Approach to Linear TV Attribution

Last Updated: February 18, 2025 10:44

Purpose

This article describes our approach to linear attribution and outlines how the probabilistic spike model works vs. the linear impression-based model.

Introduction

Innovid’s unified approach to linear TV attribution combines different models to deliver unparalleled accuracy and insight into overall TV performance for advertisers.

Our linear spike model measures the immediate response to traditional broadcast linear campaigns and works for any advertiser regardless of market size or campaign spend. It uses linear ad occurrence data in conjunction with online response data.

Using household-level data, our impression-based model measures the delivery and impact across linear and streaming campaigns. It uses a range of impression-based datasets, including ad serving data from certified OTT/CTV publishers and platforms and online response data.

Our platform is a useful tool for marketers looking to improve their understanding of the impact of their advertising campaigns and optimize their media spend accordingly.

Probabilistic spike model

Our spike model is used for broadcast linear campaigns. It works by assigning each TV spot a weighting based on the probability it drove a viewer to engage with the advertiser’s website or app.

How spike model linear data is processed

We use a three-step process:

Collect data and track response
Website: Our first-party Response Tracker collects data from the advertiser’s website from visits through to onsite actions and sales.
App: The advertiser activates InnovidXP as a conversion endpoint through an app partner, and this integration brings data through to the platform.
Calculate and generate the baseline
Continuously measure website traffic.
Filter out traffic that is highly unlikely to have been driven by TV.
Evaluate the baseline for every market, every minute of every day.
No need for historical data.
Attribute and assign response
Isolate and attribute above the baseline activity.
Assign a probability score to each user session.
Deal with overlapping spots.

The diagram below outlines this process:

Spike model data is processed daily - modeling is run on all spots aired in the 24 hours since it last ran, and results are published in the platform. This processing deals with any overlapping spots, i.e., spots that are aired at the same time on different channels.

The example below outlines this process:

Spike Processing Example
Example weekday	Description	Always on
Before spot airs	Evaluate the baseline.	Continuously measure and filter website traffic.
Wednesday, June 1st	Spot airs.	Run every day.
Period in between spots	Calculate the lift of activity above the baseline after spot airs.	Deal with overlapping spots.
Wednesday, June 8th	Spot airs again.
After each spot airs	Attribute above the baseline activity. Assign a probability score to each user session.

Impression-based model

Our impression-based approach uses granular viewership and ad occurrence data sourced from smart TV data providers (Inscape or Samba), ACR, post logs, our ad server, or integrations with third-party OTT/CTV publishers and platforms; alongside response data from the advertiser website and/or app.

Once we have captured all the impression and response data, we cleanse, filter, and model it, mapping viewership to households and online responses. From there, data is put through numerous extrapolation processes to ensure it provides an accurate representation of the total TV viewing population. The final result is actionable reach, frequency, incremental reach, and outcome insights.

How impression-based linear data is processed

We use the following process:

Collect smart TV, ad airing, and website/app response data.
Cleanse and filter the data.
Process 1:1 matching.
Calculate TV Influence (i.e. we use a control group method to isolate TV uplift).
Upscale the results (because the Inscape/Samba datasets only cover a small, but representative. fraction of the total TV viewing population).

The diagram below outlines this process:

Impression-based linear data is processed weekly. All linear impression-based spots undergo processing twice since this method only processes the previous two weeks’ worth of data. After roughly two weeks, you can view provisional impression-based results. These are usually finalized around three weeks after spots have aired.