Growth marketers’ biggest miscalculation in the AI era has been overinvesting time and effort into categories that should be automated, and not spending nearly enough time shaping the inputs that actually impact performance. Hours still go into bid rules and dayparting that the platforms already took over, and the levers that actually moved performance went untouched.

It’s not surprising, and it’s not laziness. We’re inclined to orient our work around where we’ve seen payoff for years. But the job changed categories. You are no longer optimizing. The platform is optimizing. Your job now is to define the space it optimizes inside: what counts as winning, what each version of winning is worth, which users it can bid on, and which creative it can put in front of them.

The levers that still need your hand on them are the controllables.

Three conditions decide what you hand over

There’s a simple test for which side any given lever falls on. Handing an optimization to the platform requires three conditions:

Let’s take ad copy in Google Ads and apply this rubric to the question of, “should I allow Responsive Search Ads (RSA) to automate ad composition and serving?” Utilizing RSAs allows Google to mix and match 15 headlines and 4 descriptions to compose ad combinations. Ad copy is most frequently evaluated against its ability to generate clicks effectively (i.e., CTR), so there’s a fast signal back to Google on which combinations are successful vs. not. You are sampling clicks against a universe of impressions, so most ad groups with any sort of budget tick the box for sufficient volume at the decision-making layer. Finally, there are 32,760 permutations of ads that are possible: this is a decision space that is way too combinatorially large for manual rules.

Now run the same test on the campaign objective and the conversion values behind it. A B2B advertiser pointing Smart Bidding at closed-won deals is asking it to learn from a signal that lands months late and a few dozen times a quarter. The platform will bid on it. It just won’t learn much. So somebody has to decide which upstream event stands in for revenue, and what each one is worth relative to the others. An MQL isn’t a demo request isn’t an SQL. That decision isn’t an optimization, it’s a judgment about your own funnel, and it’s the first controllable on the list.

The platform earned these

As much angst as there is around automation, it’s a welcome reprieve from our digital marketing past.

In 2010, bidding entailed separate values for individual keywords with tweaks made daily based on performance by match type, time of day, day of week, device, keyword word length, the modifier word performance, and on and on. On Facebook I’d analyze performance by age, gender, interest, as well as some of the settings listed for Google.

The analysis itself took place in Excel, where you would pivot out different views to try to understand which buckets to increase versus decrease bids on before layering dozens of rules-based modifiers tied to degree of over vs. underperformance. The whole process ran multiple hours per week per client.

I can be much more brief about how this is all handled now: algorithms in each of these respective platforms, developed by thousands of the smartest data scientists, engineers, and product managers in tech, working off of the largest treasure trove of advertising data in existence, have automated all of it. And for the avoidance of doubt, it performs better against your goals than it ever has before.

This expands well beyond bidding. Broad match keywords were once the Scarlet Letter of a Google AdWords account. A Facebook advertiser who ran on a broad audience rather than targeting by interests, behaviors, or custom audiences would have been accused of lighting ad budgets on fire. Now, they’re best practices that effectively expand reach while keeping your ads in front of a relevant audience.

What are the other categories where you should defer to the platforms and let them cook?

Let the platform cook

Each passes all three conditions

Each of these checks all three boxes. They’re closed-loop optimizations inside a decision space, and the platforms have earned them.

Defining the controllables

A controllable is a decision rather than a calculation. The platform runs the calculations, and it runs them better than you do. What’s left are the calls where the right answer depends on what the business wants: which outcome is worth chasing this year, what a new customer is worth against a repeat one, which idea is worth making, what to believe when two measurement systems disagree. None of those has a correct answer waiting to be computed. Each one is a choice with something traded away on the other side.

The decision space is the cross-product of every dimension the algorithm reasons about. Which inventory, which creative variant, which bid, against which user, for every impression. The controllables are the levers that define it, and they fall into four groups.

The eleven controllables

Our top three if bandwidth is tight

Four groups, eleven decisions. Later posts in this series take each group in turn.

