On the face of it, maximise conversion value beats maximise conversions. If you can align the algorithm’s priorities more closely to yours, why wouldn’t you?
When you have any basis at all for differentiating the values of different conversion actions (always, in practice) you can feed those differences to the algorithm. You don’t have to be precise about it. I’ve often said that if you can set a value per conversion that’s even slightly more in line with reality than a flat value across all conversions (not hard to achieve, even with a finger-in-the-air estimation) then you will be guiding the algorithm better than max conversions can.
This seems like one of those logically irrefutable principles once you understand it.
What I’ve recently learnt is: not necessarily.
I was looking to improve a plateauing account a few months ago and, in line with the principles above, decided to start the move to value-based bidding.
It made sense. We had differentiated conversion actions representing different lead stages (I’ll call them Red and Yellow) with heavily differentiated values in line with the proportion that proceed to each stage.
• Red (initial lead): $3
• Yellow (qualified progression): $641
There’s a stage three beyond Yellow, common enough in the big picture but too rare at campaign level to bother with here.
So I moved the two largest campaigns over to maximise conversion value in late January / February. As I explained to a Google rep at the time:
"Beyond tCPA, I think this account probably needs to get back into the value-based bidding that it's quite well set up to exploit. The different conversion actions represent different stages of the downstream offline conversion process - but the campaign has been on a simplified footing for the last few months, and during that time, we've definitely had degradation of conversion quality... and the proportion of higher-value conversions has been decreasing."
When the ‘inevitable’ improvement didn’t happen in the first few weeks, I was unperturbed. Differentiated values would surely win out in the end…?
A month later, replying to the same Google rep (who was keen on the move and pushing for a wider rollout), I wrote:
"We're not seeing any noticeable benefit from the VBB change. I'm fully holding with the advantages in principle, but not in a rush in this case, because none of the KPIs have improved while CPC is a little up where we're trying it. Not giving up on it, but not rushing to roll out."
Well. March rolled on, then April, and the lag turned slowly but indisputably into a reversal.
I consolidated the two campaigns I switched to MCV against two comparably-sized campaigns that stayed on max conversions throughout:
CPCs in the switched campaigns jumped 61% in the month in which we switched strategy. The campaigns left on max conversions saw around 14% over the same period. Red volume dropped a third. And Yellow, the thing we were actually chasing, also declined. The switched campaigns produced just 2.5 Yellows in February, falling to 2 in March, while the others held at 3.
How could we be doing so much worse? Not just failing to leverage those values, but falling back from where we were on max conversions?
The mechanism
The value you assign to a conversion tells the algorithm how much to care about achieving a certain kind of outcome. But the frequency of the conversion is what tells it how to achieve it. A higher value does not compensate for a lack of frequency.
Maximise Conversion Value bids on expected value. For every click it considers, the algorithm forms a prediction: this user, this query, this asset combination, this time of day, has an X% chance of producing a Yellow and a Y% chance of producing a Red. It bids accordingly. If the click matches its best bet as to what leads to a Yellow, it’ll happily pay $20-30, because $20 against an expected value of $40 is a trade worth making.
That’s a beautiful machine when the predictions are any good. The problem is the predictions need data, and Yellow at five per month per campaign isn’t real data; it’s anecdote. So the algorithm builds its prediction model on noise. It pattern-matches based on whatever happened to coincide with the last few Yellows, a particular device, a particular hour, a particular asset, and pays Yellow-priced CPCs for traffic that doesn’t justify them.
This is what causes the problem. Those higher CPCs mean fewer clicks for the same budget, which means fewer conversions of every kind, including the Reds, which were doing perfectly well on the old strategy. The algorithm isn’t malfunctioning; it’s doing exactly what it’s designed to do, but without the raw material to do it effectively. You’ve asked it to chase a pattern that doesn’t exist at this volume, and in the attempt it’s bidding up the clicks it thinks will lead to a Yellow, which, given how little it has to go on, is a fairly arbitrary subset of its traffic.
So you end up losing on both fronts: fewer Yellows (the thing you were optimising for) AND fewer Reds (the collateral damage of paying premium prices for ghost-Yellow traffic). The algorithm is trying its hardest. It just can’t find what you’ve told it to look for.
This is worth pausing on. Whenever we’re choosing what to feed the algorithm, we need to consider not just “what do I care about?” but also “what can the algorithm learn to predict?” Caring about something doesn’t make it learnable, and when the thing you care about is too rare to learn, telling the algorithm about it doesn’t get you closer to it. It just makes the algorithm worse at everything else.
Say you’re hunting for Easter eggs in the garden. You’re finding plenty. Then someone tells you there are golden eggs hidden somewhere out there too, worth two hundred times more. You don’t know what makes the golden ones different, or where they are, but you know they exist. So you change your approach. You stop following the trail that was working. You start guessing, second-guessing, wandering into corners of the garden you’d normally ignore. You find fewer chocolate eggs. You don’t find any golden ones either. The promise of gold didn’t help you. It just made you lose sight of what was already working.
That’s what happened here. The algorithm had a perfectly serviceable model for finding Reds. The moment we told it about the enormous value of Yellows, it lost sight of what it knew and started chasing what it couldn’t see.
What to do about it
Two ways to bridge this gap.
1. Insert a stepping stone.
Our Yellow is a mid-funnel step, but it’s still too rare to guide bidding reliably. An intermediary event could help: something frequent enough for the algorithm to learn from, but directionally closer to Yellow than Red is. A ‘hot lead’ action, for instance, that fires when a lead does something that correlates with eventually becoming a Yellow. Not as valuable as Yellow, but occurring often enough that the algorithm can actually model what drives it.
This is exactly what we’re now implementing in this account.
Why is this better than aiming straight for Yellow? Because although Yellow conversions are truer to your real destination, the algorithm can’t see the path to them. A well-chosen proxy conversion is something it can actually see, actually learn from, and actually navigate by. A visible stepping stone beats an invisible destination.
It’s like Ariadne’s ball of string. The Minotaur is the goal, but it’s hidden at the centre of the labyrinth. And if Theseus heads vaguely towards the centre, he will inevitably get lost. The string is what gets him there, not the destination, but the path. Your proxy conversions are the string. Without them, you’re just wandering expensively in the dark.
(Did you know this is the origin of the word ‘clue’, which originally meant a ball of string?)
2. Compress the value range.
The value ratio between Yellow and Red in this account is 213:1. The problem is the combination of the size of the gap and the scarcity of the Yellows. Their huge value dominates the expected value of any click, while their scarcity gives it no real directional steer.
The size of the prize compounds this. A Yellow worth 213 Reds doesn’t just nudge the algorithm in Yellow’s direction, it dazzles it. The algorithm gets so disoriented by the potential value of Yellow that it abandons what it can do reliably in favour of what it can’t.
If you compress the range, say, make Yellow five to ten times more valuable than Red rather than two hundred times, you give the algorithm something it can work with. It will still steer towards Yellow-likely traffic, but it won’t bet the farm on a prediction it can’t substantiate.
(It’s a bit like the coach telling his players before a tournament final that this is ‘just another game’. It isn’t. But they’ll play better if they believe it is.)
Both fixes are versions of the same move. In each case you’re trading a bit of theoretical accuracy for a usable learning signal. That trade is worth making almost every time, because an accurate instruction the algorithm can’t act on is worth less than a rough one it can use.
Â
