Meta's ad auction selects winners using a total value score — not simply the highest bid. The score multiplies an advertiser's bid by a machine-learning-estimated action rate (how likely a specific person is to take the desired action), then adds an ad quality score. Ads with lower bids often win if Meta's models predict stronger user response.
Meta runs an ad auction for every ad impression across Facebook, Instagram, and its other surfaces. The winner is determined by total value score, not by bid alone. The claims below come from Meta's official business newsroom, where Meta explained how machine learning drives its ad delivery system.
Meta selects the top ads to show a person based on which ads have the highest total value score — a combination of advertiser value and ad quality.
Meta, 2019Meta calculates advertiser value by multiplying an ad's bid by the estimated action rate — an estimate of how likely that particular person is to take the advertiser's desired action.
Meta, 2019Ads with the highest bid do not always win the Meta auction — lower bids often win when Meta's models predict a person is more likely to respond to them.
Meta, 2019Meta uses machine learning to generate both the estimated action rate and the ad quality score used in the total value equation.
Meta, 2019When Meta determines which ads to show someone, it first gathers all ads that include that person in the advertiser's chosen audience before moving them to the auction stage.
Source: Meta, Good Questions, Real Answers: How Does Facebook Use Machine Learning to Deliver Ads? (2019)
Meta selects the top ads to show a person based on which ads have the highest total value score — a combination of advertiser value and ad quality.
Source: Meta, Good Questions, Real Answers: How Does Facebook Use Machine Learning to Deliver Ads? (2019)
Meta calculates advertiser value by multiplying an ad's bid by the estimated action rate — an estimate of how likely that particular person is to take the advertiser's desired action.
Source: Meta, Good Questions, Real Answers: How Does Facebook Use Machine Learning to Deliver Ads? (2019)
Ads with the highest bid do not always win the Meta auction — lower bids often win when Meta's models predict a person is more likely to respond to them.
Source: Meta, Good Questions, Real Answers: How Does Facebook Use Machine Learning to Deliver Ads? (2019)
Meta advertisers choose their target audience through self-service tools, building audiences based on categories like age and gender plus actions people take on Meta's apps.
Source: Meta, Good Questions, Real Answers: How Does Facebook Use Machine Learning to Deliver Ads? (2019)
Meta uses machine learning to generate both the estimated action rate and the ad quality score used in the total value equation.
Source: Meta, Good Questions, Real Answers: How Does Facebook Use Machine Learning to Deliver Ads? (2019)
Meta's machine learning models consider both on-platform behaviours (clicks, likes) and off-platform behaviours (website visits, purchases, app installs) to predict a person's likelihood of responding to an ad.
Source: Meta, Good Questions, Real Answers: How Does Facebook Use Machine Learning to Deliver Ads? (2019)
Meta scores ad quality using machine learning models that consider the feedback of people who view or hide an ad, as well as assessments of low-quality ad attributes.
Source: Meta, Good Questions, Real Answers: How Does Facebook Use Machine Learning to Deliver Ads? (2019)
No. Meta explicitly states: "Ads with the highest bid don't always win the auction. Ads with lower bids often win if our system predicts a person is more likely to respond to them."
Meta states: "Facebook selects the top ads to show to a person based on which ads have the highest total value score — a combination of advertiser value and ad quality." Advertiser value equals bid × estimated action rate.
Meta uses machine learning: "Machine learning models predict a particular person's likelihood of taking the advertiser's desired action, based on the business objective the advertiser selects."
Meta's models consider both on-Facebook behaviors (clicks, likes) and off-Facebook behaviors (website visits, purchases, app installs) to estimate how likely a person is to respond to a specific ad.
Meta states: "Machine learning models consider the feedback of people viewing or hiding the ad, as well as assessments of low-quality attributes" to generate the ad quality component of the total value score.