OpenAI Turned Your Daily Ad Budget Into an Average. There's No Stated Ceiling.

Alex Thilén put $60,000 through ChatGPT’s ad platform on a client account and got about $89,000 back. He also watched the return swing from roughly 0.2x on the worst day to about 2.9x on the best. Every media buyer running this platform has a version of that chart, and until last week they all had one thing steadying it: a daily budget that meant what it said.

That guarantee is gone.

In a weekly product update to beta advertisers, OpenAI told them the control was changing.

“Daily budgets will change from fixed daily limits to average daily budgets: the average you want to spend per day over seven days,” the company wrote.

The switch landed the week of July 27, applied to campaigns already running, with no opt-out. OpenAI’s reassurance was that nothing else moves: “No action is required, and your total budget is not changing.”

The seven-day total doesn’t change. The thing that changed is what a single day can do to you.

Daily budgets had existed on the platform for nine weeks. OpenAI introduced them on May 22 as a hard per-calendar-day cap, then converted them to a rolling average 63 days later. A control that meant “never more than this today” now means “roughly this on average, and we’ll tell you the week’s number.”

The sentence Google publishes and OpenAI doesn’t

Google runs the same averaging model, and has for years. The difference is what it commits to in writing. Google’s help documentation states you will never pay more than “your daily spending limit (two times your average daily budget for most campaigns) on any particular day.” It caps the month too, at “30.4 times your average daily budget for most campaigns.”

That’s a contract. Two times, stated as a number, published where anyone can read it.

OpenAI says spend “may vary by day while staying within daily and seven-day limits” and names no multiplier. The seven-day total is bounded. What any given Tuesday can cost you isn’t, at least not anywhere an advertiser can point to.

For a team running $2,000 a week, the practical difference is small. For a team whose card gets declined at a certain number, or whose finance lead approves spend daily, an unbounded single day is a real operational problem. And it’s the kind of detail that only becomes urgent after the invoice.

Why the change is defensible anyway

The mechanism behind this isn’t a cash grab, and it’s worth being straight about that.

Hard daily caps produce a specific failure. Campaigns spend their allotment early, usually in the morning, then go dark for the rest of the day. You buy the cheapest hours instead of the best ones, and your ads miss whoever shows up after lunch. Averaging exists so the system can spend more on a high-intent day and less on a dead one. OpenAI also added intraday pacing in the same release, saying daily budgets “now pace automatically throughout the day, helping distribute spend more evenly.”

That’s a genuine improvement for performance. It costs you a guarantee to get it.

Which brings back Thilén’s chart. He’s co-founder and VP of digital strategy at the agency Opascope, and he published the 15-day run in June, before any of this changed.

“Day to day, it was choppy. ROAS swung from roughly 0.2x on the worst days to about 2.9x on the best.”

Blended across the window, that account worked. The day-level spread between his worst and best was roughly 14x, at about $1.72 per click, and his read on a serious test is to “plan for roughly $5K to $10K to get a read you can trust.”

Now let the daily number float on top of that swing. On a day the system reads as high-intent, it can spend more, and high-intent by the model’s reckoning isn’t the same as high-converting by yours. The averaging is designed to lean into exactly the days that produced 2.9x. It has no way to know in advance which ones produce 0.2x.

What this says about the platform

Adthena counted 7,378 distinct advertisers inside ChatGPT during the week of July 13 to 20. That’s a real market, nine weeks into having budget controls at all, on a surface where we’ve already covered ad formats that replace the landing page with a conversation and AI Mode keeping ads separate from citations.

Google didn’t publish a two-times commitment at launch either. That number is the residue of years of advertiser complaints, billing disputes, and eventually a written policy. OpenAI is nine weeks into the same road, moving fast enough that a control introduced in May got redefined in July, on live campaigns, by email.

That pace is the story, not the averaging, which is standard. The fact that a budget mechanic can change underneath running campaigns, on a platform where the advertiser still carries the liability for what automation does, is the operating condition of buying here right now.

Three things to do this week.

1. Reconcile at seven days, not at one. Your daily number isn’t a number any more. Pull spend on a rolling weekly basis or your pacing report is measuring something that doesn’t exist.
2. Ask your rep for the single-day ceiling in writing. Google publishes 2x. If OpenAI’s answer is a range or a shrug, cap it yourself with a lower average and accept the tradeoff.
3. Watch the days after a spike. Averaging pulls spend forward. Check whether the high-spend days are the ones your conversions came from, because that’s the only test of whether the pacing is working for you or on you.

The budget line still says the same thing. It just means something different now, and nobody’s campaign got paused to explain it.

Update, 2026-08-06. Two days after the budget change, OpenAI added product carousels to ChatGPT ads, letting several products from one retailer appear in a single placement at the bottom of a conversation. Until now the format ran one product per advertiser. Digiday reported the change today, noting the carousel pulls product data straight from retailer feeds rather than from individually built ad units, and that OpenAI’s own systems decide whether a given impression gets a single product or a carousel.

Read it next to the pacing change and the direction is consistent. Both moves take a lever out of the advertiser’s hands: you no longer set the hard daily ceiling, and you no longer decide whether your ad shows one product or six. What you supply is a feed and an average.

The timing is the tell. Ashley Fletcher, CMO at Adthena, told Digiday that “Q4 is going to be a key period for them,” and an unnamed US agency executive in the same piece put the motive plainly: “Q4 is coming, the holiday shopping season is coming, so OpenAI wants to make sure that these products are discoverable in ChatGPT conversations.” A platform building feed-driven inventory ahead of the holidays is a platform trying to book seasonal retail budget, which raises the stakes on the missing single-day ceiling rather than lowering them. Our earlier piece on OpenAI’s $100 billion ads forecast covers how far that ambition runs ahead of what analysts size the market at.

So add a fourth item to the list above. Audit your product feed before Q4, because it’s now the creative. If the feed has stale prices or thin titles, that’s what a carousel will render, and you won’t be choosing when it does.