Earlier this month, Acquisio’s Co-founder Marc Poirier and I, Brad Geddes of Adalysis, got together for a fascinating discussion on the impact of artificial intelligence on paid search, the framework campaign managers need to leverage AI, and what goes into auditing this emerging technology.
Did you miss the webinar? Watch the recording here!
Marc and I received some great questions from the audience, but unfortunately we ran out of time before we could answer them all. But, as we promised in the webinar, you can check out our answers to all your questions below.
1 – How do you feel about Google’s machine learning? Have you ever set up an A/B experiment that put Acquisio against Google’s machine learning?
Brad: Google’s machine learning, when it comes to ads, is terrible. It doesn’t matter if it’s the creation of ads, the RSA ad serving, their optimize (which is horrendous) – not a big fan. When it comes to similar lists, I think it’s amazing. But my tolerance level is much different. For similar lists, if it gets me 10% new clients I never knew about, I’m really happy about it. I would never blanket statement that they’re good or bad. Google is very good at math, they’re very bad at creativity. I segment them a little bit in those areas.
Marc: In terms of the comparison, we do it all the time. We’re trying to get more disciplined around it and get a large sample of customers who use CPA bidding, for example. We don’t have an official research piece on that – yet! But Google solves a different problem. We’re trying to spend your money (at the optimal price of a click or conversion, instead of aiming for a set price) but not go over budget while maximizing conversions.
2 – If you have a long sales cycle, say 3-6 months, how is it best to use that data, combined with the leads that are coming in day-to-day? Should you only be feeding back in the truly converted leads if the data is months behind the current bids?
Brad: This is a philosophical question as much as a data question. I try to qualify leads as soon as possible. My MQL – my marketing qualified leads – should be done within 2-7 days at most. Machine learning, working off a 7 day lag, no problem. If we can sales qualify them within 7-14 days, I’d rather work on that data. Would I work on a closed sale 6 months later? No, I wouldn’t.
In cases of data delay, you should be using a two-pronged approach to machine learning. The first is to give the machine recent data to work with in making decisions. This can be leads (marketing qualified if possible) or signals of quality visits. This helps in the day to day bid management.
Then you should feed back actual sales data and see how the sales data lines up with the qualified lead data. If it lines up closely, then you can rely on short term signals. If it does not; then you want to dig into the data to see why it does not line up. There may be keywords, locations, times of day, or other signals that lead to poorer quality leads. In these cases, you want to make adjustments to your campaigns so your short term and long term data align with each other.
Marc: We do a lot of lead gen stuff for ourselves at Acquisio, and we have that problem. People will consider the product for some time before they purchase and it can be long. Agencies are more likely than advertisers to use our product and we know the number of accounts they manage matters. If they have more accounts under management, they are more likely to use our product. We have a good understanding of our market and where we’re going to be successful, so the forms we have for demo requests and so on capture that information for us. Do we use it for bidding? We don’t. The algorithm is not designed to capture this (it could, but it doesn’t right now).
Brad: For all lead gen companies – what data do you let the machines view versus what do you use internally? You would want to know ‘here’s our total leads and here’s our qualified leads’, so you can look at what percentage is qualified and get that number higher, but then you might only push one of those numbers back to your accounts for machines to work from.
3 – How do you consider seasonality without using long lookback window / historical data?
Marc: Acquisio Turing observes short to medium trends in the data and makes decisions based on those trends. Our algorithms only make decisions on current data, reacting to seasonal changes in the auction as they happen to avoid over or under spending. Check out our interview with Acquisio’s other Co-founder Richard Couture and Jason MacDonald, about managing PPC seasonality.
4 – With Google being one of the most advanced players in the AI and ML space, what is to keep them from making the agency model almost obsolete? If they can create a truly efficient self-serve platform that spends the client’s money so efficiently… where do agencies fit in the future?
Brad: When we compare what humans vs computers are good at, we see a few big trends. The first is strategy. A machine has no idea what your company wants to accomplish. They don’t have the data for how users buy from you, what awareness is worth, and how to grow your business. This firmly sits in the world of the agency and in-house marketer.
So far, computers have failed dramatically at anything creative. This spans from what the ads look like, offers, and website content. When it comes to the ad campaigns, from strategy to creation to execution, this firmly sits in the agency and in-house marketing world.
Data insights to drive your marketing efforts forward come from human interpretation. Machines can automate reports and show data trends; but they don’t know why those trends are occurring. Data interpretation, data storytelling, and data insights should be part of the human world for a long time.
When it comes to bidding, automated reports, and doing repeatable work – then machine learning is fantastic. When it comes to strategy, creativity, storytelling, and why something is happening and how to react to it – that’s where humans sit in the marketing ecosystem.
5 – Are you mainly considering cost/revenue to optimize Google Shopping bids?
Marc: Acquisio Turing can optimize shopping campaigns for the same goals as other campaigns, either CPC or CPA. Last year around this time, we did a webinar with Seer Interactive with tips on how to optimize your last minute Google Shopping campaigns.
6 – How do you handle a low traffic problem with your ML solution?
Marc: When faced with low traffic, Acquisio Turing uses a combination of adaptively adjusting in order to gather enough information for the task at hand, as well as pulling from a pool of low traffic data to help it make an informed decision.
Interestingly, our algorithms perform very well with small budgets, so lower-traffic paid search marketers shouldn’t be afraid to give our machine learning a try.
7 – As far as I know, Google Ads is not sharing the data real-time, you can only get daily data. So how is it possible to run bid optimization real-time?
Marc: We leverage the Google API, which luckily does have real-time data available. Acquisio Turing digests that data, learns from it, and optimizes your bids every 30 minutes – which in turn can lead to increased clicks and conversions.
8 – How many conversions do you need to optimize for conversions?
Marc: In order to properly optimize for conversions, it’s recommended to have at least one conversion per day during the last 30 days. But of course, it would be better if the campaign was yielding 5-10 conversions per day, as this would give the algorithms more data to work with.
Good to know: conversions also need to be tracked with Google Ads conversion pixel, and only one conversion should be included in the conversion column (with conversion tracking set to unique versus every).
9 – Does this work on display ads as well?
Brad: Yes. Data is data. If your display ads are driving impressions and conversions; then you can automate this type of bidding and management as well.
Feature Image: Unsplash / Zach Lucero