Huge News - Facebook will no longer allow third-party data for targeting ads

Dave Lemmon · Mar 30, 2018

Alexander Lau said:
Nice work, you should float your own services here, very relevant!

We're that weird hybrid of software and agency services, so we poll the DMS and can run multiple campaigns. We can do loyalty, targeting purchase customers with buy back or OEM specials to existing customers who had been in 18+ months (or whatever time frame) or service specials (including one that has a scheduling app built right into the ad). We also do "conquest", building a look alike audience and doing either VIN specific or OEM specials to a broader audience (this is where the loss of the in market shopper data MAY hurt a little). What's cool is then doing daily polling starting 28 days out and every day after, matching back to both impressions (kind of fake number) and clicks (really strong number) and then the actual profit on those transactions. Gives us a cost per conversion that blows me away every time I look at it, even to the click to purchase data.

Dave Lemmon · Mar 30, 2018

We have to stay on our game, btw...I'm in the same city as Kerr and Steve and Matt. Can't slack when those guys are around.

jon.berna · Mar 30, 2018

Dave Lemmon said:
Looks like a cool platform (and very much like what we're doing). The machine learning/predictive layer sounds cool, but I'm assuming they're really just looking at time from purchase or time between repeat purchases of either sales or service...not that you don't need that.

The prediction input features are fundamentally based on the data you have available and how clean it is. We have all of the CRM, sales and service data and it is in a unified environment. This means we build models across all CRM/DMS providers for all dealers. This infrastructure allows us the ability to develop more accurate predictions. We also don't need to rely on 3rd parties. We evaluated and tested a number of them last year. We noticed the 1st party data was really good (especially with CRM) to get accurate predictions, it also was better for training the models. This is because everything is 1:1, with zero attribution that is reliant on hit or miss (mostly miss) identity resolution. With 3rd party data you also get a lot of noise and not much signal in terms of regression analysis.

The 3rd party data is not as great as everyone pumps it up to be either. You can check what one of them knows about you here. Among a number of large errors they are pretty sure I have an extra 2 kids, a car long since gone from 2011, and that I am a huge fan of NASCAR. Individual results may vary...

We started our journey in ML using Microsoft's ML studio back in 2016 to build our 6 month out sales forecasts. Early last year we went full custom Python and R. We have built tool after tool to enrich this layer of the process including our own out of bag testing and data training.

I will show anyone dealer/vendor the input features we use and our results. I am never afraid to show how we do things as all of this is available to everyone you just need to dive in.

Dave Lemmon · Mar 30, 2018

jon.berna said:
The 3rd party data is not as great as everyone pumps it up to be either. You can check what one of them knows about you here. Among a number of large errors they are pretty sure I have an extra 2 kids, a car long since gone from 2011, and that I am a huge fan of NASCAR. Individual results may vary...
.

On this we agree wholeheartedly, Jon. What they had on me was correct, both cars I owned, education, type of employment. They had my income substantially low and nothing at all under several categories.

jon.berna · Mar 30, 2018

Alexander Lau said:
I once had a discussion with a guy within an agency promoted post over at Automotive Digital Marketing and said someone would do just this one day (can't find thread, their search sucks). So smart!

I started giving this more serious thought in June/July of last year and we started working on it in Oct. It seemed obvious for a number of reasons, as the process to import is unbearably painful. The data prep alone is a huge barrier and massively impacts the match rates. It takes 72 hrs for a manual import to match, this is <30 mins when done through our app via hashed values. Ad costs to 1st party data is easily 50%-100% of the in-market data and typically every metric is better. Adding and removing daily improves the relevancy scores as everything stays fresh. Protect your owner base as well. Lookalike audiences on predicted / high CLTV base files are amazing.

Edited: Forgot to add CRM leads within the last 90 days in an active status.

I can say it is a really interesting project and my favorite so far. I feel like we are just scratching the surface.