Five Labels, One Dial: A Brief History of Meta Bid Strategies

Abstract

Five bidding strategies, one underlying mechanism

Today there are five Meta ad bidding strategies:

1. Highest Volume
2. Bid Cap
3. Cost per Result Goal
4. Highest Value
5. ROAS Goal

Meta offers some descriptions, and there is no end of blog posts and videos happy to regurgitate them. But how do the strategies actually work? What can we say for sure? This paper uses history as a guide: it follows the bidding options from their origins (circa 2015) through every rename and retirement, and reproduces each one's behavior in a simple auction simulation.

The record is surprisingly legible. Each name change looks cosmetic on the surface, but the sequence of changes reveals the machinery underneath.

• The "bid" Meta asks for has always meant one thing: the conversion cost you are willing to pay, whether you specify it explicitly or not. Meta multiplies it by an estimated conversion probability to produce the score that decides each auction.
• Spending a budget means scaling the bid, and therefore the score, by yet another multiplier, a technique called bid shading, or discount pacing. And for every budget there is exactly one bid and multiplier that spends it all.
• Bid caps, cost goals, and ROAS goals are not different algorithms. They are different ways of pinning the bid and multiplier.
• Target Cost was the one strategy built on a different technique, probabilistic pacing, and was the one that Meta retired.

Bottom line: Highest Volume, Bid Cap, Cost per Result Goal, Highest Value, and ROAS Goal are five labels for one dial: a multiplier applied to the bid, which is the conversion cost you are willing to pay (even if you don't explicitly specify it). At any moment, either your budget or your cap (or floor) sets that dial, never both, so one of the two numbers does nothing.

Section 1

The five strategies, as Meta describes them

Here is the current menu, in Meta's own terms, lightly edited:

Paraphrased from Meta's advertiser help documentation, 2026 naming.

That's five names representing two "spend it all" options, two caps, and a floor. The descriptions say what each strategy aims for, but not how any of them works, and not how they relate to each other. To answer that, it helps to go back to the beginning.

Section 2

Before 2015: where the "bid" came from

In the early days of real-time ad auctions, advertisers bid by cost per mille (CPM), the cost per thousand impressions, or by cost per click (CPC). Neither was satisfying, as both put the burden on the advertiser to work out what an impression, or click, was actually worth.

As time went on, Meta, née Facebook, became quite good at predicting the estimated conversion rate (eCVR) of a given ad impression, or the probability that a user would perform the desired action. That prediction enabled a better question. Instead of "what will you pay for an impression?" Meta could ask "what is a conversion worth to you?" and then compute the effective CPM (eCPM) on your behalf. This new "bid" was the optimized CPM (oCPM), and it represented the cost you were willing to pay for a conversion, or your desired cost per action (CPA).

eCPM  =  oCPM  ×  eCVR  +  quality

The oCPM stayed constant, while the eCPM was calculated for each auction, based on Facebook's eCVR. The quality was, and still is, a subjective measure Meta uses to penalize bad ads, or reward good ones; it can be ignored here because it doesn't change the overall conclusions. Eventually the term "oCPM" was dropped and the user interface simply asked for your "bid." Likewise, this analysis will use newer, simpler terms: the oCPM is the bid, the eCVR is the conversion probability (p), and the resulting eCPM is the score that decides each auction.

Note that "bid" was, and still is, a misleading term. It's not the number you call out to the auctioneer, the one that ultimately decides the auction. Here, that's called the score.

The "bid" you enter is the cost you're willing to pay for a conversion, which stays constant, and Meta scales it, in each auction, by the probability it expects a particular user to convert.

Section 3

2015: two options, one mechanism

In 2015 there were only two bidding strategies:

Manual bidding asked for a bid, or what you were willing to pay for a conversion. Automatic bidding didn't ask for a bid at all. It set one for you, and it was the one bid that would spend the entire budget.

Consider a simulated day of 1,000 auction slots, or opportunities to show an ad, with only two ads competing for every slot: yours and a competitor's. As each slot becomes available, Meta estimates a different p for each ad. The score that determines the winner is p × bid. Assuming both your bid and your competitor's bid are $20, Figure 1 shows the outcomes for the first 100 slots of the simulated day.

100 auction outcomes

Both ads begin with a $20 bid. Move your bid to see which slots change hands.

