Lead Source Performance: Rank by Conversion, Not Volume

Why this matters

A source table arrives sorted by lead count, because count is the column your records fill in by themselves. Sorted that way it answers exactly one question: which channel is loudest. The question you opened it for is which channel is working, and those two orderings disagree often enough that reading the wrong one is the normal case, not the exception. The cut that follows is expensive and quiet. The row that gets dropped is usually a small one with a strong conversion rate, because sitting next to a row four times its size it reads as a rounding error.

What the two columns are actually counting

Fix one meaning of converted and hold it for the whole table. Here it means a lead that became a booked job on the schedule, not a lead that got an estimate out the door. Now look at what each column is built from:

• The count column is leads created in the window carrying that source. Created, not converted. A lead created in March and booked in May sits in March's row, and its booking lands in March's numerator.
• The rate column is that row's conversions divided by that row's own count. The denominator is the number sitting beside it, so a row with 9 leads has a rate built on 9 things.
• Leads with no source recorded are not in this table at all. Not as a row, not in the totals, not in the overall rate. They were dropped before the table was drawn.

The third one is the largest of the three problems and it is invisible by construction. Come back to it below.

The same quarter, sorted both ways

One quarter, one shop, every lead that carries a source:

Sorted by leads, the order is search, paid directory, repeat call-in, referral, yard sign, trade partner, social. Sorted by conversion it is trade partner, repeat call-in, referral, yard sign, search, paid directory, social. Search is first of seven by count and fifth of seven by rate. Paid directory is second of seven by count and sixth of seven by rate. Trade partner is sixth of seven by count and first by rate: an eighth of search's leads at 3.4 times search's conversion rate.

An owner reading down the count column sees two rows carrying most of the volume and one row that barely registers. An owner reading down the rate column sees the opposite table.

A rate is not a rank until the row is big enough

Before you act on either ordering, apply one floor: at least 30 leads in the window before a row is ranked at all. Below that you do not rank it, you roll the window forward until it clears.

The reason is arithmetic, not caution. At a conversion rate near a third, 30 leads carries roughly 9 points of ordinary sampling wobble at one standard error, so two rows of that size sitting within 9 points of each other are tied, not ranked. The wobble is smaller at the extremes, nearer 5 or 6 points on a row converting around one in ten, which is why a genuinely terrible row shows itself sooner than a merely mediocre one.

Run the floor against the table above. Three rows clear it: search at 96, paid directory at 61, repeat call-in at 44. Four rows are struck: referral at 22, yard sign at 17, trade partner at 12, social at 9.

Among the three that survive, the inversion holds. The smallest of them converts at 65.9 percent against the largest at 19.8 percent, 3.3 times the rate on fewer than half the leads, 44 against 96. That one is not close and you can act on it. Search at 19.8 percent against paid directory at 13.1 percent is under 7 points on rows of that size, which is not a ranking, it is two rows that need another quarter before you separate them.

Notice what the floor did not do. It did not make the trade partner row go away. It told you the row is a candidate rather than a finding, which is a different instruction and a much cheaper one to follow.

The biggest channel in the table is not in the table

That quarter created 372 leads. The table above holds 261 of them. The other 111, 29.8 percent of every lead created in the quarter, carry no source at all.

Put that number beside the rows. At 111 leads, the blank bucket is larger than search at 96 and nearly twice paid directory at 61. If it were a row it would be the top row. It is not a row, so nobody argues about it, and every rate in the table is computed as though it does not exist.

That matters in two specific ways:

• The overall 31.8 percent is a rate over recorded-source leads only. You do not know the shop's real conversion rate, because 111 leads are outside the denominator and their outcomes are outside the numerator.
• The blank bucket is not a random sample of your leads. The source field gets filled when intake is calm and skipped when it is not, so the blanks skew toward busy days, after-hours calls, and whoever was covering the phone. Those are exactly the conditions under which a channel's leads behave differently.

Fixing this is not a reporting job. It is one field at intake, made mandatory, with a short closed list rather than free text, and an "asked, would not say" option so the person taking the call has something honest to pick instead of leaving it blank. Watch the blank share monthly, and treat any month over about 10 percent of leads created as a table you should not be sorting yet.

Three things that stop two rates being comparable

Conversion rate is only a fair comparison between rows selling the same thing to the same kind of buyer. Three things break that, and all three are live in the table above.

Different buying situations. Repeat call-in converts at 65.9 percent because those people already chose you once. Search converts at 19.8 percent because those people are comparison shopping, often for the first time. Ranking the two against each other reads as "repeat customers are a better channel", which is true and useless, since you cannot buy more of them directly. The useful read is that your highest-converting demand is demand you already earned, and the marketing question becomes what feeds it rather than which of the two wins.

Different work. A trade partner sending commercial change-out work and a social post producing small residential service calls are not on the same scale. If one source's leads are systematically bigger, slower, or more technical, its conversion rate answers a different question than the row above it.

Different definitions of the same word. If "converted" means booked job for the phone rows and signed estimate for the web rows, the table is comparing two constructs under one heading and the difference will look like performance. Pick one, write it at the top of the table, and check it against how each row is actually being recorded.

There is a fourth effect that is not the table's fault: leads created near the end of the window have had less time to convert, so every rate here is understated, and understated most on the sources with the longest decision cycle. That is cohort truncation and a sibling card owns it.

What to do with the row you cannot rank yet

The trade partner row is the one an owner will want to act on immediately, and it is the one the floor just struck. That is not a dead end. There are three moves available and none of them is a budget decision.

Roll the window. Recompute that row over two or three quarters rather than one. If the source runs at 12 leads a quarter, it clears 30 leads somewhere in the third quarter of data. Say so in the record, with the window stated next to the rate, so the next person to read it does not compare a three-quarter row against one-quarter rows.

Ask the qualitative question now. Twelve leads is a small number to rank and a fine number to read. Pull all 12 and look at what they had in common: who sent them, what they were asking for, what the first conversation looked like. Eight of twelve booked. That is a pattern you can learn from even though it is not a rate you can bank.

Make the cheap version of the bet. Growing a partner channel usually costs time, not money. One call a month to the three partners already sending work is an investment you can make before the row clears the floor, and it is reversible in a way that reallocating a paid budget is not.

The failure mode here is the opposite one, and it is common: a spectacular rate on a tiny row gets a budget shifted onto it, the channel scales, and the rate collapses toward the shop's blended number because the original 12 were the easy ones the partner already had. Scaling a channel changes who it sends you. Expect the rate to fall and decide in advance how far it can fall before the bet is wrong.

Sort on the column that answers the decision in front of you

There is no single right ordering. There is a right ordering per question, and most arguments about this table are two people answering different questions from the same rows.

Search runs 96 leads to 19 bookings, about 5.1 leads worked per job won. Trade partner runs 12 to 8, about 1.5. Both of those are counts of leads against counts of jobs in the same quarter, and that ratio is what predicts how much of the office's day a channel consumes, which the conversion rate alone never shows you.

References

• See related: What a Good Lead-Gen Report Actually Tells You, for building the report and the inputs it needs
• See related: Estimate Conversion Rate and the Cohort Problem, for why a rate measured over a short window is structurally understated
• See related: Lost Reasons Are Only as Good as the Field Being Filled, for reading the losses behind a rate
• See related: Handling a Referral Source Whose Leads Don't Fit, for a high-rate channel sending the wrong work