Traffic arithmetic: sessions, conversion and the cost of a customer

Paid traffic is bought by the click and consumed by the order. The bridge between the two is the conversion rate, and it is a multiplier, which is why it dominates.

cost per order = CPC / conversion rate

£0.45 CPC at 1.0% CVR  = £45.00 per order
£0.45 CPC at 2.0% CVR  = £22.50 per order
£0.45 CPC at 3.0% CVR  = £15.00 per order

Against a £18.82 ceiling: the first two lose money on every
order; only the third is viable.

Note the shape of that table. Halving the CPC from £0.45 to £0.225 at 1% conversion gives £22.50 per order — the same as doubling the conversion rate. Both are hard. But the CPC is set in an auction against every other bidder, and the conversion rate is set on a page you control entirely. That asymmetry is why the page usually deserves the attention first.

Is your conversion rate a number or a rumour?

A conversion rate computed on small traffic is mostly noise, and this can be quantified rather than felt. For a proportion p observed over n sessions, the standard error is the square root of p(1−p)/n, and a 95% interval is roughly ±1.96 standard errors.

n =   200 sessions, 4 orders  → p = 2.0%, SE = 0.99%
                                95% CI ≈ 0.1% to 3.9%
n = 2,000 sessions, 40 orders → p = 2.0%, SE = 0.31%
                                95% CI ≈ 1.4% to 2.6%
n = 20,000 sessions           → p = 2.0%, SE = 0.10%
                                95% CI ≈ 1.8% to 2.2%

At 200 sessions your cost per order could be anywhere from about £11 to a figure large enough that it cannot be computed reliably at all. The normal approximation used above is itself unreliable when the expected number of orders is below about ten, which is exactly the regime most new stores are in. The practical consequence: at a few hundred sessions you do not have a conversion rate, you have an anecdote, and decisions taken on it are coin flips wearing a percentage sign.

Sessions needed to see a difference

To detect a change from 2.0% to 2.5% with conventional confidence and power requires, approximately, 16 × p(1−p) / δ² sessions per variant — about 16 × 0.0225 × 0.9775 / 0.005² ≈ 14,000 per side, or 28,000 sessions in total. That figure is the honest reason most small stores cannot A/B test their way to answers, and should instead make changes that are large enough to be visible or justified on grounds other than measurement.

Figures in this lesson are illustrative inputs chosen so the arithmetic can be checked. They are not measurements. What FlowFinds Solutions actually measures is published, with its artifacts, in research.