ChatGPT Ads: Six Months In, Our Cost Per Lead Beat Paid Search

The number everyone quotes about ChatGPT Ads is a 0.21% conversion rate against 3.71% for Google Ads, published by Search Engine Land in August 2026. It is real, and our first month was worse than that implies. Cost per approved lead came in at a little over three times our Google Search benchmark on the same account.

Six months later the same campaigns deliver an approved lead at roughly 35% below that Google benchmark.

We have been running ChatGPT Ads since March 2026, about a month after OpenAI announced the programme, for a US financial advisor matching platform. This is what the curve looked like, and what actually moved it.

The curve

The maroon line is our cost per approved lead as a multiple of the paid search benchmark. The grey line is that benchmark. Months one and two are what waiting looks like. The bend starts in month three, and the lines cross at the end of month four.

One caveat we would rather state than have pointed out to us. The search programme is eleven years old and runs at roughly three and a half times the monthly budget of the ChatGPT account. Beating a mature, heavily worked search programme at under a third of its spend is the result we have. It is not the same as proving the channel holds up at that scale, and cost per lead usually rises as you scale, because you exhaust the best inventory first. We have not tested it at search-level budget yet. Ask us again when we have.

Two things in that table are worth pausing on.

The paid search benchmark did not move. This is not a story about search getting worse. The comparison holds still while the new channel improves against it, which is the only way a before-and-after like this means anything at all.

Click-through rate and conversion rate barely moved either. That surprised us. The gain did not come from a better click rate or a better landing page, both of which sat at unremarkable numbers throughout and roughly in line with what the same client sees on Meta. It came from reaching better qualified conversations, which is a targeting and measurement problem rather than a creative performance one.

What actually happened, month by month

Months one and two. We did what we were told.

The guidance we were given was to leave the campaigns running and let the system find relevance on its own. So we added conversion tracking, assumed that was the missing piece, and waited. Month one came in at over three times our search benchmark. Month two barely moved, and by the middle of it we could see the account was not quietly correcting itself. Waiting was not a strategy. It was just waiting.

Month three. We stopped waiting.

We did our own research, asked everyone we could find who had actually run these campaigns, and ran small experiments of our own. Three things came out of that, and they are the whole of what changed.

• Conversion tracking was table stakes, not the answer. It is necessary. It is not what moves cost per lead.
• Ad copy carries the relevance that keywords carry in search. There is no keyword field, so the words in the hint and the words in the ad are doing the targeting. In search you can write mediocre copy and let match types save you. Here you cannot.
• Creative has to be legible in one glance. Text-heavy infographics underperform badly in this placement. What works is a single centred image that carries the whole value at a glance, even a plain stock photo, as long as it genuinely fits what the ad says. Small and obvious beats detailed and clever.

Month three is where the line started to bend.

Month four. Crossover. By the end of month four, cost per approved lead had come down to the search benchmark.

Month five. Just below it, and holding, at roughly 12% under search.

Month six. The current level, about 35% below search.

Month seven is running now.

The ninety day review is the whole argument

If our client had reviewed this channel at 14 days, or at 30, they would have killed it. Month one cost over three times what search costs them. Every instinct and every dashboard would have said stop.

So set a 90 day review, and agree the date before the first dollar is spent. Then be clear with yourself about what that review is for, because this is where most tests die badly. At 90 days we were still behind search. The channel had not yet won anything. What we had was three months of slope, and the slope is the decision, not the number sitting in the cell that morning.

If cost per lead is falling meaningfully month on month at day 90, you are on our curve and the crossover is roughly a month out. If it is flat at day 90, you have a structural problem rather than a patience problem, and another quarter will not fix it.

A channel that punishes short review cycles is not unusual, and paid social behaved this way for years. What is unusual here is that the published benchmarks are all drawn from short tests, which means the figure the entire market is quoting describes month one of a six month curve.

Context hints are not keywords

There is no keyword field. None. What you get instead is a freeform text box at ad group level where you describe the conversations your product belongs in, and that is the whole targeting apparatus apart from geography and first-party audiences.

That is a different skill entirely. A keyword is a match. A context hint is a brief. You are not buying a query, you are describing a moment, and the system decides whether your ad belongs in the conversation someone is currently having.

The teams that struggle are the ones who paste a keyword list into the box. The ones who do well write the way a planner writes an audience definition, describing the situation, the problem the person is working through and the decision they are approaching, which in practice means running a handful of ad groups with sharp descriptions instead of the hundreds of tightly themed groups a mature search account would carry.

Here is the same theme written both ways, from our own account. First, the way most teams start, which is a keyword list wearing a sentence's clothes:

fiduciary financial advisor, fiduciary near me, registered fiduciary
near me, what is a fiduciary financial advisor, fee only fiduciary,
fiduciary advisor cost

And the version we actually run:

People asking what a fiduciary financial advisor is, whether their
current advisor is a fiduciary, and how to find a registered fiduciary
near them. They have $250K+ invested and want advice that is legally
required to be in their interest.

