Causal attribution
Causal attribution, measured on profit
Ad platforms grade their own homework. MTA only credits clicks. Causal factor attribution corrects both with an explicit multiplier — keep 0.6× of what Meta claims, or 0.9× of what MTA measured — grounded in incrementality tests, set per channel, funnel, and market.
Keep 60% of what Meta claims
Trusted by commercial teams at leading brands
The problem
Two sources, two different answers, no way to choose
Meta says it drove the sale. Your multi-touch attribution says organic search did. Picking one means accepting its bias. Averaging them means pretending the bias is the same everywhere — which it is not.
2 claims · 1 order
Both are describing the same order
Wrong for both
One correction factor hides more than it fixes
What MTA can’t see, it can’t credit
The untracked share keeps growing
Calibration multipliers
Keep 0.6× of what the platform says — you set it
For each channel you pick a base — the ad platform's reported value or your MTA's — and put a multiplier on it: keep 60% of what Meta claims, or 90% of what MTA measured. Below 1.0 pulls inflated channels down; above 1.0 lifts the impression channels click-tracking undervalues. You can see every multiplier, change it, and explain it to your CFO — and totals always conserve, so the split changes, never the sum.
Cost per acquisition, calibrated
Multiplier = the share of claimed conversions that are real
Branded search looked like the cheapest channel on the account. A ×0.15 multiplier means most of those conversions would have happened anyway — so an incremental customer really costs €147, not €22. That’s the whole feature. It is this easy to read.
Why it's different
Profit as the outcome, experiments as the evidence
01
Two channels, same revenue — only one made money
The correction doesn't stop at revenue: Dema applies it to contribution margin and new customers too. Two channels can drive identical incremental revenue while one sells full-price products customers keep and the other sells discounted products that come back. On corrected revenue they look equal. On corrected profit, one gets scaled and the other gets cut.
Same corrected revenue — opposite decisions. TikTok sells discounted products that come back.
02
Where does ×0.71 come from? Experiments.
You don't need to run a single test to start: every channel gets a multiplier from Dema's benchmark, built from finalized geo experiments across all our customers. Then every test you run tightens the estimate around your own business — after three Meta tests, that ×0.71 is measured on your customers, not an industry average.
No tests yet? You start on the benchmark. Every test tightens the estimate around your business.
Channel level
Meta, Google, TikTok and affiliate do not over-report by the same amount. Each gets its own multiplier.
Funnel level
Prospecting and retargeting deserve very different multipliers. Branded search usually deserves the harshest.
Market level
Reporting bias varies by country, driven by tracking coverage, consent rates and channel mix — so multipliers do too.
The stack
Where causal attribution sits in the stack
Ad platform data and multi-touch attribution are inputs. Incrementality testing is the evidence. Causal factor attribution is what turns that evidence into a corrected number you can act on daily, and marketing mix modeling is what covers the whole mix continuously. Causal attribution and MMM are the destination — the raw platform numbers are just where you start.
Trusted by category leaders
Real teams, real numbers — from cutting acquisition costs to doubling profit to automating whole planning cycles.
They tested scaling Google spend — and incremental profit went up with it.
Axel Arigato is a contemporary fashion and sneaker brand born at the intersection of streetwear, music, and culture. As growth accelerated, the team wanted to ensure that scaling their digital investments would translate into sustainable, profitable growth, not just higher ad spend. That's when they turned to Dema.
>100% Incremental epROAS on Google scalingEnterprise-ready
Built for production.
Where your data is processed, how it is protected, and what an agent is allowed to change are controls you set — not defaults you inherit.
EU & US inference
Choose where your data is processed. Model inference and data pipelines run on European or US infrastructure with full residency control.
Never used to train models
Your data is never used to train models. Processing stays in your chosen region and runs in isolated, encrypted sessions.
Always the latest models
New reasoning models roll out as they ship, so your agents keep improving automatically without migration work.
Permissions and approvals
Each agent has its own scope and integrations. Actions can require a named approval before anything changes.
What you can create
Real setups, linked to how they run.
Every example below is a real prompt from the library. Open the use case behind it to see the full run — what it checks, what comes back, and where it lands.
Questions
Frequently asked questions
Causal attribution corrects conventional attribution using evidence about what marketing actually caused, rather than what was merely recorded. In Dema this is implemented as causal factor attribution: each channel gets an explicit calibration multiplier applied to its attributed value — the MTA figure or the ad platform's own claim — and those multipliers are derived from incrementality experiments instead of being assumed.
Yes, and that is the point. Most causal measurement reports incremental conversions or incremental revenue. Dema applies the correction to contribution margin and to new customers as well, because two channels with the same incremental revenue can have very different margins once cost of goods, fulfilment and returns are accounted for. Optimising incremental revenue can quietly reduce profit.
Multi-touch attribution divides credit between the touchpoints it recorded, so it inherits every gap and bias in that recording. Causal attribution takes a recorded number as its base — you choose the MTA value or the ad platform's reported value per channel — and applies a multiplier measured through experiments: keep 60% of what Meta claims, or 90% of what MTA saw. MTA's relative split can stay useful while its level gets corrected.
From geo-based incrementality tests on your own business wherever you have run them. Before that, from Dema's platform-wide benchmark — a distribution of incremental factors built from finalized experiments across all customers — so every channel starts with a defensible multiplier rather than a placeholder. A Bayesian model combines the benchmark with your own results: each test narrows the distribution and shifts it toward your business, with statistically stronger tests carrying more weight.
Yes, and they should. Multipliers are set at channel, funnel-stage and market level, because reporting bias is not uniform: branded search typically claims credit for demand that already existed, prospecting rarely does, and tracking coverage varies country to country. A single global correction factor hides more than it fixes — you steer each one.
No. Totals always conserve. When one channel's attributed share goes up, another's goes down — the split changes, the total does not. That property is what makes the corrected numbers safe to budget against.
The calibration depends on whatever attribution inputs you have, and those may be cookie-dependent. The evidence behind the multipliers is not: incrementality testing and marketing mix modeling both run on aggregate data. As tracking coverage falls, the sensible move is to lean further on MMM and incrementality, which causal attribution is designed to accommodate.
Bring one commercial job