How We Save AudioMob $100K Every Month
AudioMob was overpaying on every ad impression it bought. The bidding pipeline had no way to tell a valuable impression from a worthless one, so it paid close to the same price for both, and any attempt to simply bid less put campaign delivery at risk.
The Challenge
AudioMob buys audio ad impressions programmatically, at a volume where a small per-impression error becomes a large monthly number. The bidding logic could not distinguish a placement likely to convert from one that was nearly worthless, so it paid a similar price for both. The obvious fix, bidding less across the board, was not actually available: bid too low and you stop winning the good inventory too, campaigns underdeliver, and the client relationship is the thing at risk rather than the margin.
What We Built
We treated it as a pricing problem rather than a budgeting one. Instead of one bid rule applied everywhere, each incoming bid request is scored on its own characteristics against the goals of the campaign it would serve, and the bid moves with that score. That way spend is withdrawn specifically from impressions that were never going to pay for themselves, and left in place on the ones that carry the campaign.
Decisions Worth Explaining
Scoring at request time, not in a nightly batch
A bid request is worth nothing once the auction closes, so the model had to return a score inside the auction timeout rather than inform a report the next morning. That constraint drove the model size and the choice to run inference inside the existing bidding path rather than as a separate service call.
Holding delivery constant as the success metric
Cutting spend is trivial in isolation and would have been the wrong win. We measured against campaign delivery rather than cost alone, so a saving that came at the price of underdelivering would have registered as a regression instead of a result.
The Result
Monthly ad spend fell by more than $100K with campaign delivery held flat, which is the part that mattered. The work paid for itself within three months, and bid efficiency improved by roughly 40 percent.
Under the Hood
A TensorFlow model trained on impression-level data, scoring each bid request against campaign goals in real time inside the Beeswax DSP pipeline, so the bid moves with the actual value of the placement instead of a fixed floor.
Technology Stack
Want Similar Results?
Let's discuss how we can help solve your technical challenges.