GROUP 01 OF 04

Setting the target

Objective selection (i.e., sales vs. leads vs. traffic) is the ultimate representation of defining the decision space. “Show me a goal and I’ll show you an outcome” rings true within platform automation. But while the reward signal is fast and unambiguous to Google, it’s ambiguous to your P&L. Inside that objective, the bid strategy is the second coarse dial: Maximize Clicks, Maximize Conversion Value, Target CPA or ROAS. Both dials come down to the same act, translating a business goal into the best proxy the platform will accept.

The easiest wins we’ve notched in the last several years came from exactly that translation.

Two of our easiest wins went in opposite directions. The right proxy depends on the funnel, not on a rule.

Value orchestration is the next dial in: what counts as a conversion, and what each one is worth. Do an eBook download and a webinar signup from the same person count once or twice? Does a first-time buyer’s purchase count the same as a repeat buyer’s? The answer can be deployed directly into platform rules through value-based bidding or similar products. Or you compute the value in your own system and apply adjusted CPA or ROAS thresholds by campaign.

GROUP 02 OF 04

Feeding the model

We’ve all heard the phrase “garbage in, garbage out.” That applies doubly when feeding the ingredients that will fuel the algos. Bad data has a compounding effect that will make ongoing optimization worse, not just the initial performance. Signal quality is the act of providing the outcome data and raw materials that the algorithms learn from. This can include offline conversion imports, putting in place server-side tracking, de-duplicating tags, DQ’ing unqualified form fills and anything else that enhances the training set.

Creative inventory is the most important lever on this list, and it’s here for a different reason than the rest. The platform isn’t incapable. It will generate concepts, assets and variants all day, and the variants are often good. The problem is where they converge. Generative creative optimizes toward the mean of what has already worked, using the same models and the same priors as every other advertiser running them. That’s excellent for producing the ninth version of an idea and structurally incapable of producing the first one, because the first one isn’t in the training data yet.

So the work is the concept, and the judgment about which concepts are worth the production. The raw assets, the copy and imagery, enough diverse options to give the search something worth finding, and the call on which assets get tested against which audiences. Producing a swath of variations is table stakes now and getting cheaper every quarter. Deciding what’s worth varying is not.

Data legibility is a major lever that shows up most prominently in product feed management. In nearly every retail account we ran at Metric Theory, GTIN coverage and attributes were incomplete, while titles and descriptions were frequently devoid of usage context or situational background. These all allow you to enter auctions more frequently with higher success rates. Structuring your site in a way that allows AI to derive targeting and copy ideation cleanly (i.e., for AI Max campaigns in Google) also falls into this theme.

GROUP 03 OF 04

Drawing the boundaries

These elements all help form the contours of the decision space: how big and in what shape are the gates around what can be optimized. The algorithm is as good as what it’s allowed to see: audience eligibility is that end pool of users. You form it with seed audiences, exclusion lists, in-market versus custom intent settings, and the match rates on the lists you upload. We tested lookalikes built from all-customer seeds against last-90-day purchasers against high-AOV purchasers, and they rarely performed the same. For a women’s retailer running holiday gifting, we built the lookalike off the previous year’s male purchasers, and it beat the all-customer seed.

Campaign architecture might not carry the same weight it did in 2010 when growth marketers were sculpting very granular targeting and ad combinations. However, the campaign structure, flight dates, budget caps, and pacing determine which decisions get pooled into one signal pot versus kept separate. There’s a constant see-saw between pooling data density that allows for more signal for the algorithm vs. more control over which audiences receive more budget: it’s a worthwhile place to intervene to tip the scales.

GROUP 04 OF 04

Crossing the walls

Cross-channel work fails condition 1 by definition. Platform automation is excellent inside its own walled garden and blind outside it, so anything whose reward signal spans two of them lands back on your desk.

