How do we build the predicted customer lifetime value we send to Meta, and how do we check it is good enough to optimize on?
Our answer
Predict each customer's value from your own history. Confirm on past customers that it outranks first-order value and gives at least 5 distinct values, the highest at least 3 times the lowest. Then test it beside your value campaign.
- Confidence
- Medium
- Strongest proof
- 2 Meta rules
- 1 dataset, 6 practitioners, 3 opinions or claims
- Sources
- 8 sources
- 12 statements
M
EWLE
New here? Take the 1-minute tour
Why this is the answer
3 reasons and our judgmentMeta sets entry rules for predicted lifetime value but no accuracy bar. Its guide requires a model validated in some way, at least 5 distinct positive values with the highest at least 3 times the lowest, and 100 Meta-attributed conversions a week for 4 weeks.
A prediction must beat ranking customers by first-order value (Wang, Liu and Miao, Google researchers). Exactius, a predictive LTV modeling agency, scores past customers at day 30 against their value at day 180 and 365, and calls a rank correlation above 0.5 useful.
Angler AI's own models landed close to actual value. In a held-out check, 92 of 100 new-customer buckets fell within 10%. Webtopia, an ad agency, tests predicted lifetime value (pLTV) for 2 to 3 weeks and judges cohorts over 60 to 90 days.
Our judgment
- Predict contribution margin rather than revenue when you already compute each order's profit. Then the pLTV, the profit value and your margin target share one definition. Angler AI lists contribution margin among the values to predict, with costs as of the order date (Sending profit instead of revenue for costs and The business metric that governs spend for the margin target).
- Treat a model as ready to test when it passes three checks on a held-out past cohort scored as of the first purchase. It ranks customers better than first-order value alone, its rank correlation with actual value is above 0.5, and its predicted value by band lands close to actual. Exactius calls a Spearman correlation above 0.7 strong, and Meta sets no threshold of its own.
- With short history, start from averages that differ by segment, not one average for everyone. Segment by acquisition channel, first product or first-order discount, check against Meta's 5-value and 3 times spread rule, and move to a per-customer model once the data allows, since Angler AI says an average carries no signal. Exactius suggests cohort averages below 10,000 customers, and in Adapty's case graded values by age group shifted delivery for one app.
- If the model needs time, send the value through AppendValue within Meta's 7-day window rather than building for minutes. Meta's own rule ranks above a vendor's advice.
- Run the Meta test only after the offline check passes, as a mirrored campaign beside your current value campaign. Judge each side's cohorts on later margin and repeat rate over 60 to 90 days, or six months for a subscription (Optimizing a subscription or repeat business), not on reported ROAS, since Meta says ROAS can't be compared across optimization settings. Webtopia runs 2 to 3 weeks of stable spend.
Our judgment, built from Angler AI, Exactius, Wang, Liu and Miao, Adapty, Aimerce, Meta and Webtopia: our reading of these sources, not a finding from any one of them.
What the sources say
12 statements from 8 sourcesIn one subscription app's uncontrolled before-and-after, valuing under-25 trials at 20% cut their spend share from about 25% to under 10%, and paid subscriptions from users 35 and older rose 3x.
Adapty says Meta's predicted lifetime value bidding needs enough conversions and a real spread in predicted customer values, with 30, 90, 180 or 365 day horizons. It gives no lift data.
“pLTV needs enough conversion data and enough variation between predicted customer values for the optimization system to learn effectively.”
Angler AI's own held-out checks of its predicted-value models: in 877,735 sessions the top band converted 12.6% against 13.4% predicted. Returning-customer predictions fell within 5% in 97 of 100 buckets.
“The charts below are from held-out data with no overlap with the training period.”
Angler AI, which sells the models, says predicted lifetime value used for bidding must be per customer, forward-looking and delivered within minutes. It puts activation on Meta at about two to three weeks.
“A cohort average tells the auction that every customer acquired in March is worth $180. That is not a signal, it is a constant”
Elevar, a tracking vendor, describes selling a predicted lifetime value score built from industry data or, on its Tailored Start, a store's own order history. It gives no accuracy figures or results.
“a real-time Predicted Lifetime Value score”
Aimerce, a tracking vendor, says Meta has stated that sending predicted lifetime value faster does not improve performance. It gives no method or data.
“Meta has stated that sending pLTV data faster doesn't actually improve performance”
Webtopia, which sells pLTV setup, says skip pLTV if repeat purchases are negligible, volume is too low to learn from, or the LTV model is guesswork, since a confident wrong prediction sends delivery to the wrong people.
“Feeding a confident but wrong prediction into a bidding system does not produce neutral results, it produces efficient delivery towards the wrong people.”
Webtopia advises testing predicted-LTV bidding against standard value optimization for 2 to 3 weeks of stable spend, and judging it on cohort repeat rate and early LTV over 60 to 90 days, not platform ROAS.
“keep your existing value optimisation running as a control, launch a mirrored campaign using pLTV signals, and give both at least two to three weeks of stable spend before judging anything.”
Meta's predicted lifetime value guide needs 100 Meta-attributed conversions a week for 4 weeks and at least 5 distinct values, and says to allow about 2 weeks of training and compare against a baseline.
“Comparing ROAS across different optimization settings is not accurate”
Meta's pLTV guide lets advertisers optimize to their own predicted lifetime values. It needs a web Conversions API setup, at least 5 distinct positive values, and 100+ Meta-attributed conversions a week for 4 weeks.
“a minimum of 100 conversion events per week attributed to Meta for each of the last 4 weeks.”
Exactius says judge a predicted lifetime value model by rank correlation with actual value: above 0.5 is useful, above 0.7 strong. It suggests about 2 years of data and 10,000 to 20,000 customers.
“Spearman rank correlation between predicted and actual LTV above 0.5 indicates useful predictive power”
Three Google-employed authors, in an academic study on public data with no ad test, recommend the normalized Gini and decile charts to evaluate a lifetime value model, with a first-purchase Gini as a lower bound.
“recommend the normalized Gini coefficient to quantify model discrimination”
In the full answer
For readers who sign in- What to do this week3 steps
- The numbers behind it3 figures from cited sources
- Where each source standsA chart placing each source on the question
- The strongest disagreementThe best case against the answer, from a cited source
- How sure we areWhat its confidence rests on, and what would change the answer