Signal Engineering - How & when to start?

[Jack811]

Recently, I’ve been seeing a lot of articles about signal engineering, and I’m curious about how effective this approach actually is compared to BAU campaigns that rely on standard MMP subscription events.

From what I understand, there are two main approaches to signal engineering:

1. Changing the volume and frequency of signals
2. Enhancing signal quality through value AKA pLTV

However, I also see some potential risks:

• If you send too few signals or send them too late, campaigns may not have enough data to learn and optimize properly.
• When attaching value to events, accuracy becomes critical. If the value model isn’t reliable, the algorithm may optimize toward the wrong type of users, resulting in lower-quality traffic.

I’d love to hear from anyone who has hands-on experience with this:

• How effective have you found signal engineering compared to standard event-based optimization?
• What’s the best way to get started?
• What prerequisites (data, infra, modeling, etc.) are needed to make this work well?
• Which approach is suitable for Meta, TikTok, and Google Ad?

 

[davamosk]

Pretty complicated to answer this with an "unique" solution for everyone.

You have to adapt signal engineering to your product and user behavior so the signals you create will be only usable for you.

You are completely right with the risks that you mentioned, Signal Eng takes time until you nail the conversion and it's an ongoing process of A/B testing different signals against BAU events.

From my experience, I normally start with engagement signals, combining different events that are very correlated to the potential trigger of a purchase/subscription. I don't like starting directly with pLTV and purchase behavior is way more sensitive and the possibilities of failing are higher.

The requisites is to have a team/dev that is capable of creating these events for being fired specifically when you determine, either with the MMP SDK, the ad network SDK or if you go with S2S event so nothing extraordinary. You'll also need the tools to analyze the pre vs post performance with the events ofc but that's obvious.

I normally test these events with Google Ads for Android campaigns (using firebase) and with Meta for iOS (using either MMP SDK or Facebook SDK)

 

[WanaC]

The risk/reward balance is real, you pinpointed at 2 potential caveats that are important to acnowledge
Done right, I've seen tweaking conversion data give a substantial return improvement.
But I've also seen (and been part) of several cases where either it wasn't big enough for the effort, and/or we completely fucked it up.

day30, voyantis, churney are providers in the space you might want to check out.

 

[Jack811]

Pretty complicated to answer this with an "unique" solution for everyone.

You have to adapt signal engineering to your product and user behavior so the signals you create will be only usable for you.

You are completely right with the risks that you mentioned, Signal Eng takes time until you nail the conversion and it's an ongoing process of A/B testing different signals against BAU events.

From my experience, I normally start with engagement signals, combining different events that are very correlated to the potential trigger of a purchase/subscription. I don't like starting directly with pLTV and purchase behavior is way more sensitive and the possibilities of failing are higher.

The requisites is to have a team/dev that is capable of creating these events for being fired specifically when you determine, either with the MMP SDK, the ad network SDK or if you go with S2S event so nothing extraordinary. You'll also need the tools to analyze the pre vs post performance with the events ofc but that's obvious.

I normally test these events with Google Ads for Android campaigns (using firebase) and with Meta for iOS (using either MMP SDK or Facebook SDK)

Got it, thanks! I’ll see how I can make it work for me

 

[Jack811]

The risk/reward balance is real, you pinpointed at 2 potential caveats that are important to acnowledge
Done right, I've seen tweaking conversion data give a substantial return improvement.
But I've also seen (and been part) of several cases where either it wasn't big enough for the effort, and/or we completely fucked it up.

day30, voyantis, churney are providers in the space you might want to check out.

yeah, I’ve contacted churney before. To be fair, they charge a huge amount and require a long period for algorithm training and testing, which makes it difficult for most companies to try their service except for large players in the industry