Bid Strategy and Pacing Rebuilt from Log-Level Auction Data

Most underperforming display programmes in this sector are not bidding too low. They are bidding into the wrong auctions and pacing so evenly that they miss the hours their audience is at a desk. The first number I ask for is win rate, split by deal and by exchange. Under two per cent you are being shaded out or outbid by someone who values the impression more. Over forty per cent on the open exchange you are usually the only bidder, which means you are buying inventory nobody else wants, which in practice means made-for-advertising sites. What I need before touching a bid: Log-level data. DV360 Data Transfer files or Trade Desk Raw Event Data Stream, landed in BigQuery or anywhere I can query them. The platform interface reports what it charged you; it does not report the auctions you lost, the bids that were shaded, or the deals where your maximum bid sat below the floor and so never entered the auction at all. That last one is the single most common cause of budget sitting unspent in a deal while the report shows healthy pacing elsewhere. What usually changes: Bid rules stop being one number per line. Under first-price auctions the useful work is separating the inventory worth a higher bid from the inventory you are currently paying the same for by accident, which means bid multipliers built from measured post-click behaviour rather than from a media plan's assumptions. Pacing moves off even for anything with a date attached and off ASAP for anything always-on. Dayparting gets rebuilt around when engineers and procurement people actually browse, which in South Africa is a weekday pattern with a hard shoulder, not the evening curve consumer campaigns are tuned to. On load shedding: Evening delivery in this market is unstable in a way the platforms do not model. When the grid is under strain, residential desktop impressions fall away and the inventory that remains skews mobile and cheap. Buying the same bid into that window on the assumption it is comparable inventory is a slow leak, and it is visible in log data within a fortnight. Automation I will not switch on: Target CPA or maximise-conversions bidding on a programme producing fewer than roughly fifty conversions a month. Below that the algorithm is fitting noise and you will spend a quarter learning nothing. In low-volume industrial programmes I build a weighted proxy signal instead, scoring the on-site behaviour that precedes a real enquiry, and bid against that with the platform's custom bidding tooling. It is slower to set up and it is the only version of automated bidding that behaves at this volume. I also will not run bid automation across a set of deals whose floors I have not seen. Automation cannot bid below a floor and will quietly reallocate away from your best inventory to satisfy its own target. Not included: Creative testing, landing page work and anything downstream of the click. Deal negotiation, which is separate work. Attribution modelling; I will tell you what the auction data supports and no more. Not for you if: You cannot get log-level exports enabled, or your agency holds the seat and will not share them. Without that, this engagement is me reading the same dashboard you already have, and you should not pay for that. Also not for you if the programme is under about twenty thousand dollars a month in working media. The gains here are real but proportional, and below that threshold the fee eats them.