Article

Automation, Arrogance, Liquidity & Optionality

December 18, 2023

I hope you’re all having a wonderful start to the Winter holiday season – whether you celebrate Hanukkah, Kwanza, Diwali, Christmas or just enjoy a festive time of year. As 2023 slowly draws to a close, I wanted to share the principles I’m using to think about and manage ad accounts moving into 2024. This has certainly been a year of evolution across the advertising landscape, and I think we’re going to see even more of that in the coming year.

While we’re still in the early stages of automation across the ad ecosystem (yes, it feels like it’s been here forever, but the reality is that we’re ~5 years in, and only ~3 of them with legitimately good machine learning). From a (very) high-level perspective, ML uses sequential algorithms and predictive analytics – combined with reams of user, publisher, creative and advertiser data – to determine the expected value of a given impression for each eligible advertiser, along with whether (or not) that expected value is sufficiently high for the advertiser to bid on the placement.

Despite the fact that this technology is still very much in its infancy, there’s no denying that it is here to stay. This is an objectively good thing for the media ecosystem, for much the same reason as the rise of trading algorithms has been for the financial sector: it allows advertisers (and traders) to spend more of their time thinking strategically vs. lever pulling.

This is a paradigm shift in the media buying world – one which is used to teams (in-house, agency, hybrid, whatever) buying specific placements and adstocks based on predetermined audience profiles and targeting criteria (like keywords, interests, topics, channels, behaviors). It was a hands-on, incredibly manual process of number crunching, comparison and probabilistic guesstimation – but one that media buyers are intimately familiar with. After all, there’s a certain comfort in knowing what went into a given buying decision. And historically, when a campaign (digital, broadcast, CTV, linear, whatever) wasn’t working, we relied on our ability to pull specific levers – to select ad units, to identify specific channels or placements, to change our bids – in order to get things back on track.

ML-powered bidding algorithms have changed all that, and the vast majority of advertisers haven’t adjusted.

One of the major red flags I see across the digital advertising landscape are advertisers (in-house and agencies alike) spending way, way too much time micro-managing campaigns – tweaking, changing, honing, manipulating, whatever you want to call it. While the exact nature of that behavior varies from account to account, it usually includes never-ending creative rotation, constant bidding adjustments (I recently reviewed one account with ~5,100 bidding changes in a single month – that’s an average of 170 changes every single day for a month), perpetual audience rotation, way-too-soon targeting changes, new campaign builds and continual strategic pivots (among other things).

All of this harkens back to days gone by, when advertisers needed to be “in the arena” pulling levers in order for campaigns to perform. But those are days gone by, and I’m far more interested in what’s going to work today than what worked yesterday.

This behavior is arrogance masquerading as expertise. It is substituting activity for outcomes.

Perhaps the clearest way for me to articulate this is an exchange I had with the always-insightful Andrew Faris on Twitter last week:

All of those activities cited above are the concrete manifestations of a belief that you have a better understanding of the expected value of a given impression than a supercomputer.

And that belief is wrong.

The simple reality is that no one – not me, not you, not Terance Tao, not Magnus Carlsen, not Finlan Serral – is going to out-trade a machine, just like no Wall St stock picker is going to out-trade a trading algorithm. Not only do we humans lack the mental capacity to do so, we don’t have access to the same volume of data as Meta, Google, Amazon, Walmart and our view is limited to our ad account (or, for agencies, a client portfolio’s ad accounts). Less data + less visibility across an ad ecosystem + less compute capacity = a staggeringly poor capacity to make accurate predictions about the expected value of a given impression.

Now, I’m certainly not a fan of mindlessly trusting ad platforms (I’ve written extensively about that) – but I think the proper role for an advertiser is to set targets and guardrails for automation, not to fight against it. The sooner we (as advertisers) embrace this new, more strategic and more forward-looking role, the better off we (and our ad accounts) will be. It’s that simple.

ML-driven media buying is here to stay. Automation isn’t going away. So, let’s talk about approaches and solutions.

Successful advertising in this new reality is built on four core principles:

Insight:

As platforms actively remove and restrict traditional targeting levers (something we’ve seen with ASC, Demand Gen, Performance Max, etc.), the natural outcome is a shift in the optimization layer for campaigns, from the target lever layer to the data layer.

We’ve seen this before in sports with the advent of Sabermetrics – when every club has access to the same basic data (statistics) and the same-ish qualitative analysis (a very homogenous group of scouts), the advantage goes to the team(s) who can identify the value that those sources don’t capture – either by supplementing that data with other, outside sources and/or by combining various pieces of public data in novel ways. Bill James did this through Sabermetrics – combining public statistics in novel ways to surface hidden value. Jim Devellano started scouting Russian players based on their skating and raw skills, vs. production and physicality (and drafted Sergei Federov…which worked out quite well for Detroit).

In digital ad platforms today, we all have access to the same platform-level data (interests, behaviors, demographics, locations) – so there’s no advantage in using those. Platform data is table stakes.

Where there’s an advantage to be gained is in either (a) combining that data together in novel, interesting ways (like Custom Segments and Combined Segments in Google Ads), or (b) in sharing specific, actionable business data to platforms (as in Custom Audiences in Google + Meta, or Lookalikes built off Custom Audiences, or value rules, or offline conversion imports).

When you provide automated bidding with data it can’t get anywhere else, it can find value that it wouldn’t anyway else.