01

You (lost)You (won)Competitor (lost)Competitor (won)

Your bid

$20.00

Your ad does not simply win the slots where its own p is highest. It wins where its p is highest relative to the competitor's. Your lowest possible bid, the bid that wins only one slot, doesn't capture the slot most likely to convert; it captures the slot most likely to convert relative to the competition.

Section 4

One budget, one bid max

If you increase your bid in this scenario, you'll win more slots. But your total spend and average cost per expected conversion will rise with it. Figure 2 shows what happens when your bid sweeps from $1.00 to $70.00.

One budget, one bid max

Sweep the bid from $1.00 to $70.00, or play the original sequence.

02

Slots won672Spend$399.46Average cost$12.26Budget$400.00

You (lost)You (won)Competitor (lost)Competitor (won)

Bid

$30.61

Play bid sweep

Three things fall out of this picture:

• The optimum exists and is unique. With this day's auctions, there is exactly one bid, the bid max, that spends the budget ($400), or at least as much as possible without going over. Below it, budget is left on the table; above it, the budget runs out before the day does.
• Everything moves together. Spend, wins, and average cost all rise together with the bid. Fix any one of bid, spend, or average cost and the other two follow. There is no separate lever for each.
• This is the target every pacing system hunts. A real delivery system doesn't have perfect foreknowledge of the day. Nevertheless, the optimum it is chasing exists, and there is only one. The job of a real-time pacing algorithm is to home in on this one optimal value while the landscape shifts under it.

This is what "Let Facebook set the bid" actually meant: let the platform pick the one bid that spends the entire budget for the day. Automatic bidding was just Manual bidding with the number chosen for you.

If you don't choose your bid, Meta will choose it for you.

Section 5

What the bid (cap) actually did

Manual bidding required a budget, too, which meant that it also had an implicit bid max, the bid that would spend the entire budget, just like Automatic bidding. In fact, it was the same bid max that Automatic bidding would choose given the same budget (see Figure 2). The bid you entered acted, instead, as a bid cap, a name Meta would later use explicitly. That created two possible regimes, depending on where the bid cap was set relative to the bid max:

• Cap-constrained: bid cap at or below the bid max. The cap becomes the operative bid. Every score = p × bid cap. Below it, scores are lower across the board, fewer auctions are won, average cost falls, and the budget goes partly unspent. In the simulation, a $10 bid cap wins 250 slots and spends $87.33, roughly a fifth of the $400 budget (see Figure 3).
• Budget-constrained: bid cap above the bid max. The system scales the bid down by a multiplier (λ) to avoid blowing the budget. λ = bid max / bid cap, so every score = p × λ × bid cap = p × bid max. At a $50 bid cap, λ = $30.61 / $50.00 = 0.61 (see Figure 3). In the literature this is called bid shading, discount pacing, multiplicative pacing, or just pacing. Note that in this regime, the bid cap has no effect at all. The outcome is exactly the same as Automatic bidding on the same budget.

Bid cap relative to bid max

Move the bid cap across the point where the budget becomes the binding constraint.

03

Regimecap-constrainedλ1.00Operative bid$30.61Spend$399.46

Bid cap

$30.61

Note that, in other contexts, the term "bid shading" may be used to describe the separate practice of bidders lowering bids to gain an edge in first-price auctions. However, here it always means the multiplier that paces a budget.

If the bid cap is set high enough, Manual bidding behaves the same as Automatic bidding.

Section 6

2016 to 2019: the renames hint at the mechanism

After 2015, the number of available bid strategies increased and several were renamed. Each change to the menu provided a clue about what runs underneath. The table below describes the evolution.

The 2016 addition of Manual (average) confirms that the 2015 Manual used a maximum, as Sections 4 and 5 describe. The 2017 renaming looks odd at first: you would expect the two Manual strategies to be grouped together. Instead, the new names paired Automatic and Manual (maximum) under the term "Lowest Cost," and left Manual (average) on its own, renaming it Target Cost. A 2017 Facebook guide, "Facebook Bid Strategy," explains the reason for the grouping: Lowest Cost "uses ‘discount’ pacing to spend budget on results with the lowest costs." Target Cost "uses ‘probabilistic’ pacing… to maintain consistent cost per result over time."

