Most accounts optimise toward the wrong number. The platform reports revenue at the moment of purchase, the bidding algorithm learns from that number, and the account gradually shifts budget toward whichever audience buys cheaply once and never returns. It looks efficient in the dashboard and it slowly starves the business of good customers.
Bidding to lifetime value fixes the incentive. Instead of telling Google and Meta what an order was worth today, you tell them what that customer is likely to be worth over twelve months — and the algorithm goes looking for more of the profitable kind.
Why first-touch revenue misleads
Consider two customers acquired in the same week. Customer A buys a £45 starter product on a discount code and never returns. Customer B buys a £45 product at full price and reorders four times over the year, spending £310 in total.
To the platform, those two conversions are identical. Both report £45. Bidding optimises for the cheaper of the two to acquire, which is almost always Customer A, because discount-led audiences convert faster and at lower cost per acquisition. Six months later your blended return on ad spend looks stable, your repeat purchase rate has quietly fallen, and nobody can point to the decision that caused it.
The same pattern is worse in B2B, where a form fill from a student researching a term paper and a form fill from a qualified buyer both report as one lead. Optimise to lead volume for a quarter and you will have taught the algorithm to find students.
What you need before you start
This work has prerequisites, and skipping them produces confident nonsense.
Clean conversion tracking. If your conversion counts and your order records disagree by more than a few percent, fix that first. Everything downstream inherits the error.
A customer identifier that survives the funnel. You need to join ad platform click IDs to orders and then to subsequent orders. In practice that means storing gclid, gbraid, wbraid and fbclid at first touch, persisting them to the customer record, and keeping them.
At least twelve months of purchase history, or enough for a genuine cohort. Predicting lifetime value from three months of data mostly predicts your recent promotions.
A margin figure per product or per segment. Revenue-based bidding still optimises for turnover; margin-based bidding optimises for the business.
Building a usable LTV model
You do not need a data science team. Start with the simplest model that beats the status quo and improve it.
Level one — historic cohort value. Group customers by acquisition channel, campaign and first product, then calculate average 12-month revenue and contribution margin per group. This alone is often enough to reveal that one campaign produces customers worth double another at the same cost per acquisition.
Level two — early-signal prediction. Identify behaviours in the first 30 days that correlate with high 12-month value: second purchase within 30 days, subscription signup, category purchased, average order value band, acquisition discount used or not. Assign each new customer a predicted value from the historic average of their matching group.
Level three — modelled probability. A gradient-boosted model or a standard buy-till-you-die model produces a per-customer prediction. Worth doing above roughly 50,000 customers; below that the added accuracy rarely changes a bidding decision.
Whichever level you use, sanity check it the boring way: hold out a cohort, compare predicted to actual after six months, and report the error. A model nobody has validated is a rumour.
Feeding the value back
Google and Meta both accept offline conversion data. The mechanics matter more than the theory here.
Google Ads. Upload offline conversion adjustments keyed to gclid, or use enhanced conversions for leads with hashed email. Send a conversion at the point of purchase with the actual order value, then send an adjustment 30 days later restating the value as the predicted 12-month contribution margin. Use Target ROAS bidding against that adjusted value, not the original.
Meta. Use the Conversions API to send purchase events with a custom value, and value rules where segments differ predictably. Meta’s learning phase is short, so the 30-day adjustment window is less useful — send predicted value at purchase time using the level-two early-signal groups instead.
Timing is the trade-off. Adjust too early and the prediction is weak; adjust too late and the algorithm has already spent a month learning from the wrong number. Thirty days is the usual compromise, shorter for high-frequency categories and longer for considered purchases.
What changes in the account
Expect the structure to shift, not just the numbers.
Budget moves toward higher-value audiences, which typically means a higher cost per acquisition and a lower in-platform return on ad spend. This is the point, and it is where most programmes get cancelled by someone reading only the platform dashboard. Agree the new success metric with finance before you start.
Discount-led campaigns shrink. Once discount-acquired customers are valued at their real contribution, the algorithm stops chasing them. Some of those campaigns still deserve to run, at a smaller budget, for the customers they genuinely start.
Prospecting gets more room. Retargeting flatters first-touch reporting and usually looks worse under value-based measurement, because those customers were often going to buy anyway. Correctly valued, prospecting frequently deserves more budget than it had.
Creative testing changes target. You stop testing for click-through rate and start testing which concepts bring in the customer types your value model rewards.
A realistic timeline
- Weeks 1–3: audit tracking, persist click IDs, join ad data to order history, and build the level-one cohort table. Present the finding that some campaigns produce customers worth twice others — this is what funds the rest of the project.
- Weeks 4–6: build the level-two prediction, validate against a held-out cohort, and agree the reporting metric with finance.
- Weeks 7–8: implement offline conversion uploads and Conversions API values in a single campaign group as a pilot.
- Weeks 9–14: run the pilot without structural changes. Compare blended contribution margin and new-customer quality against the control group.
- Month 4 onward: roll out to the rest of the account, and rebuild reporting around predicted value.
The reporting you need afterwards
Three numbers, reviewed weekly:
- Blended contribution margin after media — total gross profit minus total ad spend, ignoring platform attribution entirely.
- New customer cost of acquisition versus predicted 12-month value, by channel and campaign.
- Predicted-to-actual variance on maturing cohorts, so the model stays honest.
Platform-reported ROAS stays on the dashboard as a diagnostic for in-channel decisions. It stops being the number anyone is judged on.
The uncomfortable part
Value-based bidding usually makes your headline metrics look worse for the first quarter and the business better from the second. Nobody enjoys presenting that. The way through is to agree, before the pilot starts, exactly which number defines success and who signs off on it — because the alternative is a well-built system switched off in week five by someone who only saw the ROAS column go down.