Demystifying Incrementality: How Causal Inference Stops Platform Double-Counting

The Flyer Handout Analogy

Picture a store owner hiring a promoter to hand out 10% discount coupons outside a busy boutique.

The promoter doesn't stand three blocks away trying to convince uninterested pedestrians to turn around. Instead, he positions himself directly in front of the shop entrance. Whenever someone grabs the door handle to walk inside, he shoves a coupon into their hand. At the end of the day, 90% of shoppers redeem the coupon, and the promoter demands a hefty commission for driving "90% of all sales."

In the digital marketing world, this is exactly what retargeting campaigns, branded search, and affiliate coupon sites do every second of the day. They step in front of high-intent users millimeters before the checkout line, drop a tracking cookie, and report astronomical 10x+ ROAS inside their dashboards.

The Mathematics of Causal Inference

In statistics, the golden definition of causal impact is governed by the potential outcomes framework (Rubin Causal Model). For any marketing intervention $W \in \{0, 1\}$ on customer outcome $Y$:

The fundamental problem of causal inference is that we can only observe one reality for any given customer at time $t$. If Sarah saw an ad and purchased ($Y(1)$), we cannot simultaneously see what Sarah would have done in a parallel universe where she never saw the ad ($Y(0)$). That parallel universe is the counterfactual.

The Famous eBay Branded Search Experiment

In 2015, economists at the National Bureau of Economic Research (NBER) partnered with eBay to test the true incrementality of their Google AdWords budget. For years, eBay spent tens of millions bidding on their own brand term ("eBay"). Standard Last-Click attribution showed ROAS exceeding 20x.

The researchers ran a bold experiment: they turned off branded search ads completely in 30% of US markets.

The result? Sales did not drop by even 0.1%. When the ad disappeared, consumers simply clicked the natural organic search result right beneath it. The incrementality was virtually 0%. eBay had been paying Google hundreds of millions of dollars to acquire users who had already decided to buy from them.

Platform Double-Counting: How $100 Turns Into $300

Consider a standard direct-to-consumer purchase path:

1. A customer views a TikTok video on Monday (view-through attribution window starts).
2. They click an Instagram ad on Wednesday (7-day click attribution window starts).
3. They search your brand on Google on Friday morning and click a search ad.
4. Right before buying, they search for a coupon and click a Rakuten affiliate link.

When the order clears, all four platforms take credit. If your average order value is $100, your aggregated platform dashboards report $400 in revenue from that single human transaction. When calculating blended ROAS using channel reports, marketing teams pat themselves on the back while the CFO wonders why bank balances aren't growing.

How Social Vriddhi Implements Incremental ROAS (iROAS)

Modern brands cannot make budget decisions based on platform ROAS. You must steer by incremental ROAS (iROAS):

Where $\kappa \in [0, 1]$ represents the causal discount factor. A channel with reported 6.0x ROAS but a 20% incrementality factor ($\kappa = 0.20$) has an actual iROAS of only 1.20x. Meanwhile, an upper-funnel channel with reported 1.8x ROAS and 90% incrementality ($\kappa = 0.90$) has a true iROAS of 1.62x!

Social Vriddhi MMM solves this automatically. Our platform ingests your raw revenue and spend streams, continuously computes baseline organic trends, and calibrates each channel's incrementality coefficient using randomized holdout experiments.