debates miss: your model is only as good as your data. And your data is only as good as your instrumentation.

Teams that can't reliably track the full funnel - impression to click to signup to activation to revenue - are arguing about models built on incomplete inputs. The model choice is almost irrelevant if the event data has gaps, naming inconsistencies, or drops in the middle of the funnel.

Good measurement infrastructure has three components.

Event tracking. Every meaningful action a user takes, captured consistently. The failure mode is inconsistent taxonomy: sign_up in one place, signup in another, user_registered in a third. That inconsistency creates debt. Queries break. Funnels show false drop-offs. Teams lose confidence in the data and stop using it.

Build a clear event taxonomy before you scale your tracking. Name events consistently. Document them. Treat this like a schema - changes need review, not just a quick fix.

Identity resolution. A user clicks an ad on mobile, browses anonymously, signs up on desktop three days later. Without identity resolution, that's three separate users in your data. With it, it's one journey. Multi-touch attribution only works when you can stitch sessions to a single person. Server-side identity resolution, logged-in state, and first-party cookies are the tools here. Third-party cookies are gone. Build without them.

Data pipelines. Clean data flowing from your product into the tools that need it. Server-side conversion APIs (Meta CAPI, Google's offline conversions) feed your ad platforms first-party data instead of relying on browser pixels that get blocked by iOS, ad blockers, and privacy settings. This matters more than it did two years ago and will matter more again in two years.

First-party data is the compounding asset

Post-iOS14, post-cookie, the teams with rich first-party data have a structural advantage. Not a temporary one.

The flywheel is straightforward. More users generate more behavioral data. Better data improves acquisition targeting. Better targeting brings in more users. The loop compounds. But it only starts compounding when you instrument it properly and accumulate enough signal.

Six months of behavioral data is exponentially more useful than one month. Not twice as useful. The signal density increases enough to power models and audiences you can't build at lower volume. That's a reason to start the infrastructure investment now, even if your data volume is still modest.

The privacy angle matters here too. Collect what you need, be transparent about it, give users control. That's not just regulatory hygiene. It builds the trust that makes people willing to share the behavioral data that makes your models better. The teams that treat first-party data as something to extract will run into limits the teams that earn it won't.

Build the measurement layer before you optimize

The most common failure in growth teams is shipping without instrumentation. A new feature goes live. Traffic flows through it. Three months later, someone asks what impact it had. Nobody knows. The data isn't there.

The discipline is simple: instrument first, optimize second. Before you build the feature, define the events you'll track to measure its impact. Before you run the campaign, confirm your conversion tracking is clean. Before you trust a ROAS number, run the holdout.

This slows you down slightly at the start. It speeds you up enormously over time. Teams that can measure reliably run more experiments, make better decisions faster, and build compounding advantages. Teams that can't measure keep resetting. Measurement infrastructure also enables segmentation that actually moves revenue - without it, you're optimizing for the average customer who may not exist.

Accepting the uncertainty

Here's the practical framing: the cost of imperfect attribution is real but bounded. The cost of no measurement is unbounded.

Imperfect attribution means you'll sometimes over-invest in a channel that's getting credit it doesn't deserve. You'll sometimes under-invest in one that's harder to measure. You'll run geo-holdouts that are hard to interpret cleanly. That's the job.

No measurement means you're flying blind. Budget decisions become political. The channels with the best advocates win the budget, not the channels with the best returns. You can't learn from what you can't see.

Build the infrastructure. Run the calibration tests. Use multiple models and triangulate. Accept that none of them are right - they're approximations of a complex, multi-touch reality that no single number can capture.

The teams that build durable measurement systems now are building a compounding asset. The models will improve. The data will accumulate. The decisions will get sharper.

You don't need certainty. You need a system that gets less wrong over time.

[^1]: Measured (formerly Recast), "Incrementality Testing", 2023. Platform-reported ROAS is overstated by 20-50% on average across major ad platforms. Corroborated by Meta Conversion Lift Studies, which show 20-30% inflation in self-reported conversions (2022).