Ad Server Migration With a Dual-Run and Reconciled Cut-Over

The month you switch ad servers is the month your revenue reporting is least trustworthy, and the plan should be built around that rather than around a go-live date somebody promised a board. We run migrations with a dual-run of four to six weeks, in which both servers carry an agreed split of the same inventory, so the difference between them is measured on your own traffic instead of debated in a meeting. Publishers who skip the dual-run generally spend the following quarter unable to say whether the new server is worse or the market was. When to schedule it: Ad Manager builds availability forecasts from delivery history. Cut over in the middle of the quarter your direct book was sold in and your sales team spends the next thirty days quoting availability they cannot check. So the calendar comes from the direct book, not from the contract end date of the old server. If the old contract expires at the wrong moment, the cheaper answer is usually a short extension, and we will say so even though it delays our own invoice. What actually breaks: Object models do not map one to one. A legacy campaign, insertion, creative hierarchy from Equativ, Xandr or Revive has to be rewritten into orders, line items and priority bands, and the priority band is where migrations quietly go wrong, because a legacy weighting that behaved like a guarantee becomes a standard line that loses to house inventory. Passback chains from the old setup keep firing into partners nobody has spoken to in two years, and the migration is the only sensible moment to end them. Third-party tracking has to be reissued: CM360 placements, IAS and DoubleVerify wrappers, and every macro in every VAST tag, since a click macro that survived the copy-paste unencoded is a campaign that reports zero clicks to an advertiser who will notice. Mid-flight direct campaigns: Signed IOs running across the cut-over get individually planned, not batch-imported. Each one gets a delivered-to-date figure agreed with the buyer before the switch, so the invoice at month end has one story and not two. This is slow, manual, and the reason clients call us instead of a scripted import. Reconciliation: We close the cut-over month against advertiser-side numbers with a stated tolerance. Ten percent is the discrepancy the industry tolerates between publisher and advertiser servers; we hold our side under five and document every line that misses, with the cause, rather than presenting a total. The old server stays readable, in export, for thirteen months so year-on-year reporting survives. Rollback: The conditions under which we go back are written and signed before the first line item moves, with named thresholds on fill, revenue per thousand sessions and direct delivery pacing. A rollback plan invented during an incident is not a plan. Not included: Renegotiating your demand contracts. Video player replacement, which people try to bundle into a migration and which doubles the number of things that can fail at once. Order management or billing system changes. Sales team training beyond the ad ops function. We turn down: Migrations chosen to escape a billing dispute, because the dispute follows you. Migrations where the destination was picked before anyone measured the current one. Anyone who needs it live in three weeks, which is achievable and which we will not do, since the corners you cut are the reconciliation and the rollback, and those are the two things you will want in week five.