How to Train Google's Bidding Algorithm

What's the best Google Ads account management structure? How can you build a quality audience if the business does not have any lookalike audiences to leverage? What does Google's change to broad match keywords mean for your campaign structure? These questions answered, and more.

‍

Full Q+A Transcript Below

Patrick Gilbert:

Hello everyone, and welcome to another Q and A for Isaac Rudansky’s Digital Advertising Superstars Facebook group. I’m Patrick Gilbert, joined as always by Danielle Immerman. And this week, we brought on Esti Nadoff, one of our senior account managers here at AdVenture. 

Isaac is out of the office today, but we wanted to bring an Esti to get some perspective on some topics. Esti is an excellent resource for a lot of the things that we talk about. She's one of the people that I go to very often to flesh out ideas and work through concepts. She's a huge help for a lot of the content that we've put out over the years. To plug the book, Join or Die: Digital Advertising in the Age of Automation is now available on Amazon as a hardcover, paperback or Kindle. It's on sale right now on Amazon. Check it out there. The reviews have been excellent. The feedback has been excellent. I really appreciate all of the kind words that everyone has reached out to give on Amazon, Twitter and LinkedIn. It really does mean a lot and as part of it, people have been asking questions that we're going to be answering in this session today. Just to kick things off, Esti, how are you? How are things going today?‍

Esther Nadoff:‍

Good. I'm really excited to be here and join this group, so thank you for having me.‍

Patrick Gilbert:‍

It's an absolute pleasure. I wanted to kick things off, before we dive into any of the specific questions that we've gathered from the group, with a topic that we've been discussing a lot internally about Facebook account structure. This is a big topic that I think a lot of people have questions about, and there's a lot of changes that are taking place with Facebook, both in strategy and on the backend, but I'm curious as far as how this changes anything that you're looking at. What is your best practice mentality? How do you approach Facebook account structure and setting it up for not just profitability, but how to scale accounts and when we bring them on board?‍

Esther Nadoff:‍

That's a really great question. Yeah. We've been talking a lot internally with our teams about different structures within Facebook and what we've seen with best practices, and really it's a game-changer right? Set it up for scalability. So what does that take? 

It's really looking at the short-term goals and long-term goals and trying to create that playbook for your strategy to set yourself up for success. What we really like to do, and what I really like to do specifically, is take a deep dive into the highest quality audience and really starting from there and building the strategy out. From there, once we max out that high-quality audience, it's about expansion and different opportunities, but really trying to hone in based on what you know about the specific customer, what you know about the highest quality customer and really trying to hone in and expand it.‍

Patrick Gilbert:‍

How do you determine the best quality audience?‍

Esther Nadoff:‍

I think it's definitely unique for every company and every industry. So let's say, for example, it's e-commerce. You can take your highest value purchasers - people that have purchased the top 25% of the purchase value and that may be the highest quality audience for you. If you're trying to really scale up to bigger value purchases, that would be a high-quality audience. We have a client that has entrance points from multiple different areas and we know that we're trying to get people to enter for a specific advertising award. With this particular client, they have huge customer lists from years and years of running these awards. We took lists from the last couple of years because we feel like that is the highest quality and we’re trying to create those lookalike audiences based on the highest quality audience. But really trying to hone in on recent trends for audiences, as well as taking not just thousands of customers and trying to expand from there, but taking that highest quality. So again, like the advertising awards, people specifically engage with the brand or company, and from there, expand that list.‍

Patrick Gilbert:‍

I think that's a great point. And I don't know if you specifically mentioned this, but you're referring to creating lookalike audiences based on that. And I think that's really interesting. That's something we talk a lot about with clients, is how can we take this audience data and really trim the fat, because that's the idea of training your algorithm to go and figure out who another similar audience is. 

Let's take a second to think about the data that we're feeding into the algorithm first. And I think what you're describing is extremely helpful and not thought about often enough. The advertising awards client is a great example. You know, they've been around for a long time. They have a lot of years and years and years of data, but maybe data from eight years ago isn't as relevant.

There's a similar idea with some e-commerce clients if you see a change in your customer based on seasonality. So let's say for example, that you get a different type of customer during the holiday season than you generally see throughout the rest of the year. I like to kind of step back and say, maybe we should pull a list of customers that bought between, let's say November 1st and December 15th of last year, and use that audience to create a lookalike audience. And that's how we fuel our Black Friday and Cyber Monday type initiatives, as opposed to using all of our customers. And the same for like, let's say you’re a flower delivery business or anything in the gift space: instead of just using all customers, you can pull the data out specifically from your Valentine's Day customers from last year and they're probably different than your Mother's Day customers.

