Signal Engineering 101

Signal Engineering 101

Gabe Kwakyi, was CEO of Lingvano, the #1 sign language app, and is now founder of a new app, Bonsai Burnout Coach. Gabe is a mobile growth OG, founding the agency Incipia whose clients included Walmart, Coinbase, Peloton, Canva, Bird Scooters, PeopleFun, Duolingo and more.

A BMW M3 is not only a beautiful, high-performance sports car – it’s also a fun metaphor that can help you understand the basic implications of signal engineering for modern, algorithmic advertising campaigns.

So what do signal engineering of Meta/Google/TikTok ads have to do with a sports car?

To achieve its vision of being “the ultimate driving machine,” BMW M3s feature a turbocharged engine, which takes high-quality, premium gas.

What would happen if you put cheap, low quality gas into a BMW M3?

You could probably drive it – but would never reach its true potential.

What would happen if you put uber-high quality, F1 fuel into a BMW M3?

It might start and even drive around the block – but that fuel type might actually cause it to break down.

Just like how the BMW M3 performs best in the “goldilocks” zone of high enough fuel quality, your Meta, TikTok, and Google ad campaigns perform best in a goldilocks zone in terms of the data quality you fuel them with.

“Signal engineering” is a method of enriching the quality of data you post-back to your algorithmic ad campaigns from your product, so that your campaigns can perform at their best, in CAC/ROAS/ROI terms.

Upper-funnel data = poor quality fuel

These are the activities done by all or so many users (e.g. installs, signups, and even some first key actions). That means they have little-to-no correlation with good campaign performance, because they do not signal that the user acquired was high-value.

Lower-funnel data = too highly enriched fuel

These are the activities that are done by only the most valuable users (e.g. subscriptions, extremely high AOV purchases, 7+ day retention). While these do correlate very well with good campaign performance, they are not at all well-formulated for ad algorithm engines, because they either happen too infrequently or take too long after the user installs before they post-back.

Lower-middle funnel data = goldilocks zone fuel

These are the activities that are done by just enough users (e.g. trial start, or happy path completed within 15 minutes of app open). Not only are these signals not completed by low-value users – they also happen frequently and early enough after install, which makes them well-formulated for ad algorithm engines.

Note: “early enough” in most cases means within 24 hours of install, because modern ad algorithms don’t like to wait longer than that for a reward signal.

Here are 2 examples of signal engineering used in apps:
‍

• Many financial apps have found that, rather than optimizing for all signups or trials, it is better to send a "qualified" signal, based on what is available in the user’s profile as well as their behavior early from their app usage. For example, signing up and then making a first deposit or transaction, rather than simply signing up is one way to signal a qualified signup.
  ‍

• When managing marketing for the casual word puzzle game, Wordscapes, we found that what worked best was creating a combination signal, which only posted-back when a user 1) used 4 in-game hints and 2) completed 20 game levels. Pouring this fuel into Google app campaigns enabled us to scale dramatically more than simply optimizing towards a single level completed, or even a purchase event.

If your app is subscription-monetized with a free trial, try starting with a qualified trial event, which only fires when the user does NOT cancel their free trial within the first 1 hour of starting their trial.

To refine your qualified trial event, pull some data and fine-tune the following inputs of your qualified trial event:

• How long to wait before posting back the signal – if most users who cancel their trial within 10 minutes, then send the signal at 10 minutes instead of waiting a whole hour.
• Adding additional actions after the trial, which increase the confidence that the user will convert into a subscription.
• Adding onboarding signals that indicate the user is high-value, which you captured when asking them onboarding questions.

If your app is not subscription monetized or doesn’t have a free trial, try starting with a qualified signup event, which only fires when the user ALSO completes a key action within 1-24 hours after signing up.

Note: be sure not to immediately replace your primary trial or signup events with these new signals. Instead, create a new event and map it to an unused standard or custom event slot in your ad network, so that you can test the efficacy of your new signal, and don’t have to switch back if it performs worse than your primary event.

Lastly, keep in mind some of these challenging nuances of signal engineering:

1. Sending multiple signals.

‍When sending signals, modern ad algorithms generally prefer not to receive a range of signals, but rather one single, strong signal. This means, rather than sending 10 separate events, or an event with 10 parameters for 1, 2, 3, and up to 10 levels complete, it’s best to focus on a single level complete, of say 5 or 10.

2. Assuming organic trends will replicate under paid marketing.

While using regression models can be a great way to find good signals “in the lab,” be aware that these insights don’t always translate well “into the wild,” as correlation doesn’t always imply causation.

Also, trends that exist organically may not respond well to paid optimization. This means for instance, that a behavior that is correlated with 20% higher LTV in an analysis may lead to no discernible change once you factor in the paid algorithm incentives to maximize users likely to reach that behavior marker, but not necessarily to go further.

3. Using too narrow a signal.

Using very precise signals from regression or decision-tree models can produce powerful predictors (e.g. predict subscription conversion at 30% vs the average of 3%), and still be entirely the wrong formula for ad algorithms (e.g. only .2% of all users match that behavior combination).

While it may be true that females ages 35-54 in Santa Barbara county on iPhone 17 Pro Max convert to subscription far more than other users, trying to target an ad campaign to that segment is guaranteed to fail.

If your interest is piqued to dive deeper into this topic, you can continue learning from two of the best experts in the growth marketing space:

• LinkedIn post and podcast episode by Eric Seufert
• Slide deck by Thomas Petit

‍