The data you share will take many forms, from customer data (good) to business data (goals + targets, nuances, etc.), but at the end of the day, the best data wins. The better your data, the more likely it is that you’ll win out in the long term.

Liquidity:

In finance and economics, liquidity refers to the ability to convert an asset into cash (or other readily-marketable security). In advertising, it has much the same meaning – the ability for core assets (budget + data) to flow freely throughout an account.

The more liquidity in an ad account, the more likely it is that you’ll be able to capitalize on market opportunities and maximize value capture.

In practical terms, that means:

Unconstrained by Budget: I wrote extensively about daily budgets killing ad account performance here. That point still stands. If your account (or a campaign in it) is constantly hitting your daily budget, then liquidity is low. That means you’re likely to end up in situations where there are impressions available that meet your targeting and expected return criteria, but which you can’t purchase due to lack of funds. To put a finer point on this: customers don’t buy based on your fiscal schedule, so ensure that an arbitrary budget doesn’t prevent you from giving them the opportunity to do business with you.
Budget Can Flow Freely Across Your Account & Campaigns: it isn’t enough to simply have ad dollars available in your account; those funds need to be able to be deployed into the Ad Sets / Ad Groups where they are needed. This is where Campaign Budget Optimization (Meta) comes in handy: it allows Meta to dynamically re-allocate your Ad Set budgets to your top performers. Google Ads does this automatically (unless you have ad group spending restrictions) – so as long as you’ve satisfied #1 above, you’re good to go.

This Applies to Data, Too: Just as your ad budget must be able to flow freely, so too must your data. In practical terms, this means segmenting only when the expected value of doing so justifies it (i.e. no more segmenting for the sake of segmenting). My test is always “maximum sustainable” – how large can a given Ad Set / Topical Cluster get before the need to segment becomes an imperative? The more consolidated a given account, the higher the level of data liquidity, the higher the overall expected performance.

Optionality:

If liquidity is the turbine that generates electricity, optionality is the grid that distributes that power across an entire ecosystem.

I think of optionality as a diversity of:

Targeting
Placements
Audiences
Campaign Types
Landers
Creative
Offers

Which gives machine learning more possible routes through which to invest your capital. The more optionality you provide, the higher the probability that the targeting algorithms will be able to find impressions that satisfy your targeting criteria.

If you’ve been following thus far, you’re probably thinking, “Well, not every audience / product / creative has the same expected value for us – some customer/prospect types are objectively better than others. Some offers have higher LTVs than others.”

And this is the crux of why I’m in favor of automation: because leaning into it gives you (the smart, savvy advertiser) more time to set the right targets and more time to think through which angles/offers make sense and more time to create the assets (creative, landers, etc.) necessary to take advantage of machine learning.

The more optionality you provide to the system, the higher the probability that it will find impressions that satisfy your constraints and meet your objectives. Those might not be the ones you initially thought would work, but at the end of the day, we’re in a results-driven business. If we check our ego at the door and focus on what’s actually moving the needle inside our ad accounts, we’ll all be a bit happier.

Constraints:

If the last two sections have made it seem like I’m a fan of unrestricted spending and unfettered machine control; that couldn’t be farther from the truth.

I’m a proponent of responsible media investment and reasonable constraints.

To put it simply, constraints are one way in which we (advertisers) communicate to machines what is important to us and what we know.

There are a number of constraints in an ad account, but the core ones are:

Direct Exclusions – whether keyword or audience or placement, there are some things that a human knows aren’t relevant, but a machine might not. The better you (as an advertiser) are at eliminating the garbage, the better off you’ll be.
Efficiency Targets – setting ROAS/CPA targets is a constraint – perhaps the most important constraint – on an account. The more stringent the target (High tROAS / Low CPA), the fewer potential impressions will have the expected value required for an ad to serve.
Positive Targeting – providing machines with helpful signals as to what is most likely to work can dramatically accelerate machine learning (and no, Meta / Google don’t like when you do it). There’s a reason I still use interest stacks and LALs in many ad accounts: because they give an advantage. In fact, some of the highest performing ad sets during BFCM 2022 + 2023 were LALs off of top customers.
Creative + Lander – we’ve all heard the “creative as targeting” thing over the past year, and for good reason: it’s true. Ad platforms have gotten incredibly adept at pattern recognition between ad creative, lander and audience, to the point where they can identify potential targets for an ad based exclusively on the creative added to the account. Does this work better when combined with other exclusions above? Yes. But can it work in a vacuum? Also yes.
Actual Data – Optimization events are constraints on smart bidding. There’s a reason I advocate for targeting the thing you actually want (purchases or qualified leads or signed cases or whatever) vs. some interim milestone: because there are a LOT of people who have relatively high probabilities of completing interim milestones and relatively low probabilities of completing the actual thing you want. Those are the exact people that you (the advertiser) don’t want, but are the exact people that the algorithm will serve your ads to, because they will appear under-valued. Your data serves as a constraint on advertising algorithms; the better that data, the more effective the constraint.

This is a brave new world for all of us – and navigating it is challenging. What worked in the past is unlikely to work in the future (or, at least, work in the same way). In the words of Taleb, “Adapt or Die.” It’s true for risk management, it’s true evolutionarily, and it’s true here.

The sooner you learn to embrace the principles above and integrate them into your account management, the better your accounts will perform, the happier your clients will be and the more time you’ll have to do the real, important strategic work required to keep your accounts performing.

Until next time,

Sam

Related Insights