The ChatGPT Ads accounts that report more revenue than spend, and the arithmetic behind them

Among the results reviewed here, two ChatGPT Ads accounts report more revenue than spend. One had orders averaging over $1,000; the other paid $1.72 a click, well under what most advertisers report, and still only reached 1.49×. The rest of the detailed results I found got few or no conversions, or didn't publish enough to check.

I read every result I could find from June to late September, about a dozen with enough detail to compare. Most were published by agencies or vendors, and none of those reports describes a randomized holdout, so every return below is attributed revenue.

The arithmetic

Revenue per click is conversion rate times order value. Divide that by what you pay per click and you have ROAS:

ROAS = conversion rate × average order value ÷ CPC

ChatGPT's inputs are the unusual part. The two ecommerce accounts paid $1.72 and $4.41 a click. Most B2B tests paid around $9 to $10, one paid $3.14, and one test found New Zealand the most expensive market at $22.89. Click-through rates in the published tests mostly sit between about 0.5% and 1.1%; one agency reports 0.5% to 2.5% depending on intent, and the desktop-software vendor below saw 7.9% early on. The same agency, pooling client accounts across several verticals, found paid clicks converting at roughly 2%, which it put at about half what organic ChatGPT traffic converts at.

With those inputs, this is the order value you need:

1× only covers the ad spend. With a 40% margin you need 2.5× before the campaign pays for itself.

For a subscription, use revenue over a stated payback window in the ROAS formula, then apply your margin to judge whether the ads pay for themselves. Keep forecast lifetime revenue separate from money already received.

The two accounts with revenue above spend

The ecommerce agency's client spent $9,620 at $4.41 a click and a 0.94% click-through rate, and went from $7 a day to about $1,000 a day in roughly a month. The agency says its average order is over $1,000, which is what makes a $4.41 click affordable: at that order value the revenue from one sale exceeds the cost of two hundred clicks, before fulfilment costs. It reported between $19,000 and $38,000 in attributed revenue depending on the attribution model. Search Engine Journal put the ROAS at 3.3× to 6.8×, and the agency's own post says about 5×. The spend and revenue as published divide to roughly 2× to 4×, and neither source explains the difference, so I use it as "positive under both models" and leave the multiple alone. They checked results in Triple Whale alongside OpenAI's reporting.

Opascope, which runs paid media for clients, published one scaled account: about $60,000 over 15 days, a $1.72 click, a 2.35% conversion rate, about $89,000 in revenue, 1.49× blended. Working backwards, that's roughly 34,900 clicks and 820 orders at about $108 each. With orders that size it only covers its ad spend because the clicks were cheap. That early account's CPC isn't a rate to budget on: Opascope says part of the return came from being in before the auction filled up. Daily ROAS swung between about 0.2× and 2.9×. The figures cover only the campaigns they moved budget into after those showed signal, not everything they launched, and they tell readers to judge on rolling windows because even a one-week read misleads.

The pooled agency (Q1Media) didn't publish a return. What it did publish is what worked across its accounts: a specific problem with a direct fix, hints with decision wording in them ("reviews", "comparisons", "near me", "best options", "which is better"), UTMs on every ad, and at least $5,000 to $10,000 a month so there's enough data to read. It also writes exclusions into its hints ("define who you are not targeting"). OpenAI's hints page says hints can't enforce exclusions, so that's the agency's experience rather than a setting you can rely on.

On the ad itself, David Dugan, OpenAI's head of global ads solutions, told reporters at a Cannes Lions briefing in June that response rates are "much higher" when the call to action reads as a benefit or as a response to the question the person originally asked, compared with an ad that just names the product. None of the accounts above published their ad text, so I can't show you one that worked.

The tests that got nothing

At least five published tests ended with no conversions or close to it. Three were agencies advertising their own services. That's a considered business purchase, often made by someone on a paid ChatGPT plan, which is one agency's own explanation, and two of those tests paid around C$7 and US$9 a click. Plus, Pro, Business, Enterprise and Edu accounts don't see ads. A separate B2B advertiser traced 336 paid clicks back to 146 organisations; five matched their customer profile.

The write-ups don't agree on why. One agency used a dedicated landing page, three ad groups each with its own hint, and a bid four times OpenAI's suggestion, got no lead from 53 clicks at about $9 each, and put it down to the platform: with only reach and clicks to bid on, "the system can't learn its way toward your best prospects." Another spent $2,319 on 739 clicks without a qualified lead and concluded that ChatGPT reaches people who "may not be ready to commit." The third had OpenAI's pixel and Conversions API connected and blamed the audience outright: "The mechanics worked. The audience did not."