Same theme, same landing page. The second one names a person, a state of mind and a threshold. It reads like a brief you would hand a media planner, because that is what it is.

Four rules we now work to. Roughly 280 characters per ad group, which is enough for two or three sentences and no more. One theme per ad group, never two personas in the same box. The landing page has to match the hint, or the description you wrote is a promise the page breaks. And mirror the hint's language in the ad copy itself, because consistency between the two appears to help matching.

One open question worth knowing about. Naming a competitor directly in a hint, as in describing people weighing a switch away from a named firm, matches noticeably better than describing the situation generically. Google restricts competitor names in ad text and OpenAI has not published a settled equivalent, so we describe the switching behaviour without naming anyone until that is clarified.

Your usual bid management levers also do less work here, because you cannot narrow your way to relevance the way you can in paid search.

The mechanics

An agency cannot open the account for you

An agency cannot open a ChatGPT Ads account on behalf of a client. The client creates it under their own legal business name and tax ID, and the agency is added as a user on it.

If you run paid media through an agency, four things change, and all four are better decided before launch than after. Billing sits with the client rather than the agency, so there is no pass-through invoicing. Access is granted and revoked by the client, which is arguably how it should always have worked. Offboarding is clean, because the account and its history stay with the advertiser. And somebody on the client side has to finish tax and payment setup before anything runs. That last one is usually what delays a launch by a week.

Sort this out first. It is the most common reason a launch slips.

If your category is restricted, it is not necessarily closed

Financial services sits on the restricted list, alongside adult, alcohol, tobacco and gambling. Read that list quickly and you would conclude the channel is shut to you. It is not. Our own account is in one of those categories.

Restricted turned out to mean reviewed rather than refused. Getting cleared took two things. The first was working the process through customer support directly rather than waiting on an automated decision, which is slower and considerably more manual than anything you will be used to on Google or Meta. The second was a US partner vouching for the advertiser.

Plan for that in your timeline. If your category is on the list, treat approval as a project that runs before the campaign build rather than a checkbox during it, and start it early. Nobody advertising in a restricted category should read that list and assume the answer is no.

What launched in India, and why it matters to a US advertiser

OpenAI opened the ChatGPT Ads self-serve manager in India on 4 September 2026. More than 50 brands are live and WPP and Omnicom are launch partners. The minimum daily budget is 725 rupees, roughly $7.60, against $25 for the same thing in the US.

Scale is the reason to care. India passed 100 million weekly active ChatGPT users in February 2026, and globally the ad product went from launch to a $1 billion annualized revenue run rate in under 200 days, which is roughly the pace at which a new channel stops being optional for anyone buying media seriously.

That minimum makes India the cheapest place in the world to learn this platform, at about a third of the US entry point, and the mechanics are identical. The catch is the account rule above. An account needs a legal business name and tax ID, so this route is open to you only if you have an India entity. If you do, learning on a 725 rupee daily budget before carrying that knowledge into a US account is considerably cheaper than what our own education cost, which was three months of expensive leads.

Where this sits next to AI search

ChatGPT ads and AI search visibility are separate problems that get discussed as one.

Ads are bought. They appear below an answer, clearly labeled. Being cited inside the answer is earned, and no amount of ad spend affects it, because OpenAI is explicit that advertisers cannot pay to influence responses.

Buying the ad slot while being invisible in the answers around it is a weak position. The budget papers over it for exactly as long as you keep paying. The durable work is AI content optimization, the slower business of becoming the source an assistant reaches for when it answers a question in your category, which we have written about before in terms of the difference between ranking and being cited.

Who should run these now, and who should wait

Test it now if you sell something people research conversationally before buying. Considered B2B purchases, software, professional services, education. Our own result comes from financial advisor matching, which is about as considered as a purchase gets.

Test it now if you are a direct response advertiser measured on cost per lead, provided you can commit to a 90 day window. An earlier draft of this article advised the opposite, that CPA-driven advertisers should wait. Our own data contradicted it. The channel does work for direct response. It simply does not work in the first month, and an advertiser who cannot hold a position for a quarter should not start.

Wait if your measurement stops at last click, because a channel that reaches people early will always look worse than it is under a last click report. That is the case for multi-touch attribution before you launch rather than after the first disappointing month.

Wait if your paid search account is not yet fully worked. Search is still the cheaper lesson.

What we would do on day one

• One campaign, three ad groups, three genuinely different context hints. Not thirty.
• Clicks objective, not Reach, until you know what a click is worth to you.
• Bid above $3. Below that the platform may not spend, and a campaign that does not deliver teaches you nothing.
• Install the pixel and tag every URL before launch. Retrofitting conversion tracking after two weeks of spend wastes the two weeks.
• Send traffic to a page built for an earlier stage. A pricing page wastes it, which makes landing page optimization more decisive here than in search.
• Agree the 90 day review date in writing before you spend anything.

If you want this run properly, that is what our ChatGPT Ads management team does, and the full account story is in our financial services case study.