Modern marketing teams have one primary goal - maximizing the return of every marketing dollar spent. To achieve this, growth leaders need to understand the impact of their budgets so they can optimize the signals that power today’s algorithm-driven channels.

Traditionally, the way to measure impact has been A/B testing - analyzing how different versions of products, landing pages, or messaging affect conversion rates.

But not everything can be A/B tested. Sometimes it’s infeasible, sometimes ethically questionable, and often practically impossible.

Take paid ads, for example. Ideally, you’d want to know which signals will train platforms to drive valuable traffic before committing extensive budgets. But how do you identify high-quality signals beforehand?

This is where treatment effects come in. In this post, we’ll explore how they work - and why they’re a powerful way to generate signal inputs that improve platform performance without burning unnecessary spend.

The Blackbox Dilemma

Ten years ago, marketers still had direct levers: micro-segmentation, bid adjustments, and fine-grained targeting strategies to test and optimize reach.

Today, those levers have disappeared into the blackbox. With automated systems like PMax and Advantage+, audience selection and bidding are no longer in the marketer’s hands — they are controlled by platform algorithms we can’t see or steer directly.

The result: audience coverage, spend allocation, and even which customer segments are prioritized are all determined inside an opaque system. The only thing marketers are still in control of are the signals they feed into this blackbox - the data points that tell the algorithm who is valuable, who isn’t, and what actions matter most.

Causal Analytics to the Rescue!

The good news is: even in a world of automated targeting, marketers can still improve performance. The key is moving from observational data to signals of incrementality. So instead of relying on noisy behavioral signals, marketers can use causal analytics to generate signals that reflect true incremental value.

This is where observational studies come in. They estimate the causal effect of an action while accounting for pre-existing differences that might otherwise bias the outcome.

For example, suppose you want to know whether an ad impression actually increases the probability of a user completing checkout. By modeling the counterfactual - what would have happened had the user not seen the ad - observational methods allow us to estimate the incremental lift attributable to exposure.

In data science terms, this causal lift is known as the treatment effect. Treatment effects let us transform raw behavioral data into high-fidelity causal signals - signals that platforms can use to allocate budget toward actions and users that drive measurable incremental outcomes.

Average Treatment Effects

Treatment effects measure the average causal impact of a binary action (e.g., showing an ad vs. not showing it) on an outcome variable (e.g., revenue). The concept originates in medical research but translates directly to marketing.

Formally, we define:

• Y₁ᵢ: Potential revenue of user i if they see the ad.

• Y₀ᵢ: Potential revenue of user i if they do not see the ad.

• Dᵢ: A binary indicator (1 if the user sees the ad, 0 otherwise).

The observed revenue for user i is:

yᵢ = Y₀ᵢ + Dᵢ (Y₁ᵢ − Y₀ᵢ)

This formulation highlights the fundamental challenge: we only ever observe one of these outcomes per user.

To estimate the missing counterfactuals, we use methods like propensity score matching, regression adjustment, or machine learning–based uplift modeling.

The result isn’t just a measurement of past performance - it’s a generator of causal signals. These signals tell platforms which users are likely to show incremental lift when exposed to marketing, and which ones would convert anyway.

The Signal Optimization Framework

Once treatment effects are estimated, they can be used to generate three classes of signals that directly improve platform learning:

1. Value-Based Seed Signals

   Instead of seeding lookalikes with past customers, feed platforms seed lists of users with high predicted causal impact. This increases the chance that the platform discovers new customers instead of overfitting to historical buyers.

2. Negative Suppression Signals

   Just as important as positive signals are negative ones. By excluding users with near-zero or negative expected treatment effects, marketers reduce noise and protect budgets from wasted spend.

3. Dynamic Intent Signals

   Treatment effects can be updated continuously to reflect shifting intent - e.g., users at risk of churn or with rising incremental value. Feeding these as “next best action” signals allows platforms to adjust in near real time.

Getting Treatment Effects to Work (The Innkeepr Approach)

Reframing treatment effects as signal generators changes the marketer’s role. Instead of trying to outsmart platform algorithms with targeting rules, marketers focus on supplying the highest-quality causal signals possible.

This shift enables growth teams to:

• Increase new customer ratios by training algorithms with incrementality-based seeds.

• Reduce wasted spend through suppression of low-value signals.

• Improve retention and lifecycle marketing with intent-driven signals that trigger the right interventions at the right time.

In other words: in the age of automated targeting, the job of growth teams isn’t to define audiences - it’s to optimize the signals that shape how platforms learn.

What Comes Next

Causal analytics isn’t just a measurement tool - it’s the foundation for building smarter signals that guide today’s ad algorithms. In an automated media-buying world where audiences and bids are decided by black-box systems, the only real lever advertisers control are the signals they feed in. The stronger and more incremental those signals, the better the outcomes platforms can generate.