I'm closer to the third. If your buyer is a company, the question to answer before spending is whether the person deciding uses a free ChatGPT account while they decide. Who sees ads in ChatGPT is how I'd check.

The cheap-clicks test

The most detailed test is also the cheapest. A software vendor selling a desktop app bid between 7p and 15p against OpenAI's suggested £2, bought 2,988 clicks for £289.52, and got 14 installs: 0.46% as reported, against 5.3% from Google Ads and 4.2% from free ChatGPT referrals to the same site over the same period. Average time on page was about seven seconds. The vendor had the traffic checked for bots. That review found most visits likely human, with red flags on about 30% of clicks. The vendor concluded that poor targeting was the main problem; the review doesn't establish that every click was valid. Each install cost about £21.

It's also the closest public test to a self-serve software purchase, where one person decides and pays. OpenAI describes its auction as relevance-weighted and doesn't publish the details. My reading, which I can't prove: clicks at about a twentieth of the suggested bid (10p against £2) win the placements other advertisers passed on.

The click-counting question

You'll also read that OpenAI bills clicks that never reach the site. The published cases behind that are single campaigns of about 50 to 60 clicks, compared against analytics without server logs. In the vendor's first write-up, OpenAI's clicks didn't match Google Analytics until the vendor learned to add a utm_source tag; three weeks later they reported the two were close. OpenAI's click reference (such as oppref, which its help center tells you to keep for the Conversions API) isn't a substitute for UTMs. Without campaign tags, paid clicks can appear under a chatgpt.com referral or Direct in GA4. Clicks vs GA4 sessions goes through the other reasons the counts drift.

Our own campaign has been small. From 24 to 26 September our server recorded 8, 8 and 5 arrivals carrying the campaign's ID, the same as OpenAI's daily click counts. Three days is three days, and OpenAI doesn't publish how it defines a valid click.

Moving the numbers

I'd work on conversion rate first. On the vendor's site, free ChatGPT referrals converted at 4.2% and paid clicks at 0.46%. The site can clearly convert people who come from ChatGPT. For paid traffic I'd build one page per situation the ads are written for, pick up where the answer left off, and show the price.

Order value is worth a look if you sell bundles or higher tiers. CPC, if OpenAI's description of the auction holds, moves with relevance as well as with the bid, which is one more reason not to chase $1.72 clicks by bidding a fraction of the suggestion. None of it helps if your buyer is on a paid plan.

How I'd set up a test

1. Put your own conversion rate and margin into the table. If you're a factor of three short, fix the offer or the page first. A campaign won't close a gap that size.
2. Free or Go account for whoever decides?
3. Tracking before launch: UTMs on every URL, the OpenAI pixel, the Conversions API, and orders counted in your own system.
4. Separate ad groups where the message or the landing page differs, each with hints that describe a specific situation and its own page.
5. Ad text that answers the question the person asked, within 50 characters of title and 100 of body.
6. A bid in OpenAI's suggested range, or a conversion objective once you have the event volume for it.
7. A campaign total budget, at least two weeks, and a stop rule written before launch.
8. Judge on rolling windows of several days, and compare the paid conversion rate with what free ChatGPT referrals do on the same pages. If paid is a small fraction of organic, something in the matching or the page is probably off.

What nobody knows yet

The ecommerce agency's attribution models disagree by a factor of two on the same campaign, and its published numbers don't reconcile. OpenAI's FAQ says it doesn't yet have performance benchmarks across advertisers, industries or campaign types. Opascope credits part of its return to being early, and nobody has published how fast that fades as advertisers pile in. If you've run a holdout on ChatGPT Ads, I'd like to see it.

Sources

• Ecommerce agency's client, B2B organisation match, CPCs by country: Search Engine Journal, 6 months into ChatGPT ads, advertisers still don't know what good looks like, 8 September 2026; the agency's own post on X, 1 July 2026.
• Opascope, ChatGPT ads benchmarks, updated 15 June 2026.
• Q1Media, ChatGPT ads, part 2, 24 July 2026.
• The desktop-software vendor: My experience buying ads in ChatGPT (13 August 2026) and ChatGPT ad targeting is garbage (2 September 2026).
• Agency tests: Symphonic Digital, Workshop Digital, Choice OMG; MediaPost, Data reveals ChatGPT ads not performing.
• David Dugan: Marketing Dive, How OpenAI is positioning ChatGPT ads, 23 June 2026.
• OpenAI Help Center: Ads in ChatGPT: The Basics, Frequently asked questions, Ads in ChatGPT, Write context hints, Measure results.
• OpenAI developer docs: Conversion-optimized campaigns.
• Our own campaign: Serge server arrival records, 24–26 September 2026.