Most brands and agencies test, but within the narrow confines of an individual channel. Two teams launch a creative test and a value orchestration change in the same week and nobody can say afterward which moved the number. Testing roadmap is the call about what runs, in what order, and what gets left alone. It means weighing the risk, reward and likely learning from every test on the list against how much noise the others will throw off if they run at the same time. No platform can make that call, because no platform can see the other tests.

Measurement approach allows you to extract conclusions from that testing that drive growth from learnings over time. Platform tracking consistently double (or triple, or quadruple) counts conversions and doesn’t answer whether those conversions required the aid of an ad to generate the purchase. So while you should leave allocating credit within a channel to the machine, lift, incrementality and holdout tests are how you find out which activity actually produced a new customer. The platforms will run one for you, and both Google and Meta offer them. Two problems with taking the result at face value. The platform is designing and grading a test on its own inventory. And a single-platform study can’t answer the question you actually have, which is whether the next dollar does more in Meta than it does in Google. Somebody still has to design the test that spans both, and decide what to believe when the answers conflict.

Media mix could be automated, and the vendors will tell you it already is. Doing it well means weighing incrementality results against channel momentum, MMM output, and whatever is coming next quarter that nobody has modeled yet. Three of those four inputs don’t exist in any platform’s data. Where the dollars go is still the most consequential decision you make, and it’s still yours.

Lastly, insight distribution is what makes the other ten compound. The creative test result has to get back from the paid social team running the media to the creative team that made the ads. The product feed manager who found success appending use cases onto product titles should inform the GEO team so that they can reflect those use cases in future content. The data team that sees new customers acquired from the updated prospecting campaigns delivering a modeled LTV that’s 25% higher than historical averages needs to share that with the paid search team so that they can adjust the value-based bidding amounts. These are wildly different workflows, and insight latency between them can have a material impact on growth.

The platforms are coming for this list. Not all of it.

Several items on this list are already available to automate, and the honest read is that the list will shrink. Could and should are different questions. Advantage+ audience expansion is available and defaults on, and we’ve seen it run against a stale exclusion list often enough to know the bill: you buy your existing customers back at prospecting rates.

The bigger concession is that the data gap is closing. Offline imports, CAPI, warehouse connectors and MMM tooling all exist to hand the algorithm more of what it currently can’t see, and that only moves one direction. Any argument resting on the platform being blind to something has an expiration date. The exception is structural rather than technical: Google is never going to hand Meta its auction data, so anything spanning two walled gardens stays yours regardless of what either roadmap ships.

What doesn’t close is the judgment. A model with every number you have still can’t tell you whether this is the year to trade reported efficiency for new customers, what a first-time buyer is worth against a repeat one, or which of three conflicting reads on incrementality to act on. And there’s usually a multi-year lag between an automation launching and that automation clearly beating what a good operator does by hand. That gap is where the advantage sits.

If you only have bandwidth for three, take creative inventory, value orchestration and measurement approach.

Controllables in practice: eleven decisions on one account

Let’s say your brand sells ski and snowboard goggles with swappable lenses. The brand is facing declining growth. The underlying issue is that while revenue has stayed fairly flat the last couple of years, new customer acquisition is declining.

Working assumptions, illustrative rather than pulled from a real account: $400k a month across Google, Meta and TikTok, with the platforms reporting 8,000 purchases at a blended $50 CPA. Loaded fully against new customers only, that’s a $111 CAC.

That creates three problems, and they compound. A lens buyer converts more readily than a goggle buyer and spends $70 instead of $220, so the algorithm learns to chase lenses. An existing customer converts more cheaply than a stranger, and 55% of reported purchases already come from people on the customer file. And most of those existing customers would have bought anyway, direct or through email or organically, which makes the conversion the platform is proudly reporting one it didn’t cause.

So, what can you do? First, upload a customer list that you fully exclude from that campaign to avoid showing ads to existing customers to begin with (Audience eligibility). The match rate won’t be 100%, so some existing customers will still sneak through, but it’s a great start. Next, set up a new customer acquisition goal at the campaign level (Objective selection). Last, assign a higher value (i.e., 2x the value) to a purchase when the user isn’t on your existing customer list (Value orchestration).

