OpenAI Ads bid management: the push and pull loop
Single-intent ad groups give you a clean reading of what each piece of demand is doing. Bid management is what you do with those readings. It is the least glamorous part of the playbook and the part that compounds, a weekly loop that quietly moves money toward what pays and away from what does not.
A simple, repeatable loop you can run by rule
Clicks lie, so measure outcomes instead
Before any of this works, you have to fix what you are measuring. OpenAI Ads reports clicks and impressions by default, and clicks are a trap. They feel like progress, they are cheap to generate, and they correlate only loosely with whether anyone bought. An ad group can win the click contest and lose you money, while a quieter one with a third of the traffic pays for itself twice over. Optimize toward clicks and you will reliably feed the wrong ad groups.
The remedy is to bring your own measure of success. OpenAI Ads provides a measurement pixel and a Conversions API, and wiring one of them up is the first job in this entire playbook, ahead of any bid change. Pass conversions back, and pass a value with them wherever you can. The difference between a conversion count and a value-weighted conversion is the difference between knowing an ad group made five sales and knowing it made five sales worth forty dollars each while another made two worth four hundred. Without the value, those two look similar. With it, they are not remotely the same decision.
Then triangulate. Tag your destination links so your own analytics can tie sessions and revenue back to specific ad groups, and read that against the platform's aggregate counts and your conversion feed. No single source is trustworthy alone on this platform, but together they are enough. The rule of thumb is to treat your own data as the truth and the platform's CSV as corroboration, never the reverse.
How to read small numbers without fooling yourself
Here is the tension at the heart of this approach. The structure that gives you clean signal also gives you thin signal, because splitting traffic across many narrow ad groups means each one sees only a slice of it. A single-intent ad group might gather a dozen clicks in a week. Acting on twelve clicks is not optimization. It is reading tea leaves.
Small numbers are loud and dishonest. An ad group that converts twice on ten clicks looks like a triumph, and an ad group that converts zero on eight looks like a failure, and both readings will often reverse themselves the following week purely by chance. If you cut bids on every quiet week and raise them on every lucky one, you will spend your time chasing noise and your results will drift sideways at best.
The discipline that fixes this is patience with a number attached. Decide, in advance, how much data an ad group needs before its result counts, a minimum of clicks, or better, of conversions, and refuse to act until it clears that bar. Judge over a window long enough to absorb the randomness, not over a single day. For your thinnest ad groups, lengthen the window further or pool them with neighbours rather than pretending a handful of clicks is a verdict. The whole point of building a clean signal is to act on real signal, and real signal takes enough volume to be real.
Push, pull, pause, reallocate
Once an ad group has cleared your data bar, you have four things you can do with it, and they are deliberately simple. The art is in the patience and the measurement, not in the actions themselves.
Push the winners. When an ad group converts profitably against your target, raise its bid so you capture more of the context that is paying. This is where growth comes from, and it is the move people are too timid with. If something is working, the correct instinct is to want more of it.
Pull the losers, and prefer pulling to cutting. When an ad group spends without returning, lower its bid rather than switching it off outright. A lower bid keeps a small, cheap presence in the auction while you watch whether the early read holds, and it avoids the whiplash of killing something that was one slow week away from proving itself. Reserve the hard stop for the genuinely dead.
Pause the dead. When an ad group has spent real money over a real window and produced nothing, stop paying for it and free the budget. The trick is honesty about what dead means, which is why the data bar from the previous section matters so much. Dead is not unlucky. Dead is enough spend, enough time, and nothing to show.
Reallocate the budget. The money you stop wasting on losers is the money you use to scale winners. The bid loop decides what each ad group is worth; budget allocation makes sure the account as a whole leans toward the ad groups that earn it. Done weekly, this slow tilt toward what works is most of where compounding returns come from.
The weekly loop, step by step
Run this on a fixed cadence. Most accounts want a week; only the highest-volume ones can justify faster.
Pull the numbers
Run a report for every ad group: spend, clicks, and conversions with value, over a window that clears your data bar. This is read-only, so it never touches the account.
Judge against target
Compare each ad group to your target return or cost per acquisition, and set aside the ones that have not yet gathered enough data to count.
Adjust in measured steps
Raise bids on ad groups above target and lower them on ad groups below it, in steps of roughly ten to twenty percent rather than dramatic swings, and pause the ones that are genuinely dead.
Reallocate and record
Move freed budget toward the winners, and write down what you changed and why, so next week's result is interpretable instead of a mystery.
The rules that keep you from over-reacting
Most bid management goes wrong not through bad moves but through too many of them, made too fast, on too little evidence. A few guardrails keep the loop honest.
Change one thing at a time on a given ad group, so that when its numbers move next week you can attribute the move to your change rather than guessing among three. Hold to the weekly rhythm and ignore the daily wiggle, because intraday and day-to-day swings are mostly noise and reacting to them just adds churn. Keep a simple log of what you changed and when, which sounds fussy until the week you are trying to work out why an ad group's cost per acquisition doubled and your own past edits are the answer. And size your steps modestly. A run of ten percent nudges that you can reverse beats one heroic adjustment that you cannot, because the entire point of doing this weekly is that you get another turn soon.
Turn the loop into a rule the assistant runs
Everything above is deliberately rule-shaped, and rules are exactly what an assistant runs well. The loop is repetitive, evidence-driven, and identical across hundreds of ad groups, which makes it a poor use of your afternoon and a natural fit for the MCP.
The pattern is to state the policy once and let it run. You might say: each week, for every ad group that has spent at least a set amount, raise the bid by fifteen percent where the return beats target, lower it by fifteen percent where it has fallen short for two windows running, pause anything that has spent the threshold with no conversions, and show me everything you changed. An assistant connected to your account over the MCP pulls the reports, applies that rule across the whole grid, and hands back a summary. You move from doing the loop to supervising it, which is the only way running thousands of single-intent ad groups stays sane.
Where this fits in the playbook
Start with the structure
The bid loop only produces clean decisions if each ad group is a single intent. Build that first, then come back here.
Read the SKAGs guide →Run it from Claude or ChatGPT
State your bid policy once and have an assistant apply it across every ad group, every week, over the MCP.
See the MCP →OpenAI Ads bid management, answered
What should I optimize toward?
Outcomes with value attached, meaning conversions and revenue rather than clicks. Install the pixel or Conversions API and pass conversion value, so an ad group that drives a few large sales is not mistaken for one that drives many tiny ones. On a platform that natively reports clicks, your own conversion data is the only honest target.
How big should each bid change be?
Small and frequent beats large and rare. Steps of roughly ten to twenty percent, reviewed the following week, let you move steadily toward the right bid without overshooting on a single noisy reading. The weekly cadence is what makes small steps safe, because you always get another turn soon.
How much data before I judge an ad group?
Enough to clear the noise, which you should define in advance as a minimum number of clicks or, better, conversions. Single-intent ad groups are low volume each, so judge over a window rather than a day, and lengthen the window or pool the thinnest groups instead of acting on a handful of clicks.
Can the MCP run the loop for me?
Yes, and it is the intended way to do it at scale. You state the policy once, the target return, the step size, the spend threshold for pausing, and an assistant over the MCP pulls the reports and applies the changes across every ad group each week, then reports back what it did for you to review.
Run the bid loop with the MCP
Analytics, dayparting, audiences and the MCP in one console. Build the structure, read the signal, and act on it at scale.