The work · 03

One month of television, read two ways

A subscription brand spending about a million dollars a month across linear and streaming, and a report that could not tell the brand buys from the response buys.

One month of television, read two waysPerformance lines ranked on cost per purchase with a volume gate; brand lines set apart and read on reach rather than four-week cost per purchase.Performance lines · cost per purchase, spend-weightedRetargeting$51Sports network B$54News network$60Prospecting$65Tennis network$74Lifestyle network A$90Home network$99Cable news$102Streaming drama$116Streaming reality$124Comedy network$154Streaming reality B$448Cable reality$470campaign average $118Lines under the volume gate are shown in the workbook but not ranked. Faded bars are the lines proposed for reduction.Brand lines · read on reachHeld, not cutPlayoff tentpole$505 CPABroadcast prime tentpole$221 CPALive sports A$248 CPALive golf event$120 CPAWhat they are judged onReach and cost per unique reachBranded search in tentpole weeksDirect and organic traffic vs baselineA longer window than four weeksFour-week CPA shown for completenessDid efficiency hold as spend moved?Week 1 to week 4, cost per purchaseNews network$29 to $109Lifestyle network A$57 to $176Streaming reality$139 to $129Retargetingheld
Line names are generic. Figures are from the anonymized review. Cheap lines that tighten as they scale are a normal pattern, and the reason the plan favoured the lines that held.

The situation

Eight hundred and fifty-five spots across thirty-nine networks and three kinds of inventory, four weeks, seven creatives, with attributed visits and purchases from the measurement vendor. The standard read ranked every line on cost per purchase, which put a playoff tentpole at the bottom and a handful of very small, very cheap cable lines at the top, and pointed the next month's budget toward the small lines.

What I did

Separated the buy into two kinds of inventory that had been reported as one. Performance lines, most cable, streaming direct and programmatic, ranked on spend-weighted cost per visit and cost per purchase, with a minimum-volume gate so that lines too small to read were shown but not ranked. Brand lines, live sports, tentpoles and broadcast prime, read on reach and on branded search and direct traffic in tentpole weeks, over a longer window than four-week cost per purchase. Then a check the standard report does not run: did each line's efficiency hold when its spend moved week to week? Several of the cheapest lines tightened sharply on modest increases, which is the pattern that makes cheap inventory look better than it can scale to. The next month's plan was written as the whole budget under one proposal with a likely range for cost per purchase, cost per visit and purchases, and a note on what a good month would look like for the brand lines.

What changed

A reallocation of about seven percent of the budget inside the performance lines, with the tentpoles held and instrumented rather than cut, and a plan the client could see the arithmetic of. On the observed costs, blended cost per purchase moves from one hundred eighteen dollars to somewhere between one hundred eight and one hundred fourteen, depending on how much erodes as lines scale.

What I would do differently now

This one is current, so the honest answer is what I already did: the first version of this method, built years earlier, had a composite index with a weighting error and an inverted term in it, and it recommended the tentpole. Rebuilding it is what produced the two-kinds-of-inventory read. Methods should get audited the way budgets do.

Back to the work