Do not fund your ChatGPT ads test by cutting a Meta campaign that works. A working campaign is not spare budget. It is the control you are about to break.
I keep seeing the same proposal: shift 20 or 30% of Meta into ChatGPT for a month and “see what happens.” Say you spend $20k a month on Meta and it books qualified jobs at a cost you can live with. Move $5k and you take on two costs: the output you give up and the experiment you buy. You pull money from a campaign that has learned how to find buyers and hand it to a channel that has not proven it can find one for your business.
A new audience is a reason to consider a test, not a reason to raid a winner. Any agency that will not say that out loud should not hold both budgets.
Early practitioner reports point to the risk: advertisers who moved budget from Meta reported noticeably higher costs per conversion. I treat those reports as directional. The samples are small and self-selected, not a benchmark to paste into a forecast.
The mechanism matters more than the average. Meta had months of conversion history on those accounts. ChatGPT had none. Moving the money meant giving up established performance to buy uncertain performance.
Say a $5,000 Meta ad set produces 100 qualified leads a month. Cut it to fund a $5,000 ChatGPT test, and those leads become the opportunity cost against which the test must be judged. They were not guaranteed next month. They were still the productive use of that money you chose to abandon. Most proposals show the test budget without showing what it displaces.
Contextual proximity is not the same as finding a buyer
Meta’s conversion history and ChatGPT’s conversational context are not interchangeable. Four ad buyers told Adweek that ChatGPT Ads are a black box because advertisers cannot see exact user prompts or exclude unwanted intent the way they exclude junk queries in Google. Their example stays with me: a premium refrigerator advertiser matching a cheapest-refrigerator query. Same category. Wrong buyer.
An analysis of more than 50,000 prompts found ads appeared in 25.94% of commercial queries, but 14.35% of placements were off topic. In 96.37% of placements, the advertiser was not cited in the answer copy. That is contextual proximity, not proof that the conversation contains a customer for you. Do not price it as an even swap for an established campaign.
Workshop Digital’s published B2B test makes the distinction concrete. The team spent $2,319 of its own money for 69,630 impressions and 739 clicks across services, manufacturing, and financial services:
- ChatGPT: 1.1% CTR, $3.14 CPC, zero marketing-qualified leads (MQLs).
- Google Search: 5.2% CTR, $6.76 CPC, two MQLs during the same period.
The authors explicitly say this is not a direct comparison because intent differs. Good. The point is not that two MQLs settle the channel debate. It is that cheap clicks do not establish cheap customer acquisition.
Specificity helped the clicks, not the pipeline. In the same test, keyword-based context hints delivered 1.2% CTR versus 0.9% for conversational hints, at nearly identical CPCs of $3.13 and $3.15. Better targeting bought more clicks per thousand impressions. The test still produced zero MQLs.
If you cut Meta prospecting that books sales calls to chase a 1.2% CTR with no MQLs attached, diversification is doing a lot of work in that sentence. Keep the productive campaign until the alternative earns its budget.

Equal platform CPA is the wrong hurdle
I hear the opposite mistake, too: ChatGPT must hit the same CPA as Meta or the test dies. That sounds disciplined. It can still reward the wrong channel.
Platforms can claim the same customer. One channel may create demand while another collects the conversion credit. A channel-mix analysis puts Meta prospecting at 55 to 70% incrementality and retargeting at 15 to 35%, with total attributed conversions inflated by 40 to 70% when platforms claim overlapping journeys. Whatever those ranges look like in your account, comparing dashboard CPAs alone does not resolve that overlap.
Write down what a qualified lead and a closed customer are worth before spending one dollar on ChatGPT. Then judge the channels against those economics in one system: your CRM.
Workshop Digital’s $3.14 clicks produced no MQLs in that test. Google’s $6.76 clicks produced some pipeline. The lower CPC did not make ChatGPT the better purchase. Nor does a higher platform CPA automatically make it the worse one if the customers justify the cost.
I want equal economic discipline, not identical dashboard numbers. The test must earn its place through qualified pipeline and customers, not through whichever conversion definition makes the report look cheapest.
Three managers, one budget tradeoff
The agency question follows directly from that measurement problem: who owns the cost of moving the money?
Your incumbent Meta agency has a conflict worth naming. A successful ChatGPT test could threaten its budget. A failed test lets it say it told you so. Ask it flat: “Will you recommend moving money away from Meta if the business numbers support it, even if that leaves less spend under your management?” A team protecting its own account cannot fairly grade a replacement.
