ChatGPT ads now start at $25 a day, and clicks look cheap next to Google. Payback is still unproven. Here is how the cost model works, what the early evidence says with each source named, and the decision rule I would use as a challenger brand: audit organic AI visibility first, then test small.
What ChatGPT Ads Actually Cost, and How to Decide If They Are Worth It Yet
By Stephen Honight, Founder of Lmo7
The prices for ChatGPT ads are now public, and they are small. What nobody can show you yet is whether the clicks pay back. My sense is that anyone who answers that question with confidence, in either direction, is working from a sample of one.
This piece exists because the question I keep hearing is practical. Founders and heads of ecommerce can see that the ads exist. They have no read on what a test costs or how to judge it. So this covers the cost model as OpenAI publishes it, the mechanism underneath, the early evidence on payback with each source named, and the decision rule I would use for a challenger brand. It updates our earlier piece on why ChatGPT is becoming the biggest product search surface, written before the ads launched.
The short version
- What it costs. A floor of $25 a day per campaign (GBP 15 and EUR 15 in those currencies) and a suggested starting maximum CPC of $3 to $5. OpenAI publishes no average prices (OpenAI Help Center, checked October 2026).
- What is unproven. Payback. The early evidence is one partner-measured result for a large advertiser and one agency report of ChatGPT ad visitors converting at 0.21% against 3.71% from Google Ads.
- What I would do. Audit organic AI visibility first. Test small on the prompts you lose. Judge on incremental enquiry and branded lift over at least 72 hours.
What the published numbers say
OpenAI sets a minimum daily budget of $25 USD per campaign. It is set per billing currency across 30 currencies, which makes it GBP 15 and EUR 15 for UK and European buyers. The pilot-era six-figure commitments are gone (all figures in this section from OpenAI Help Center, checked October 2026).
For Clicks campaigns, OpenAI suggests a starting maximum CPC of $3 to $5. That figure is a bid ceiling and says nothing about what you will pay on average.
There is no published average to compare it with. OpenAI publishes no average CPM or CPC, and its own FAQ says benchmarks across advertisers, industries and campaign types do not exist yet. That means any industry CPM table you see without a named source is invented. I would ignore it.
What we can do is set the suggested bids against the channel everyone already understands. WordStream’s 2026 benchmarks put the median Google and Microsoft search CPC at $5.42 across 13,474 US campaigns. ChatGPT’s suggested bids sit at or below that median click price.
Then there is what operators are seeing. A Search Engine Land field report from one agency running the platform puts average CPCs at $2 to $5 across industries. It also carries a delivery warning: set your max CPC below $3 and the ads struggle to serve.
Our own ballparks sit in the same range. In an OpenAI onboarding session Lmo7 attended, the economics discussed were a CPM around $45 and CPCs of roughly $3 to $5. These are ballpark figures from that session and from live work on the platform. I would not budget on them as fixed rates.
How the auction works, and why the Google mental model misleads
The auction is relevance-weighted and second-price. Your max bid sits at ad-group level. You pay per 1,000 impressions or per valid click, never per conversion. Relevance shapes who wins, and OpenAI’s inputs for relevance are the conversation, your landing page, your title and your copy. A sharper ad can beat a bigger bid.
That last point is where the Google habit breaks. On Google you buy a keyword and bid against everyone else who bought it. Here the context hints are free text. You are briefing the model on who the ad is for and what moment they are in. OpenAI’s onboarding recommendation was to run 100+ ad variations. My read of why is that there is no fixed keyword list to exhaust. Every variation is a different description of the buyer, and the relevance weighting decides which one earns the placement. That workload is for brands scaling the channel. For a small first test you do not need 100 variations; a focused handful written against your sharpest buying prompts is the right size. The point to take is that creative volume, not bid size, is the lever this platform rewards, so plan the workload before you plan the budget.
There are three objectives to choose from:
- Views, which bills on CPM.
- Clicks, which bills on CPC.
- Conversions, which uses oCPC or oCPM and optimises to a standard pixel event.