The execution mechanism varies by platform. In Google, you’d use Maximize Conversion Value or target ROAS against the custom conversion rule. In Meta, you’d pass a parameter on the purchase event indicating new vs. returning (matched against your CRM) via CAPI (Signal quality) and use “maximize value” bidding. Again, these algorithms will do what they’re told. Now they’re told to maximize value, and you’ve set the value of new customers higher.

To date, the highest performing ads have been assets that highlight swappable lenses paired with either situational settings (spring skiing vs. storm day) or a promotional offer. This is a self-fulfilling prophecy of the campaign objectives favoring lower-AOV repeat purchasers. You decide to develop new concepts focused on new customer acquisition: UGC ads extolling the virtues of the goggles, a lifestyle concept called “See the Stoke Clearly” and product-focused ads showing the goggles in use (Creative inventory). This diversity of concepts increases the probability of you finding a step-change improvement in new customer acquisition performance via creative.

A thorough audit of product feeds also reveals a litany of issues. GTINs for newly added products are missing. Titles don’t uniformly reinforce that new goggles come with one free swappable lens. Product descriptions are light and rarely pair rich context that helps mirror realistic prompt sets (most durable category 1 lenses for heavy fog conditions). Addressing feed health increases the eligibility to show on key products within DABA, Smart+ and PMax campaigns, improves performance when you do show, and shifts the composition of impressions to goggles that are more often purchased by first-time buyers (Data legibility).

You realize campaign structure is not allowing for the right type of budget controls either. Prospecting and remarketing audiences are blended in Meta and TikTok. On Google, one search campaign contains keywords for lenses, goggles, full sets, and even apparel. Segmenting these allows you to allocate more budget to prospecting campaigns that are excluding past purchasers, as well as prospects searching for goggles and sets rather than lenses. As you approach the holidays, it also makes it easier to run a promotion specific to swappable lenses for past purchasers who already own goggles (Campaign architecture).

You’d want to stagger the new creative testing with the adjustment in campaign objectives and conversion values to try to isolate impact. Feed optimization, meanwhile, would impact a different set of campaigns, so it could run concurrently. Volume, performance and creative volatility all factor into the decision to launch new creative on TikTok rather than Meta first (Testing roadmap).

A geo holdout test found that ~70% of repeat-buyer purchases happened anyway through direct, email or organic. That means the true incremental CPA is $81, not $50, and that dashboards understate the real acquisition costs by ~60% (Measurement approach). Among the platforms, Meta delivered a higher share of incremental purchases and the lowest CAC for new customers, even though overall CPA trended higher. Top-performing creative achieved an even more significant delta over average performance there as well, leading you to shift 10% more budget away from Google and TikTok to Meta (Media mix).

The product feed enhancements yielded stronger performance overall, but products tagged in titles and descriptions as being “perfect for low light powder days” saw a particularly large lift. The feed manager coordinated with the SEO/GEO, creative and influencer teams to use more of that language in site copy, YouTube channel content, new creative tests and UGC scripts (Insight distribution).

Here’s what that adds up to. Reported blended CPA gets worse, from $50 to $59. New customer CAC improves from $111 to $90. You will have to defend that first number to someone, and being able to is the entire point of the second one.

Everyone gets the same algorithm

Notice that in every step in the example above, the actual bid execution, real-time auction decisioning and audience matching is still automated. You’ve redefined the decision space, not the real-time decisioning.

Every advertiser running Smart Bidding gets the same Smart Bidding. Every advertiser on an AI Max campaign gets the same logic, trained on the same data. The automated layer is a commodity. The differentiator isn’t how you allocate your time, it’s the space you hand it. Your value definitions, your exclusions, your feed, your concepts, your measurement. That space is the only part of the system a competitor can’t buy.

The platform will happily generate the 32,761st ad. It will look a lot like the other 32,760, and so will your competitor’s. Everything that makes your account different from theirs is something a person chose.