The reason this needs stating is that every platform reports the metric that flatters it. Advertising platforms report impressions and clicks. Social platforms report reach and engagement. Analytics reports sessions. None of those are the business outcome, and a channel can improve on all of them while producing nothing.
Set up the measurement before spending anything, because this cannot be reconstructed. Conversion tracking configured and verified, and a record of enquiry sources collected manually. A month of activity without either is a month you cannot learn from, and the second month repeats the first because nothing distinguished them.
Give each channel a fair timescale rather than a common one, since they work on genuinely different horizons. Paid search can show something within weeks. Content and search visibility take months and then continue without further spend. Judging the second by the timescale of the first is the most common reason work gets abandoned shortly before it would have started producing.
Decide the threshold in advance, which is what separates a test from an ongoing expense. Write down what result would justify continuing, what would mean stopping, and the date you will decide. Businesses quit at three weeks when the test needed eight, or continue for a year because stopping would mean admitting the spending was wasted.
Compare against cost per acquired customer rather than cost per click or per lead, and include your own time where it is significant. A channel that appears expensive per click can be cheaper per customer, and one that produces leads cheaply can be the most expensive of all if none of them convert.
Watch for the channel that produces poor quality rather than low volume, since that is the failure people misdiagnose. Enquiries arriving that are consistently wrong for you indicate a targeting or messaging problem rather than a channel problem, and switching channels carries the same problem to a new place.
Be careful about attributing everything to the last thing somebody clicked, which is what most tools do by default. Somebody who read three articles, followed you for a month, and then searched your name arrives as a direct visit, and the work that actually produced them receives no credit. That is why the manual record matters alongside the automated one.
Then check whether the answer changed rather than treating a conclusion as permanent. A channel that worked at one scale or in one season frequently behaves differently later, and a quarterly review against the same four questions is enough to notice before a year of spending has followed a stale finding.
Compare against what you would have got anyway, which is the comparison nobody makes. Some enquiries would have arrived regardless of a campaign, and a channel credited with those is being overstated. Pausing something for a period and watching what happens is the only reliable way to establish the difference.
Look at the quality of what arrived alongside the count, since a channel producing enquiries that never convert is costing you time as well as money. Recording what happened to each enquiry, not only that it arrived, is what makes that visible.
Account for the seasonal pattern before concluding anything, since a channel assessed across a naturally busy or quiet period will look better or worse than it is. Comparing against the same period last year removes most of that distortion once you have the history.