Optimizations

Where Do Learnings Take Place?

• Learning takes place at ad level, NOT ad set level.

• Adset level learnings have some effect bunch are much less important

• Learning means predicting click-through rate and post-click conversion rate.

• Post-click conversion rate needs to be at the very least 0.5%, preferably over 1%.

• Each ad must have 25 lifetime conversions for delivery optimization to work.

• Each new ad needs to run for 36 hours until learning takes effect

What Will Reset Ad Learnings?

• Changing bid and budget is OK and will not reset learnings.

• Changing just about anything else (title, message, photo, …) will reset learnings because FB considers this to be a new ad.

• If the ad goes to review process that will definitely mean that learning for that ad starts from scratch.

Account level learnings

• Accounts also collect learnings (used as priors for new ads).

• It is not fair to compare performance against an existing account. All cross-account bake-offs MUST use new accounts.

• New accounts use country/objective level priors.

Delivery issues

• Most common cause of delivery issues:

  • CTR is low

  • or CR is low

• What to do if you do not get delivery:

• • Massively increase bid (or budget if the adset if budget limited).

  • If using autobid increase budget.

  • Relaunching the ad might also help because there is some randomness in getting initial conversions.

Deduping

• Deduplication happens also on object level (page/app).

• If another account shows ads for the same page, only one of the accounts will be able to show an ad for each user.

• This is why you absolutely must use an ad study with bake-offs

AutoBid

• Autobid will keep increasing your bid so that the budget is spent; in theory it should always spend the entire budget.

• Bid can go up to 95th percentile of bids in country/objective; there is no fixed dollar cap otherwise.

• The bid distribution is not updated in real time, so when bids suddenly increase (e.g. on Black Friday) the cap might be much lower than 95th percentile and hence auto-bid fails to spend the budget.

Fun Facts

• 25 conversions is enough to provide roughly 90% accuracy for predicting conversion rates.

• Facebook uses up to 90 days of user level data in optimization. This is based on EU regulations but the same 90 days is commonly used throughout Facebook.

• There is time decay for learning, so last days matter more. This also depends on the amount of data: if you spend over $10k per day and get tons of impressions then older data has less influence.