Here is a month that has ended a few careers. Two traffic sources, both of which improved, and a blended number that got worse.
In the first month, paid traffic brought 10,000 visits and 200 conversions, which the conversion rate calculator puts at 2.00 per cent. Organic brought 5,000 visits and 400 conversions, a rate of 8.00 per cent. Blended, that is 600 conversions from 15,000 visits: 4.00 per cent.
In the second month, paid improved to 2.50 per cent, on 375 conversions from 15,000 visits. Organic improved to 10.0 per cent, on 200 conversions from 2,000 visits. Blended: 575 conversions from 17,000 visits, which is 3.38 per cent.
Where the loss came from
Nothing in that story is a measurement error. Both channels genuinely got better, by 25 per cent each in relative terms. Total conversions did fall, from 600 to 575, and total traffic rose, so the blended rate dropped.
What changed is the mix. Organic, which converts four times as well as paid, fell from a third of the traffic to under an eighth of it. The blended rate is a weighted average, and when the weights move toward the weaker segment, the average can fall even while every segment rises.
This is Simpson's paradox, and conversion data is where most people meet it, usually while being asked why the number went down. The honest answer is that the site-wide conversion rate was never a measure of how well the site converts. It is a measure of how well the site converts, multiplied by where the traffic came from, and those two things move independently.
Why it shows up so reliably here
Segments in web analytics differ enormously in conversion rate. Branded search against display, returning against new, mobile against desktop: ratios of four or five to one between segments are routine. Any metric that averages over segments that different is going to be dominated by the mix.
And the mix moves constantly, often as a direct result of doing well. Scaling a paid campaign adds exactly the kind of traffic that converts least, so a successful acquisition push reliably depresses the blended conversion rate. The better the quarter, the worse the number looks.
The same effect runs through click-through rate when the placement mix shifts, through cart abandonment when the device mix does, and through average order value whenever a promotion changes who is buying.
What to do instead
Report rates by segment and report the mix separately. If a single blended figure is needed, keep the weights fixed at a reference period so that mix changes show up as a mix change rather than as a performance change, which is the same technique a price index uses for the same reason.
And when a blended rate moves, check the mix before checking anything else. The question is not what went wrong on the site, it is what changed about who arrived, and that question takes thirty seconds to answer and saves a week of looking in the wrong place.
For the related trap of ignoring how common something is before reasoning about it, see the base rate is the whole story, and mean against median covers the other way an average can describe a population that nobody in it resembles.