An existing campaign cannot switch objective. You clone it. So choose before you launch, because changing your mind means rebuilding.
Budget pacing also behaves differently from the Google habit. A daily budget is an average over a Sunday to Saturday week. One day can spend up to 2x the daily figure and the week caps at 7x. On a $25 campaign that means a single day can reach $50 while the week cannot exceed $175.
Then the lags. Pausing is not instant: ads can keep serving for up to 24 hours and that spend is billable. Spend reporting lags by 7 to 8 hours. The practical consequence is that you cannot run tight intra-day kill rules on this platform. Set your rules on weekly numbers and agree them before the first pound is spent.
Why the price fell, and what that tells you
The pricing history matters because it shows where the market sits. The following is Digiday’s reporting, as covered by trade press in April 2026.
ChatGPT ads launched in February 2026 on a CPM model around $60. Reported clearing CPMs fell to as low as $25 within roughly ten weeks. OpenAI then shifted to offering CPC bidding at $3 to $5 alongside the CPM option, first reported by Digiday. The minimum commitment dropped from $250,000 to $50,000 and then disappeared into the $25-a-day self-serve floor.
My interpretation, and it is only that, is that a falling clearing price means supply outran advertiser demand. OpenAI says ChatGPT reaches 1.2 billion people a week (OpenAI’s own number), so there is a lot of inventory to fill. The CPC option reads to me like an invitation to performance budgets, which want to pay for something closer to a visit than an impression.
This gives you the honest tension. Entry costs have collapsed and clicks look cheap next to Google. Cheap reach tells you the inventory is plentiful. It tells you nothing about whether the people reached buy.
What the payback evidence says so far
There are two pieces of published evidence and they point in different directions.
OpenAI’s 5 October 2026 measurement announcement says DV Rockerbox measured WeightWatchers’ attributed CPA on ChatGPT Ads at 15.3% lower than its blended paid-search benchmark. That is OpenAI publishing a partner-measured result about one large advertiser. It is encouraging and it is not an independent study. I would not carry it across to a brand with a fraction of WeightWatchers’ scale and awareness without testing it.
The Search Engine Land field report goes the other way. It describes one website converting ChatGPT ad visitors at 0.21% against 3.71% for Google Ads visitors. That is one agency and one site. Read it as a warning signal and nothing firmer.
From our side, the read is sceptical too. Across three recent client and prospect conversations, the consistent early view on payback was impressions high, conversions poor. Three conversations is a small number and I am not generalising from it. It is what operators are seeing in the first months, and it matches the gap the Search Engine Land figure points at.
I do not know how much of that gap is poor-fit traffic and how much is a measurement problem. I do not think anyone does yet. The measurement problem is worth taking seriously, though, because an assistant surface does not behave like a search results page.
Here is the mechanism as I see it, and it is my inference rather than a platform finding. A shopper asks ChatGPT for a shortlist. They see your ad or your name and they do not click. They compare prices, wait for payday, ask someone. A day or three later they search for your brand or type the URL. Your analytics records a branded or direct visit. Last-click credits that visit to branded search or direct, and the ad that started the path appears nowhere in it.
If that is happening, a same-session ROAS report will undersell the surface. It will also be wrong in a way you cannot see from the report.
OpenAI’s measurement is improving. The October 2026 update added an Event Quality Score, selectable click-through windows with a one-day view-through window, and third-party measurement partners ranging from AppsFlyer through to Triple Whale. It is a real step. It is also platform-young, and a one-day view-through window is shorter than the 72 hours I would want to look across. You will need to read your own branded search and direct traffic alongside whatever the platform reports.
The decision rule for a challenger brand
My recommendation would be to test ChatGPT ads, and to do it in a particular order. The order matters more than the budget.
Audit organic AI visibility first
The ad auction is relevance-weighted. Relevance depends on the conversation, the landing page, the title and the copy. Two things come from organic AI visibility work, and both decide whether a paid click pays.
The first is targeting data. Share-of-Model tracking shows you which prompts recommend a competitor and not you. Those are the prompts worth paying for, and they are the raw material for your context hints. We wrote up how we built ChatGPT ad targeting from Share-of-Model data if you want the detail.
