Every conversion rate discussion I have sat in went the same way. Someone proposes a change, someone else disagrees, and nobody can settle it because the store emits a single number and no structure underneath it. The fix is not a better opinion, it is better instrumentation, and that part belongs to whoever owns the code.
What The Funnel Should Actually Emit
The minimum useful set of events is product view, add to cart, checkout start, shipping step, payment step, purchase. Six events, each carrying a session id, a device class, a traffic source and a cart value.
That is enough to compute step to step drop off, which is the only view that tells you where to spend effort. A store at 2 percent overall might be losing 70 percent between add to cart and checkout start, which points straight at the cart page, or losing 40 percent at the shipping step, which almost always means the shipping cost surprised somebody.
Without per step data you are optimizing by vibes. With it, the argument ends in about ten minutes.
Averages Hide The Thing You Need To See
Overall conversion rate is a blended number and the blend is doing a lot of damage.
A typical store converts branded search near 4 percent and cold paid social under 1 percent. Averaged, that is a respectable 2 percent, and the average suggests a site wide problem that does not exist. Segment by source, device and new versus returning before you touch anything, because those three splits usually explain most of the variance.
The practical consequence is that the same page can need two different fixes for two different audiences, and a site wide test will show nothing because the two effects cancel.
The Sample Size Conversation Nobody Wants To Have
Here is where engineering teams new to this get burned. On a 2 percent baseline, detecting a realistic relative lift at normal confidence takes far more sessions per variant than most stores see in a two week window.
Run that math before the test rather than after. If the number of sessions you need is larger than the traffic you have, the test cannot answer the question, and running it anyway produces a result that feels like evidence and is not.
The failure mode is not a wrong answer, it is a confident wrong answer that ships and then gets cited for a year.
What To Do When The Traffic Is Not There
Most stores are in exactly that position, and it is fine, it just changes the method.
Ship the changes that are correct on their own merits. Removing a validation rule that rejects legitimate addresses does not need an experiment. Neither does showing shipping cost before the final step, or cutting three hundred kilobytes of tag manager payload off the mobile product page.
Then watch the trend over weeks with your segmented funnel data, and reserve formal testing for the genuinely ambiguous changes, the ones where a reasonable person could argue either direction.
Where The Rest Of It Lives
The page level checklists, the testing framework and the full list of what actually moves the number are in this complete guide to ecommerce conversion rate optimization. This post is the instrumentation half. That one covers the rest.
Top comments (0)