In other words, the grouping follows the mechanism. Automatic and Manual (maximum) are the same discount-pacing machine from Sections 4 and 5. Target Cost is the black sheep. It uses probabilistic pacing, where the bid is never shaded. Instead, a random number between 0 and 1 determines whether the ad enters each auction at all. The dial is the chosen threshold, above which the auction is entered. The higher the threshold, the fewer auctions entered and the lower the spend. But the average cost per conversion stays constant regardless of spend, one of the strategy's distinguishing characteristics. However, we won't pursue Target Cost further, because Meta eventually phased it out.

That also resolves a puzzle that arose in 2019: Cost Cap and Target Cost appear to optimize the same thing, an average cost per conversion, which would make one of them redundant. The answer is that they coexisted because they used different techniques.

Section 7

What the cost cap actually did

Cost Cap was discount pacing, yet again. But instead of capping the bid, it capped the average cost of a conversion. What often goes unnoticed is that they're two sides of the same coin. Figure 2 depicts the two curves, shows how the values move together, and demonstrates that the cost max ($12.26) and the bid max ($30.61) occur at the same point, the point where the bid spends the entire budget.

Figure 4 makes it even more explicit: pinning the average cost pins the bid, which in turn dictates the spend. This creates two regimes, similar to the ones created by the bid cap and described in Section 5. In this case, the regimes depend on where the cost cap is set relative to the cost max.

• Cap-constrained: cost cap at or below the cost max. The cost cap sets the bid outright, just as a low bid cap does. In Figure 4, when the cost cap is below the cost max ($12.26), there is no need to shade the bid, so λ = 1.00, and spend stops short of the $400 budget.
• Budget-constrained: cost cap above the cost max. The cost cap dictates a bid that would spend more than the budget, so pacing shades the bid. In Figure 4, when the cost cap is above the cost max ($12.26), λ < 1.00, which brings the bid back down to bid max ($30.61), a bid that spends the entire budget. Note that in this regime, the cost cap has no effect at all. The outcome is exactly the same as Automatic bidding on the same budget.

So the cost cap behaves exactly as the bid cap did, and for the same reason: it's a ceiling on the same dial, just from a different direction. Set it below what the budget would have produced and it binds, leaving budget unspent. Set it above and it's idle, because pacing, not your cap, decides the bid. Only one of the two can be active at a time, and raising an idle cap does nothing at all.

Cost cap pins the bid

Set the average-cost ceiling, then follow the dictated bid to its spend.

04

Regimecap-constrainedλ1.00Operative bid$30.61Spend$399.46

Cost cap

$12.26

Section 8

2026: new value strategies, same old method

By 2026, Meta had retired Target Cost, renamed the survivors, and added two new "value-optimized" bidding strategies named Highest Value and ROAS Goal. Meta hasn't revealed how they work, but they are almost certainly enabled by a new ability to predict both the probability of a conversion as well as the value of the conversion itself. In the straightforward case of a purchase, the predicted value is the predicted order value. Your bid, or your desired conversion cost, is then some multiple zeta (ζ) of that value: bid = value × ζ.

The underlying mechanism barely changes. Before, the score was p × bid, and Meta predicted p for each auction. Now, the score is p × bid = p × value × ζ, and Meta predicts both p and value for each auction. Before, for the Highest Volume strategy, Meta chose the bid max that spent the entire budget. Now, for the Highest Value strategy, Meta chooses the ζ max that spends the entire budget, which can be seen in Figure 5 (bottom).

As a bit of a speculative aside, Meta could create a sixth bidding strategy to complete the symmetry: Highest Value with ζ Cap, or just ζ Cap, which would be the counterpart to Bid Cap. But, capping ζ would simply be too abstract to present in a user interface. Nevertheless, you could set a ζ cap, just as you set a bid cap, which is exactly what the last strategy does, indirectly.

That leaves the ROAS Goal strategy. Figure 5 (top) shows how ROAS and ζ are analogous to average cost and bid, as shown in Figure 4. The difference is that ROAS moves down as ζ, and spend, move up. Consequently, there is a ROAS min that spends the entire budget, and it occurs at the same point as the ζ max. And just as Cost Cap placed a cap on increasing average cost, ROAS Goal places a floor on decreasing ROAS.

A better name for Cost per Result Goal would be Average Cost Cap, and a better name for ROAS Goal would be ROAS Floor.

Figure 5 shows the ROAS Goal strategy in one picture. The ROAS floor dictates ζ, which in turn dictates spend. Yet again, there are two regimes, depending on where the ROAS floor is set relative to the ROAS min.

