Teaching Google Ads what a good customer is — and when you can actually know it

Value-based bidding is simple in principle. Your ad platform optimises toward whatever conversion value you give it. Give it first-order revenue and it finds people who place large first orders; give it predicted lifetime value and it finds people who stay. The mechanism is well supported and not technically hard.

The difficulty is timing, and it is worth being precise about it because we got it wrong first.

What the upload windows actually require

Our first draft of this argument claimed Google Ads cannot bid on a value that arrives late. That is not right. Checking Google's own documentation:

So the real constraint is not that bidding ignores late values — it is that the click comes before the order. A customer clicks, buys three days later, and a value computed 90 days after that order reaches an identifier 93 days old. It is not rejected because it is unhelpful; it is rejected because there is nothing left to attach it to. Anything past seven days also disappears from your attribution reports, so the numbers you bid on and the numbers you report on diverge.

Which makes the practical question: how good can a forecast be in the first days, while it can still be uploaded comfortably?

What a forecast is worth, by when it is available

Available on…	Method	Top decile found	Revenue captured
Day 0	Sort by first-order value	46%	37.0%
Day 0	Model on the first order	46%	37.2%
Day 7	Model	48%	37.3%
Day 30	Model	46%	37.6%
Day 90	Sort by 90-day spend	62%	41.2%

With about 82 customers in the test top decile, two percentage points is under two customers. At every moment that fits comfortably inside the upload window, a trained model matched a plain sort. The 16-point jump appears only at 90 days, and it comes from a fact that did not exist on day seven: whether the customer came back.

That is the whole lesson. No model recovers information the business has not generated yet. “Predicted lifetime value” is marketed as a modelling problem; in this dataset it was a waiting problem, and the waiting is exactly what the bidding window cannot accommodate.

What we would recommend now

• For bidding, send first-order value. It is available immediately, it is comfortably inside every window, and in this data a model layered on top of it added nothing. If you want to send a modelled number because the platform needs pounds rather than a rank, that is a fair reason — just do not expect the model itself to improve the targeting.
• For strategy, use the 90-day figure. Budget allocation, segment priorities, which cohorts deserve retention spend. Nobody is bidding on it, so the lag is irrelevant and the extra 16 points of accuracy are free.
• Upload within seven days anyway, so your attribution reports and your bidding see the same world.
• Measure this curve on your own data before committing. Our features were transactional only — order value, basket size, country. A retailer who knows the acquisition channel, the device, the discount used and the browsing that preceded the purchase has far more signal available on day zero, and may well find a model pulls ahead. That is a test worth running before buying anything that promises predicted lifetime value.

Under the hood — for the technically curious

Four BigQuery ML linear models, one per horizon, each given only what was knowable at that point. Trained on 2,135 customers acquired December 2009 to March 2010; tested on 816 acquired April to June, so no customer appears in both. The label is log first-year revenue, because the raw distribution is dominated by a handful of wholesale accounts.

CREATE OR REPLACE MODEL `recsys.uci_clv_d0`
OPTIONS (model_type = 'LINEAR_REG', input_label_cols = ['label'], l2_reg = 1.0) AS
SELECT
  d0_revenue, d0_lines, d0_units, country,     -- everything known at the first order
  LOG(GREATEST(value_365d, 1)) AS label
FROM `recsys.uci_clv_horizons`
WHERE split = 'train'

A boosted tree on the 90-day features reached 63% against the sort's 62%, so the finding is not an artefact of using a linear model. Evaluation is the same throughout: each method ranks the test customers, takes the top 10%, and we count how many of the true top 10% it contains.

Data, licence and sources

Chen, D. (2012). Online Retail II [Dataset]. UCI Machine Learning Repository, doi:10.24432/C5CG6D, licensed CC BY 4.0. Upload limits quoted from Google Ads Help: About attribution reports and Troubleshoot offline conversion imports, both checked 12 October 2026. Google's product rules change; check them against your own account before building on them.