Every demand-side platform ships with a bidding algorithm. You pick a goal, such as clicks, completed views or conversions at a target cost, and the platform decides how much each impression is worth. For many advertisers that is good enough. The trouble starts when the goal the platform can count is not the thing your business is paid for.
The problem with default goals
A default algorithm treats every conversion as equal. Your finance team does not. A first order from a customer who will buy again for years is worth far more than a discounted one-off. A sale with a healthy margin is worth more than one that barely covers shipping. An algorithm that cannot see those differences will happily buy the cheap conversions and call it a good week.
What custom bidding actually is
Custom bidding lets you supply your own definition of value. The mechanics differ by platform. In some, you write a scoring script that assigns a value to each impression from the signals available. In others, you attach values to conversions and let the platform learn from them, or connect an external model that adjusts bids. In every case the platform still does the auction work. What changes is the objective it is working toward.
What you can optimize toward
- Margin instead of revenue, so the algorithm stops chasing low-profit orders.
- Predicted lifetime value, so acquisition is judged on what a customer becomes, not on the first basket.
- New customers, weighted above people who were going to buy anyway.
- Quality of exposure, such as attention or viewable time, for brand campaigns.
What it needs to work
An algorithm learns from examples, so volume matters. If the outcome you care about happens only a few times a week, there is not enough signal, and you need a reliable proxy that happens more often. You also need the data to reach the platform cleanly and quickly: values that arrive late or patchy teach the algorithm the wrong lesson.
How to know it worked
Run the custom algorithm against the platform default as a proper split, with the same budget, audience and dates, and judge both on the business outcome, not on the platform's own metric. Where you can, confirm the result with a lift experiment. An algorithm can get better at claiming credit for sales without causing any more of them, and only an experiment tells those two apart.
Where it goes wrong
- Optimizing toward a signal that is easy to game. If the value signal rewards cheap conversions among existing customers, that is what you will get more of.
- Set and forget. Prices, products and audiences change. A scoring model that is never revisited drifts.
- Not owning the logic. If the algorithm lives in a vendor's black box, you cannot inspect it, and you cannot take it with you.
Done well, custom bidding is not a clever trick. It is simply telling the machine what you actually want, and then checking that it listened.