And I think being able to understand that nuance is extremely important. That's a great point. Are there any other kinds of things that you look at or tips that you have for just like organizing? And I like this idea, right? So start with what you know is quality. And this is all assuming that you have historical data. If you have a brand new startup with no previous customer history, is there something that you would look at first, as far as, okay, this idea that we know that works is leveraging the best quality audiences? So what's the first thing that you would then do. Like, let's assume that there are no remarketing customer lists that we can operate on. Where are you going to start?‍

Esther Nadoff:‍

I'm going to start with as much research as possible on the client and the customer base to then determine those core correlating points of a customer that you can try to create those interest-based audiences off of and really try to qualify it that way.‍

Patrick Gilbert:‍

Totally agree. Excellent. Any other thoughts on Facebook account structure - any sort of tips or best practices that you find helpful?‍

Esther Nadoff:‍

There's a lot. So one thing that I do find is when I inherit an account that's super complex, there's a lot going on and it's just like, I need to tackle and get down to what was working, what wasn't I go through Excel spreadsheets of previous client accounts and campaigns to really try to understand what the goal was. I write down what their goal was for that account, for that campaign, what they were running, how long it was running for, what type of creatives, etc. I think that's really helpful in understanding and getting a better idea of their marketing in the past.‍

Patrick Gilbert:‍

So you're actually trying to learn from your past experience. That's a pretty crazy concept.‍

Esther Nadoff:‍

Exactly. I mean, I think it's a tip and advice because not many people do it as much as you think they would.‍

Patrick Gilbert:‍

You're totally right. So we were having, I don't know if you had a chance to read this message yet that I sent in Slack a little bit earlier, but we were discussing the idea of, like, when you have a contrarian thought, an aspect of your strategy that goes against what you would feel to be a best practice and how do you really handle that? And I think the first thing is to really acknowledge when that happens and then err on the side of being willing to test whatever it might be. And then really building in that feedback mechanism to be able to go back and say, okay, well, this worked or didn't work. And here's why it played out this way, and we'll create a lot of those things. So that maybe six months from now we'll have a similar situation and we'll be able to go about it in a more intelligent way. And this example that you're providing certainly hits that on the head.‍

Esther Nadoff:‍

Yeah. I did get a chance to read your Slack message and yeah, I was reading through some of the thoughts on Facebook and I did have that initial thought like, no, this is contrary to what we typically believe is a best practice, but it's true. It's like, no, you should be testing these things and really understanding if it works well for different things.‍

Patrick Gilbert:‍

Definitely. Awesome. Okay. So the first question that we need to get to came from a reader of Join, or Die who reached out to me on LinkedIn. Now there's a full chapter dedicated to the concept of training an algorithm in the book, but it's more conceptual, and his questions were very specific. And I really like this because what I tried to do with the book was to create a framework of topics and questions and ways that we should be thinking about things that could then lead us to exactly these types of questions.

I think I did send it to Danielle if you want to put this up on the screen. Let's say we have two examples of Google Ads accounts. Example 1: Ecomm store. Dozens if not hundreds of conversions a week. Big search volume, lots of opportunities. This seems to be a no brainer for tCPA or maybe tROAS. CPAs are mostly on the cheap side. How do you train the algorithm to work better and get better over time? (Let us say the goal is to reduce the CPA as time goes by)

Example 2: Service business. I work with a lot of roofers: High CPCs, small search volume in a small radius around their business. The KWs are pretty straightforward. And on top of that, they usually go for a budget of around $1000-2000. How do you train an algorithm to get better if the account has about 10-20 conversions a month? I usually can't even get an algorithm to work in such a cramped account.‍

Esther Nadoff:‍

My initial first thoughts are that I think it's really a matter of making efficient changes over time. Not too much at once. But for, let's say it's the CPA goal, it would be decreasing it, or if it's the return on ad spend goal, it would be increasing it slightly while adjusting budgets to match your goals. I think these concepts in tandem, adjusting budgets and adjusting goals to increase performance over time would really help.‍

Patrick Gilbert:‍

I agree. I think with that though, you need to be willing to take risks. So I think there's a balance that we often don't vocalize. Hey, listen, we're trying to make incremental changes to this campaign because it's profitable. And I'm in, based on the way that you've outlined this, it seems like this campaign is probably chugging along and is profitable. You'd like to get incremental improvement generally. Like our take there is like don't turn everything on its head, try and make some subtle improvements, but that doesn't mean that you can't take 15% of your budget or 20% of your budget and do something completely crazy with it. I would definitely encourage that. Now, if you're an agency and you're managing this on behalf of your client, I would definitely be clear about your expectations. The fact that this is a test and based on what we had just spoken about, you really want to make sure that you have the proper feedback mechanism in place to circle back and see, okay, did this work, did it not work?

