Your customer acquisition cost is judged against the wrong number

The standard way to judge an acquisition campaign is revenue from the first order, divided by what it cost to get it. It is the number the ad platform reports, so it is the number that gets used. The question is how good it is at telling you who you just acquired. We took two years of orders from a real retailer — 5,852 identified customers, every invoice — and checked.

Value is concentrated — more than most plans assume

Rank customers by what they spent over the two years and the top tenth produced 64% of all revenue. The top fifth produced 77%. The bottom half, 6%. This is a gift retailer with some wholesale buyers, so the concentration is at the sharp end of what a consumer shop would see, but the shape — a small group carrying most of the value — is the normal one, and it changes what "a good customer" means. Average customer value is a number almost no customer has: the mean two-year value was £2,969; the median was £880.

The first order is a small part of the story

Across all customers, the first order was 14% of what they went on to spend. The median customer's first order was £293; the median customer's two-year value was £880. Seventy-two percent of customers ordered more than once. A campaign judged on first-order revenue is being judged on roughly a seventh of what it actually bought.

That would be fine if the first order at least ranked customers correctly — if the big first orders were the big customers. The right-hand chart shows they partly are: the top decile by first order did end up worth the most, with 36% of revenue. The first order carries real signal. But the concentration nearly halves compared with the truth on the left, because most eventual top customers did not announce themselves.

One in three

The cleanest way to put it: take the 586 customers who ended up in the top 10% by two-year value, and ask how many were in the top 10% by first-order value. The answer is 196 — one in three. The other two looked average on day one, and an acquisition process optimising for first-order revenue would have treated them accordingly: bid less for people like them, spend less to retain them, report them as mediocre.

Customers ranked by…	Top decile's share of revenue	Avg. orders in top decile	Repeat rate in top decile
Two-year value (the truth)	64.0%	26.8	99%
First-order value (what campaigns see)	36.2%	10.0	80%

What this means for your shop

The number used to judge acquisition — first-order ROAS — is not wrong, it is early. It is measured before most of a customer's value exists, and it identifies a minority of the customers that matter. The consequence is not academic: bidding algorithms optimise for whatever value they are given. Give them first-order revenue and they will find you customers who place large first orders, which is a different population from customers who stay.

• Know your concentration. The share of revenue from your top 10% of customers is a single query on order history. If it is anywhere near the numbers above, "average customer value" is not a planning figure.
• Measure campaigns on a longer window. Even a 90-day or 180-day value per acquired customer, looked at retrospectively, is a better judge than the first order. It costs nothing but patience and a warehouse that keeps the history.
• Predict, then act. Lifetime value is only useful for bidding if you can estimate it early — from the first order, the first weeks, and what you know about the customer. That prediction, and how to hand it to Google Ads so it bids on the right people, is parts 2 and 3.

Under the hood — for the technically curious

The source is a static Excel workbook, loaded once into BigQuery and declared as a source to Dataform. Staging removes cancellations (invoices prefixed C), returns (negative quantities), zero prices, non-product stock codes (postage, manual adjustments, bank charges) and exact duplicate rows — about 20% of the raw lines. Unidentified customers (23% of lines, mostly guest checkouts) are kept for revenue totals and excluded from customer-level analysis. From there: an order-grain fact and a customer dimension with first/last order, order count, revenue and first-order revenue, plus the frequency/recency/age inputs that part 2's BG/NBD model needs.

The concentration numbers are one query on that dimension:

WITH c AS (
  SELECT customer_id, revenue_gbp, first_order_revenue_gbp,
    NTILE(10) OVER (ORDER BY revenue_gbp DESC)             AS value_decile,
    NTILE(10) OVER (ORDER BY first_order_revenue_gbp DESC) AS first_order_decile
  FROM dim_uci_customers
),
tot AS (SELECT SUM(revenue_gbp) AS total FROM c)
SELECT value_decile,
  COUNT(*)                                   AS customers,
  ROUND(SUM(revenue_gbp) / t.total * 100, 1) AS pct_of_revenue
FROM c CROSS JOIN tot t
GROUP BY value_decile, t.total
ORDER BY value_decile

And the one-in-three:

WITH c AS (
  SELECT
    PERCENT_RANK() OVER (ORDER BY revenue_gbp DESC)             < 0.1 AS top_by_lifetime,
    PERCENT_RANK() OVER (ORDER BY first_order_revenue_gbp DESC) < 0.1 AS top_by_first_order
  FROM dim_uci_customers
)
SELECT
  COUNTIF(top_by_lifetime)                        AS top_customers,
  COUNTIF(top_by_lifetime AND top_by_first_order) AS visible_from_first_order
FROM c

The same two queries run unchanged on any order table with a customer key, an order date and a revenue column — which is the point. This is not a model. It is a question your warehouse can already answer.

Data and licence

Chen, D. (2012). Online Retail II [Dataset]. UCI Machine Learning Repository, doi:10.24432/C5CG6D, licensed CC BY 4.0. All transactions of a UK-based, non-store online retailer of unique all-occasion giftware, 1 December 2009 to 9 December 2011; many customers are wholesalers. 1.07 million lines, 53,600 invoices, 5,942 identified customers before cleansing.