How does the Meta ad auction work?

Meta's ad delivery system combines an ad auction with machine learning to decide where, when and to whom it shows an ad. Each time there is an opportunity to show an ad to someone, an auction runs to pick which ad that person sees. Meta says billions of these auctions take place every day across Facebook and its other technologies.

An ad can compete only if the person belongs to its target audience. Meta's own example uses two advertisers, one targeting women who like skiing and one targeting all skiers who live in California. A woman who skis and lives in California falls into both audiences, so both ads are eligible for her impressions. When several ads from the same advertiser are eligible for one auction, Meta enters only the one with the highest total value, so an advertiser's own ads do not bid against each other.

Meta also has to cut the field down before any ad wins. Its engineering team says retrieval narrows tens of millions of candidate ads to a few thousand, and larger ranking models then choose the final ads. Meta's help centre describes that final choice as an auction on total value. The article on what Meta Andromeda is covers the retrieval stage. This article covers what decides the winner.

Meta's help centre says the winning ad is the one with the highest total value, subject to a price floor that can affect whether an ad shows and the price paid. Total value is a combination of three factors.

Meta says estimated action rates and ad quality together measure ad relevance. Both come from Meta's predictions, so the advertiser influences them only through the ad, the offer and the data it sends. The bid is the part the advertiser sets, and most bid strategies still let Meta choose the actual amount in each auction.

Meta's current help page lists the three factors but does not publish a formula. A 2020 Meta explainer was more specific. It said the system multiplies an ad's bid by its estimated action rate to get advertiser value, then adds an ad quality score to reach total value. Meta has not published the scale or weight of each term, so treat that formula as a description of the logic.

Meta's bid cap documentation shows the same multiplication from the bid side. Within the auction, Meta calculates a bid per impression as the bid per estimated action multiplied by the estimated action rate. An advertiser bids for a purchase or a lead, and Meta converts that into the value of one impression by multiplying it by the chance that this impression produces the action.

A made-up example shows the effect. Ad A bids 20 for each purchase, and Meta predicts a 1 in 1,000 chance that this person buys, so one impression is worth 0.02. Ad B bids 10 for each purchase with a predicted 1 in 400 chance, so one impression is worth 0.025. Ad B leads on advertiser value with half the bid, before Meta adds ad quality.

The current help page adds two details. A price floor, or minimum price, can stop an ad from showing and can change the price paid. Meta may also adjust auctions, for example to improve ad quality and relevance, test ad products or comply with the law. Meta says those adjustments will not make it charge more than the bid.

Meta states this directly. Because estimated action rates and ad quality are part of the auction, an ad that is more relevant to a person can win against ads with higher bids. The 2020 explainer said lower bids often win when the system predicts a person is more likely to respond or finds the ad is higher quality.

Relevance changes cost. Meta says higher-quality ads perform better in the auction and that ads with a lower quality ranking tend to cost more, which can reduce their delivery and lead to fewer results. If an advertiser repeatedly posts policy-violating or lower-quality ads, Meta's systems may start treating all ads from that Page, domain or ad account as lower quality.

Meta lists engagement bait among the low-quality attributes it assesses, and its auction page says clickbait and engagement bait do not improve ad performance. Engagement is useful to an advertiser only when it predicts the outcome the ad set optimises for, because the estimated action rate is about that outcome.

A bid strategy tells Meta how to bid on the advertiser's behalf in each auction while respecting its spend and cost goals. Meta groups the strategies into three types.

• Spend-based bidding. Highest volume aims to get the most results possible from the budget, and highest value aims to spend the budget on the highest-value purchases. Meta does not optimise for cost per action here, so costs rise and fall with auction competition.

• Goal-based bidding. A cost per result goal keeps the average cost near a target, and a ROAS goal keeps return on ad spend near a target. Meta does not guarantee either, and it recommends 50 to 100 or more conversions a week for best performance.

• Manual bidding. A bid cap sets the maximum bid per estimated action in every auction. Meta may bid lower if a higher bid would spend the budget too quickly, and it says bid cap suits advertisers who understand their predicted conversion rates well enough to calculate the right bid.

Every strategy enters the same auction. A strategy works on the bid term, while the other two terms come from Meta's predictions about the person and the ad. A tight cost goal or bid cap can also hold an ad set in the learning phase, because Meta lists a low bid or cost control as one cause of the Learning limited status.

Meta replaced its old relevance score with ad relevance diagnostics. Ads Manager reports three rankings, each based on how the ad performed in past auctions over the dates selected.

• Quality ranking compares the ad's perceived quality with ads competing for the same audience.

• Engagement rate ranking compares the ad's expected engagement rate with ads competing for the same audience.

• Conversion rate ranking compares the ad's expected conversion rate with ads that have the same optimisation goal and compete for the same audience.

Meta points to quality ranking as the way to measure ad quality against competing ads. The engagement and conversion rate rankings describe expected rates, the same kind of prediction that sits behind estimated action rates. Meta is clear that the diagnostics are not inputs into the auction, and they only appear once an ad has 500 impressions.

Meta's guide to the diagnostics reads them as a set. Average quality and engagement with a below-average conversion rate ranking means the ad is not producing conversions, so Meta suggests working on the call to action or the post-click experience. A below-average quality ranking alongside a below-average conversion rate ranking suggests the ad looks like clickbait. Meta adds that an ad already meeting its objectives does not need better rankings.