What you'll learn in this article
- Which of the performance max bidding strategies I actually start on, and why it is almost never the one the client asks for
- How Maximize Conversions, target CPA and target ROAS each behave inside the learning phase, from accounts I run every day
- The single most common mistake I inherit: a target set too aggressively, too early, and what it does to delivery
- The exact sequence I use to transition from Maximize Conversions to a constrained target without resetting learning
- The numbers I wait for before I trust a target, with real examples of what "enough data" looked like on my campaigns
Almost every account I inherit has the same problem hiding in its performance max bidding strategies: someone set a target on day one because the goal felt tidy, and the campaign has been quietly starved ever since. The bid strategy is not the hard part of PMax. Choosing the right one for the stage the campaign is actually at is, and that judgement is where I have watched more budget get wasted than in creative, feeds and audience signals combined.
So this is not a definitions post. Performance Max runs on Smart Bidding only, there is no manual CPC to fall back on, and that means the strategy you pick is the campaign's entire nervous system. What I want to walk through is how the three options behave in the one window that decides everything, the learning phase, the error I see forced onto campaigns before they have earned it, and the transition I run once the data is real. All of it comes from accounts I manage, not from a help-center paraphrase.
The through-line is simple: an early target is a constraint on a model that has nothing to constrain yet. In my experience the campaigns that end up cheapest are usually the ones I let breathe first. If you are still deciding whether PMax is even the right container for this spend, my breakdown of what Performance Max is and how it fits an account is the better place to start; here I assume the campaign exists and the only open question is how it should bid.
The three performance max bidding strategies, and what each is really for
Strip away the labels and PMax gives you two goals and two optional guardrails. You either maximize conversions or maximize conversion value, and onto either you can bolt a target: a target CPA on the conversion goal, a target ROAS on the value goal. That is the whole menu. Everything written about performance max bidding strategies is a variation on when to add the guardrail and how tight to make it.
Maximize Conversions with no target is the accelerator with the floor removed. It will spend the full budget chasing volume, and in the learning phase that is a feature, not a flaw, because volume is exactly the data the model is starved for. Maximize Conversion Value behaves the same way but weights each conversion by the value you feed it, so it is only honest if your conversion values are passing back correctly from the site; a value goal on top of broken values is worse than no goal at all.
The constrained versions are where accounts get hurt. Performance max maximize conversions target CPA tells the model to hold an average cost per action while it bids, and target ROAS does the same for return. Both are legitimate, mature-campaign strategies. The problem is never the strategy, it is the timing: a target is an instruction to leave cheap-but-scarce conversions on the table in order to protect an average, and a campaign with forty conversions of history has no stable average to protect. I treat the target as something a campaign graduates into, which is the opposite of how most of the accounts I take over were set up.
How each strategy behaves in the PMax learning phase
The learning phase is not a countdown, it is a data threshold. The model is building a map of which placements, audiences and creatives convert, and until that map has enough points on it every bid is partly a guess. How your performance max bidding strategies behave in this window is almost entirely a function of how much freedom you give the model to gather those points.
Maximize Conversions: the one that actually holds
On new campaigns I start on Maximize Conversions without a target, and it is the strategy that reliably survives the learning phase, because its only instruction is "spend and convert." It explores freely, fills the map fast, and gives me a clean read on what the account's real CPA looks like when nothing is being suppressed. On a lead-gen account I launched earlier this year, the first two weeks on unconstrained Maximize Conversions landed around a €38 CPA at full budget; that number was not the goal, it was the measurement I needed before I could set one honestly.
Target CPA and target ROAS: right strategy, wrong phase
Drop a target onto that same campaign in week one and the behaviour changes completely. The model, told to hold a CPA it has no data to hit, gets conservative: it bids down, delivery collapses, spend falls well under budget, and the campaign crawls through learning instead of powering through it. I have seen a target-ROAS PMax spend a third of its budget for three straight weeks and never stabilise, purely because the target was set before there was any conversion history to anchor it. The strategy was not broken. It was asked to do a mature-campaign job with a brand-new campaign's data, which is the core reason the learning phase misbehaves when budgets and targets are set wrong.
This is where budget and bidding stop being separate decisions. A target only works if the budget gives the model room to hit it; I set daily budget well above the target CPA precisely so the guardrail has somewhere to operate. The interplay is why I treat bidding and budget as one setting, and why my note on how PMax budget shapes delivery sits right next to this one in my own process.
The mistake: forcing an aggressive target too early
The single most damaging thing I inherit is a target set too aggressively, too soon. It usually comes from a good instinct the client knows their target CPA is €25, so they type €25 into a week-old campaign and expect the machine to comply. What actually happens is that the model reads €25 as "only bid where you are already confident of a €25 outcome," and on a campaign with no confidence anywhere, that means almost nowhere. Delivery dries up, the learning phase never completes, and everyone concludes PMax "doesn't work" when the truth is it was never allowed to learn.
The tell is always the same shape in the data: budget massively underspent, impression volume a fraction of what the auction could give, CPA on the handful of conversions that do come through looking deceptively fine because the model is only taking the safest bets. It looks efficient. It is actually paralysed. Google's own guidance on Maximize Conversions and its optional target is explicit that the unconstrained version exists to spend the budget and gather conversions that is the phase you skip when you force a target on day one.
The correction is rarely "loosen the target." It is "remove the target entirely, let it learn, then reintroduce a realistic one." The aggressive number is not just too low, it is premature, and no amount of tuning fixes a guardrail installed before there was anything to guard.
How I transition from volume to a target on my accounts
Once Maximize Conversions has done its job, the move to a constrained strategy is the part people rush and I deliberately slow down. The transition is where performance max bidding strategies either compound the learning you paid for or throw it away, so I follow the same sequence every time.
Wait for real data, not the calendar
I do not switch on a date; I switch on a number. In practice I want a campaign clearing at least 30 conversions in a 30-day window, and ideally more, before any target goes on that is roughly the volume at which the model's average CPA stops swinging week to week and starts meaning something. On the €38-CPA account above, that took about four weeks; only then did I have an average worth protecting.
Set the first target off actual performance, not the wish
When I introduce the target, I set it at or very slightly below the CPA the campaign is already achieving, not at the client's dream number. If it is running at €38, my first target is €38, maybe €35 never the €25 the client wants on day one. The point of the first target is to hold performance while the model adjusts to being constrained, not to squeeze it. Once that holds for a couple of weeks I tighten in small steps, 10-15% at a time, watching that volume does not fall off a cliff with each turn of the screw. Every tightening is effectively a small re-learning, which is why the same discipline that governs a standard campaign's bid changes applies here too: move in increments the model can absorb.
Value-goal campaigns follow the identical logic with the axis flipped. I let Maximize Conversion Value run until ROAS stabilises, read the real number, set target ROAS at that level rather than an aspirational one, then climb. The strategy label changes; the graduation principle does not. Done this way the transition costs almost nothing, because I am never asking the model to hit a target it has not already shown me it can reach I am just formalising a number the data already earned.