And why? Um, one thing I would recommend is doing some sort of like crazy experiment, very high risk, we're going to take our products and we're going to put them in a campaign and we're going to put on maximize conversion value. And we're just going to blow the doors off this thing to see what happens. What's good about that is you're essentially forcing the system back into the learning phase. And I think that's really where the major changes are unlocked. And to really simplify this, there are all these signals that, let's say, Google uses in the auction to predict the conversion rate for a given option.

And after that initial test, let's say that they find that, I don't know, there's like 15 different signals that when all the dots connect and the stars align with these 15 signals, that we think that there's a show on the likelihood of conversion. So we're going to enter you with an aggressive bid. Great. And if there are 13 of the 15 signals that meet the checkboxes, we'll have an adjusted bid. That's slightly lower than if all 15 were met. Now, that's essentially how a learning algorithm works, where it finds the trends, the data points that it can essentially rely on to replicate at scale, where it can reasonably predict results. And once it reaches that point, it's going to just really focus on say those 15 signals, and match up as many options with as much as your budget as possible. Those 15 signals, for all you know, three to six months down the road, something's going to change with the marketplace where there are five other signals that the algorithm had not previously really looked at scale that could potentially help you become even more profitable to predict better outcomes to reach more people in a more effective way.

But because you've kind of let it just sit over here and optimize with these 15 signals, it hasn't been willing to dedicate enough time and to learn to explore new opportunities. Frederick Vallaeys gives a reference in his book of like, “Hey, maybe snowfall accumulation is like a signal that could matter for our business.” And maybe it matters, maybe it doesn’t, maybe it's something that should be considering the option. Maybe it matters for a ski resort, but not a software as a service business. So like all these things that might not come up, all of a sudden can be entered into the system and learned in a new way, if you give the algorithm a little bit of freedom. So I would say in this case, if there is a lot of volume, you have to realize there are so many different trends that could be unlocked if you just let the algorithm run wild. But at the same time to Esti’s point, you really don't want to like completely turn things on their head because it might not work. And you don't want to completely turn off a revenue-generating engine in hopes of finding something better. So you need to balance the two of those things.‍

Esther Nadoff:‍

Yeah, I think those are really interesting points, Patrick, and really great ones. And I think we also talk a lot about liquidity when it comes to all of this, look at you, look at that smile. This is Patrick's baby. When it comes to training the algorithm, there's so much data and so much information across the account that we want to make sure it's leveraging as much as possible within individual campaigns. So allowing the algorithm more freedom is a really great point. And adding that liquidity throughout the campaigns and throughout the account can help train the algorithm even more.‍

Patrick Gilbert:‍

Exactly. So even untouched over time, you should see some sort of incremental improvement, theoretically. It doesn't always happen, but it should be able to learn some things along the way. You can set aside some test budget to nudge it a little bit further in the right direction, but you know, all the disclaimers there, not everything is always going to work the way we would love it to. So the second question on this was the second example for a service business. So the first question was for an e-commerce business, which is a little more straightforward. E-commerce is much easier to track because you have revenue exchanged at some point online. Hopefully, right after the click happens with lead gen, there's a sales component. There's a sales team offline conversion tracking. There are other variables that make this ecosystem more complicated.

So this is a service business. I work with a lot of roofers. Okay. So the local services are very competitive, high CVCs, small search; the keywords are pretty straight forward, right? Roofer near me, keywords are pretty straight forward. And on top of that, they usually go for a budget of around $1,000 to $2,000 per month. So it's all low budget, high CPC, tight search volume for a competitive field - it's honestly one of the most challenging sorts of accounts that you can be dealing with. And then the question is, how do you train an algorithm to get better? If the account has about 10 to 20 conversions a month, I usually can't even get an algorithm to work in such a cramped account. It's tough. It's certainly tough. And honestly, there are times algorithms work great at scale, but you're operating in a much smaller, less scalable environment, so you'll need to do more traditional marketing and advertising. Let's let the algorithm do its thing. That's where we're going to get our benefit here. If you're not doing good marketing and good advertising and good salesmanship from the very beginning, then you're ultimately never going to be able to make this work. But as for this roofing example, what do you think are some tips?‍