The second is conversion-readiness. A click lands on a page, and that page has to answer the question the shopper asked the assistant. If your product content is thin, the click is wasted money.
The proof I can give you is organic. At Trip Drinks, a 60-day Lmo7 engagement on organic AI visibility moved average position across the major LLMs from 7th to 3rd and lifted AI referral traffic 33% (source: Lmo7 client engagement with Trip Drinks). That is an organic result. I am not claiming any ads result from it. The point is that the position and the referral traffic came from content and tracking work, and that work is what any later paid test would sit on.
Content and tracking work does not replace the ads. It is what makes the ads pay.
Start small, with arithmetic you can check
The floor is $25 a day, but a test that can tell you anything needs more than the floor. Take $50 a day for 30 days. That is $1,500. At a $4 max CPC, which sits inside OpenAI’s suggested $3 to $5 range and above the $3 delivery warning from Search Engine Land, that budget buys at least 375 clicks. I say at least because $4 is a ceiling in a second-price auction and clicks can clear below it. This arithmetic comes from OpenAI’s published mechanics and is not a forecast of what you will get.
Is 375 clicks enough? It is enough to tell you whether something is badly wrong with the landing page or the targeting. It will not prove a conversion rate to two decimal places, and I would not pretend it does.
Point the spend at the sharpest ready-to-buy prompts you lose
Do not buy broad awareness prompts. Take the ready-to-buy prompts from your Share-of-Model data where a competitor is recommended and you are not, write your context hints for those buyers and run the variations against them. This is where a relevance-weighted auction favours the brand that has done the preparation.
Judge on incremental enquiry and branded lift
Set a baseline before you launch: branded search, direct traffic and enquiries for the weeks before, taken from a period with no promotions running, so a promo spike cannot masquerade as ad lift or hide it. Then read the same lines across the test, over at least 72 hours after exposure, and not on same-session ROAS. A same-session view will show you the smallest number the channel produces.
Given the reporting lag and the 24-hour pause lag, judge on weekly totals and not daily ones.
Kill or scale on a pre-agreed number
Agree the number before the first day of spend, with whoever owns the budget. I am not going to give you the number, because it depends on your margin, your basket size and what an enquiry is worth to you. What matters is that it exists in writing before the data arrives. Without it, a weak week gets explained away and a strong week gets over-read.
If the test beats the number, scale it. If it misses, stop and look at the content on the page before you look at the platform.
What to do next
The checklist above is work your own team can start this month. If you would rather do it with help, this is how I would sequence it, with prices attached so you can judge the trade. Three steps in priority order.
Baseline organic AI visibility and pick the prompts. Owner: your head of ecommerce or Amazon. Our DaaS tier is £250 + VAT per month per marketplace. It is the data layer, an organic AI visibility baseline that tells you which prompts are worth paying for. The commercial reason is that without it your context hints are guesses, and the platform charges you for guesses at the same rate as good targeting.
Run the test properly and fix the pages the clicks land on. Owner: your head of marketing or brand content, working with Lmo7. Challenger, Agentic Ads is £1,000 per month and Agentic Tracking & Content Optimisation is £1,000 per month. Both are fees for the work; media spend is paid to OpenAI on top. The first runs the test with the structure above. The second does the content work that makes the clicks convert and tracks branded and direct lift over the right window. The reason to run them together is that a click is only the start of a sale, and the content decides what happens after it.
Agree the decision rule across the business. Owner: the CEO or founder, with finance and marketing in the room. Enterprise workshops start from £5,000 and cover multi-brand paid-AI planning and stakeholder alignment. This earns its cost when several brands or teams are each about to run their own test and read the results against different numbers. The full tier structure is in our piece on what an agentic commerce agency does and what you should pay.
Start with step 1 this week. List the ready-to-buy prompts in your category and check which ones recommend a competitor instead of you.
Stephen Honight is the Founder of Lmo7, an agency working in AI search and agentic commerce.