• Floor-constrained: ROAS floor at or above the ROAS min. The ROAS floor sets ζ outright. The higher the ROAS floor, the lower the ζ, and the lower the score, meaning fewer auctions won, and less of the budget spent.
• Budget-constrained: ROAS floor below the ROAS min. The ROAS floor dictates a ζ that would spend more than the budget. Shading is employed to scale ζ back down to ζ max, and spend the full budget. In this regime, the strategy behaves exactly like Highest Value, with no ROAS floor at all.

ROAS floor pins ζ

Move the floor to see when efficiency or budget becomes the active constraint.

05

Regimefloor-constrainedλ1.00Operative ζ1.54Spend$399.86

ROAS floor

1.60

Highest Value (no ROAS floor) is analogous to Highest Volume (no bid cap), while ROAS Goal is analogous to Cost per Result Goal.

The value strategies use the same techniques as the volume strategies. And if every conversion were worth the same, then predicting value would degenerate to predicting conversion probability, which would make them not just analogous, but identical.

Section 9

Putting it all together

History and simulation suggest that Meta's bidding strategies are different facets of the same machine. The following table summarizes them. Target Cost was retired, probably because it used a different mechanism.

† cost max < cost cap and bid max < bid cap are the same threshold (see Section 7).

‡ ROAS min > ROAS floor and ζ max < ζ cap are the same threshold (see Section 8).

Meta predicts, you set, and one number, bid or ζ, is scaled, which determines the number of auctions won and the total spent. And each of the strategies can be made to behave like another: a cost cap can be chosen that mimics a particular bid cap, and either can be set high enough that the strategy behaves as uncapped (Highest Volume). Similarly, a ROAS floor can be set low enough that it behaves as unfloored (Highest Value), and if conversion values are constant, both strategies look like their older cousins.

Furthermore, this exercise has yielded several interesting insights:

• It's not about the highest conversion probability (or value). Auctions are won on probability or value relative to the competition.
• There is always a single bid max (or ζ max) that spends the entire budget. Real-time pacing has no foreknowledge, but it hunts for an optimum that exists.
• Pacing moves every score together. One variable (λ) shades the bid (or ζ) down, and every score moves with it. No auction is treated specially.
• Everything is related. Spend, wins, and the average conversion cost all rise with the bid; ROAS falls as ζ rises. Fix any one of bid, average cost, or spend, and the other two follow; the same holds for ζ, ROAS, and spend.
• Only one of your two numbers dictates spend. A budget does nothing when the cap is too low or the floor too high. A cap or floor does nothing when the budget is too low.

Section 10

Conclusion

A decade of Meta bidding strategies, including introductions, renames, and a retirement, has resulted in five related ways to control the same underlying machine. Meta predicts a probability (and sometimes a value) for every auction slot, multiplies it by a number, and moves that number until either your budget or your cap (or floor) stops it. The strategies differ only in which quantities are predicted and which constraints you set.

That simplicity is good news, because it clarifies what an advertiser controls:

• Choose the measure. Volume strategies predict, and optimize for, the total number of conversion events; value strategies predict, and optimize for, the total amount of conversion revenue.
• Choose the constraints. Set the budget, and either spend it all, or introduce a cap (or floor) to protect efficiency, possibly at the expense of the budget.
• Feed it the right number. Every path through the machine starts from the same input: what you are willing to pay for a conversion. The mechanism is Meta's; the decision is yours.

That last input is the hard part, because the worth of a conversion is not a constant, and neither is the worth of a visitor, a cart, or a checkout. Pricing those stages, live, from first-party data is the subject of our companion paper, "Pricing the Funnel with Marginal Stage Value (MSV)." This paper explains the machine; that one explains what to feed it.

---

About the simulations. Figures are generated directly from the simulation in the companion notebook: two ads competing across 1,000 auction slots in one simulated day, conversion probabilities (and values) drawn from a long-tailed (Beta) distribution, second-price settlement (the winner pays the competitor's score plus one cent), a competitor holding a fixed bid (a fixed ζ of 1 in Figure 5), and perfect foreknowledge of the day. Conversions are counted in expectation, as the sum of p across the slots won, so every cost quoted is a cost per expected conversion. Real delivery systems know less and face more bidders, but the mechanisms illustrated (a unique bid max, monotone curves, one constraint binding at a time) are structural, not artifacts of the setup. Strategy descriptions and the 2017 "Facebook Bid Strategy" quotes are drawn from Meta's advertiser documentation across the years discussed.