The lag between watching and buying

Every video is checked against one convention: el=yt-<topic><nnn> with htrafficsource, hcategory and hgoal. A video either matches it exactly or it lands in a named fault class. Nothing counts as fine merely because it looks fine.

Only the failing videos get edited. Each writes its original description to a backup ledger, then takes the corrected link. The ledger below interleaves the two so the order is visible: a backup row always precedes its edit row.

Run the rollout to see a diff.

Sales arrive from four places and none of them writes an email address the same way. Matching lowercases, trims, drops plus-addressing, and ignores dots only where the provider ignores them. Whatever still will not join is shown rather than hidden or forced.

The addresses that needed normalising before they would join:

Six buyers sampled across the whole lag range, from the quick ones to the slowest, each drawn on its own time axis. The shaded band is the attribution window measured from the anchor touch. Change the window or the touch rule in the instrument and watch credit move between them.

Leads, applications, calls booked, sales and revenue per video, under the window and touch rule currently set. Sort any column. This is the view that answers the question the channel actually has.

The console recomputes its own figures from the source rows and asserts them, including the ones that should move when you change the window. If a number above were decorative, one of these would fail.