The bid is the most tempting lever in ChatGPT Ads, run from OpenAI's Ads Manager, and the one that does most damage when misused. Not because the platform punishes the change, but because every time you move it you reset your own measurement: the previous days no longer compare with the following ones.
This is the protocol we use to move bids and budgets: one change, a waiting period and a verdict against rules set in advance.
When to touch it and when not to
Before looking at the bid, rule out everything else. If the ad doesn't fit the conversation, raising the bid only makes the same mistake more expensive: the auction weighs relevance and bid together (OpenAI's official documentation). Hints, ad and landing page come first.
Don't touch the bid if:
- The campaign has been serving for less than 14 days or is in a cool-down.
- You are looking at today or yesterday: today's figures are only half there and yesterday's can still change (checked by us, September 2026), and conversions arrive up to 7 days later.
- The campaign isn't serving for a reason unrelated to the bid: brand review, rejected ad, account spend cap exhausted.
- There aren't 5 conversions in the period you want to judge.
Caps on each move
When it is time to move, you move a little: never more than a 30% cut or a 20% increase at once. The asymmetry is deliberate: cutting reduces spend and shows quickly; raising commits money and can send cost per conversion up before you notice. A big move also leaves you without a conclusion: if you halve the bid and volume collapses, you don't know whether the problem was the price or that you stopped entering the auction at all.
⚠️ Trap: changing and un-changing. Cutting today because cost went up and raising again two days later because volume dropped is the fastest way to learn nothing and to never let automated strategies settle.
The cool-down
After touching something, you wait. And waiting means neither touching again nor judging:
| What you changed | Cool-down | Why |
|---|---|---|
| Bid strategy | 7 days | Automated ones need to rebuild their allocation from scratch |
| Maximum bid | 5 days | Delivery takes days to find its new level |
| Daily budget | 3 days | The spending pace resettles quickly |
During the cool-down the figures are a blend of before and after: if you judge them, you are measuring the transition, not the change.
How to judge: 14 days, 5 conversions and 20% of margin
The verdict is made over 14 full days, with at least 5 conversions in the period, and against a target cost per conversion decided before looking at the data (if you don't set it first, you will always find an excuse to leave everything as it is).
Tolerance is 20%: above target but within that margin counts as "borderline" and is left alone.
Discarding anomalous days (and sorting the median)
One odd day — a server outage, a promotion, a bank holiday — drags the whole average. So before judging, you drop the days whose spend falls below 20% of the median or above three times it. The median and not the mean, because a runaway day moves the mean and leaves the median almost untouched.
And the median is calculated on the sorted data. It sounds obvious and it is a classic bug: taking the middle value of the series as it comes. Seven days of spend, as they arrive:
25 · 3 · 24 · 26 · 2 · 25 · 23
The middle value of that list is 26, which happens to be one of the
extremes. Sorted — 2 · 3 · 23 · 24 · 25 · 25 · 26 — the median is
24. The 20% floor is 4.8 and drops the days of 2 and 3; the
three-times ceiling is 72, which nobody reaches. You judge on five days.
💡 Ninja trick: those discarded days aren't deleted, they are logged. Two anomalous days in a row every week aren't noise: they are a delivery problem to fix before touching the bid.
Switching strategy
The strategies are fixed bid, maximise clicks and maximise conversions (OpenAI's official documentation). Our criteria for moving between them:
- Switch to maximise conversions with at least 30 conversions in 30 days and 12 in the last 14: the sign that the campaign has something to learn from.
- Go back to clicks if, on maximise conversions, there are fewer than 10 conversions in 30 days despite spending. It isn't that the strategy is bad: it is optimising towards a signal that barely arrives, and before changing anything it is worth auditing the measurement.
Remember that a campaign's objective can't be changed after it is created, and neither can the chosen conversion event: going from Clicks to Conversions means a new campaign ("Bidding and budgets in ChatGPT Ads").
"Budget is capping" isn't "budget is spare"
Two different diagnoses that get mixed up:
- Capping: spend hits the daily cap almost every day. Only then does raising it make any sense.
- Spare: spend stays well below the cap day after day. Raising it will do nothing: the brake isn't the budget, it is that you aren't entering enough auctions (relevance, bid or reach).
And the rule that stops you burning money: you only raise the budget of something that already meets the target. If a campaign hits the cap every day but its cost per conversion is above what you can pay, raising it multiplies the problem. First fix the cost — relevance and bid — then scale.
The golden rule
One change at a time, written down. Date, what you touched, from value A to B and why. Without that log, in a month you won't know whether the improvement came from the bid, from the new ad or from the fact that it was September. With two changes on the same day, not even a log.
What to take away
- Before the bid, rule out relevance and delivery: the auction weighs both.
- Caps on each move: never more than −30% or +20% at once.
- Cool-down with no touching and no judging: 7 days for strategy, 5 for bid, 3 for budget.
- Judge at 14 days, with 5 conversions and 20% tolerance against the target.
- Drop days spending below 20% or above three times the median, calculated on sorted data.
- Move to maximise conversions with 30 in 30 days and 12 in 14; back to clicks with fewer than 10 in 30 days.
- Only raise the budget of what already meets the target.