Esther Nadoff:‍

Yeah, I have a couple of thoughts. I think when you're saying traditional marketing and that approach, I think that's a really good point. And it's about creating that brand awareness to increase that search funnel. The problem here right now is the overall volume, right? We need to get more data in order to get more data. We need more volume. So if we try to increase the brand awareness so that we can help increase the volume of people searching for our company or service, that can help with people when they're searching, they're familiar with our company, and then they'll click compared to all the other competitors. So that may help with overall volume. The second thought is I would consider qualifying different conversions with more micro conversions. Identifying what attributes across your different leads in different, um, quality conversions, whether does that mean they have been on the site, you know, they have searched the site for about two minutes or have they viewed an average amount of, I don't know, five pages on your site trying to create those micro-conversions, what we like to call them as micro-conversions, to help increase that volume can help train the algorithm first towards a higher quality audience and then help increase that overall volume as well.

My other thought is that what we have been very pro on is leaning towards more of a maximized clicks strategy to try to increase again. Since it's a volume problem, we really want to try to get as much volume as possible. So really maxing out the volume with the maximize click strategy, gather as much data as possible and then help train the algorithm with that additional data towards those higher-quality conversions as well.‍

Patrick Gilbert:‍

Totally agree. So we were historically not into the whole maximized clicks thing.I don't really think it even gets mentioned very much in the book because what you really want to be doing is a conversion-oriented strategy, like full stop. Please don't think that anything other than conversions is ultimately what matters, but there are times where your ultimate goal is to get to be able to lean on more conversion data, but it's just not available for you. So you have to figure out what other levers that you can pull in the meantime. And you have to be careful with how you do these things. But if you think about it, if you're on a limited budget and you look at a bottom-funnel search term, right? So if it's roofers near me, that's pretty high intent. And that's probably what everyone is searching for - the service that you offer.

It's not as if that can’t also mean someone’s looking for the name of a band and someone's Googling lyrics or tickets to a show for a band called roofers near me. Like, that's what I mean of like very, very specific bottom-funnel search. If that's the case, if there's very little additional signal data that you need to tell Google that this is somebody I wanted to bring to my website then theoretically, you could benefit from just saying, “Hey, I want as many of these people as possible, give me as many of these clicks,” which is going to reduce your CPC. It's going to increase your click-through rate, and you're going to get more of these people. You're going to be able to play that numbers game, but you also have to leverage the micro conversions.

You have to do everything else. Well, you have to have really good copy on your landing page. You have to be doing good marketing and advertising. None of this is going to work. If you look like you're a bad company, if you don't have social proof on your website, if you're not leveraging a mobile-friendly website, if there are spelling errors and other things that make sloppy mistakes, people are not going to hire these companies to come to their house. So all those traditional marketing things absolutely needed to be in play if you want to do any of this right. I think those are good answers. We have to wrap up in a minute, but there's this big thing in the PPC world right now about broad match modified. We don't really seem to care at AdVenture Media. Esti what do you think?‍

Esther Nadoff:‍

Yeah, no, it’s really not. I guess it's still there. It's just a different symbol.‍

Patrick Gilbert:‍

Right? Like we've been pro broad match keyword and just like automation. If you've embraced automation over the last year or two, three years, like this means literally nothing. And like, I find it so interesting to watch LinkedIn and Twitter and even like, look, this own Facebook group has a lot of people that are freaking out about the broad match, modified change. It really shouldn't change anything that you do. You should have already embraced the reality that these things are going to address the change and you should already be thinking about, “Okay, well, what is it going to look like when we don't have keywords? Okay, well, I have to really understand how bidding algorithms work. And I have to really understand how good landing page copy works.”

That's the way that we all need to start thinking - we should have been doing it for the last two years. And Google is making these changes that are forcing us further and further down that path; they’re ripping it off like a band-aid. We just need to be ahead of it. You need to predict the fact that we're ultimately going to lose all keywords. So what can you do to put yourself in a position to succeed if and when that actually happens - the broad match modified thing is a very, very small step in that direction. I don't think it's going to change literally anything we do in any accounts.‍

Esther Nadoff:‍

And I think another step to embracing all this change is reading Patrick's book because Join, or Die.‍

Patrick Gilbert:‍

Amazing. Join, or Die: Digital Advertising in the Age of Automation is available now on Amazon.