A separate ChatGPT specialist has the opposite conflict. It owns the new spend and none of the loss elsewhere. Its dashboard shows its clicks, CTR, and CPC. It does not show the qualified Meta leads you gave up to fund them. That is why I argue that a separate ChatGPT ads agency is usually the wrong hire for B2B. Two owners can leave nobody accountable for the tradeoff.
Coordinated search management earns its keep only if one team owns that tradeoff. Search experience has something useful to transfer: query structure, negative-keyword thinking, and landing-page discipline inform how a team approaches contextual hints. I lay out those distinctions in Stop Hiring ChatGPT Ads Specialists.
But transferable skills are not enough. The search team still has to reconcile its plan with whoever owns Meta. If groas runs your search, it should put three things in writing:
- What pipeline the proposed budget cut puts at risk.
- What replacement outcome the test must produce.
- Who has authority to stop it.
No shared ledger, no test. An agency does not get to call its own campaign a success while ignoring the productive campaign that paid for it.

That is why I hire on one question: can this agency recommend leaving the money where it is?
A serious operator looks first to new money or weak campaigns, not branded search or top performers. If Meta must fund the experiment, the marginal ad set goes under review, not the winner. Any proposal that starts by raiding the winner is a spend transfer with your logo on it. Make the agency defend the source of the budget before it pitches the destination.
Buy a test with a cap, not an invoice with a story
A real test has a media cap, a kill rule, and a fee structure that does not reward spending more. Require written approval for any increase and quote management separately from media. Otherwise, “the test budget” can mean whatever the proposal needs it to mean.
The reported OpenAI floor is $25 per day per campaign, with a suggested $3 to $5 maximum CPC. That makes roughly $750 a month sound like an accessible experiment. Accessible is not the same as informative.
Take an illustrative $4 CPC and 2% conversion rate. Thirty conversions require about 1,500 clicks, or roughly $6,000. A $750 budget buys about 190 clicks and three to four conversions at that assumed rate. That is a thin basis for a verdict, especially when the conversion you need is a qualified lead rather than a form fill.
There is a pacing issue, too: daily spend can reach twice the daily budget under a seven-times weekly cap, while spend reporting can lag seven to eight hours. The agency needs to explain how it manages that exposure. A low daily setting is not a substitute for a clear total cap.
Then add the fee load. AdVenture Media recommends $3,000 to $5,000 per month in media for a first test, plus a 15 to 25% management fee and $3,000 to $8,000 in upfront creative. Those are additional costs, not more clicks. If your total budget is fixed, fees and creative reduce what remains for media.
I used to tell clients a small shift was harmless. I was wrong. A small shift from a winner into an unproven channel, with a separate fee load, can be an expensive way to buy an inconclusive answer. Get the full cost quoted before Meta loses a dollar.
The calendar needs the same discipline. Only Free and Go users see ads; Plus, Pro, Business, Enterprise, and Edu users do not. That limits the reachable audience from the start, particularly for B2B. The test does not get to ignore that constraint because the pitch promises a new audience.
Practitioners describe allocating 5 to 10% to a new channel for 8 to 12 weeks, with a geographic holdout and a preset scale rule. The useful discipline is the bounded allocation, the comparison, and the decision rule, not treating those numbers as permission to cut a winner.
For lead generation, judge qualified pipeline in the CRM and allow the sales cycle to catch up. If your sales cycle is 45 days, a two-week Meta raid cannot deliver a closed-customer verdict. It can deliver an invoice. Do not confuse the two.
Three conditions that can justify moving money
I am not arguing that Meta owns its budget forever. I am arguing that a working campaign does not lose its budget merely because another channel is new. Here is where I would consider a transfer.
1. The Meta money is genuinely marginal
Frequency is climbing. Incremental CPA exceeds your acceptable acquisition cost for qualified pipeline. The ad set has not produced a closed customer in two sales cycles.
That is a campaign to challenge, not a productive campaign to protect out of habit. Cut the weak allocation, not the winner. The distinction belongs in the proposal, with the numbers beside it.
2. The test has a specific new-demand argument
The largest practitioner test shared so far reported stronger results for home services, car rental, international real estate, and personal injury law, and weaker results for SaaS, dental, and criminal defense. I would use that as a hypothesis to test, not a forecast to sell.
If you sell an urgent consumer job, the case might be that a ChatGPT conversation can surface intent you cannot reach on Meta at the same cost. That is more specific than “a new audience.” It gives the experiment something to prove.
Name the demand the test is meant to reach. A channel launch is not a customer-acquisition strategy by itself.
3. You deliberately pay for learning
Sometimes the purpose is research. Say so, price it, and accept the cost rather than hiding it inside a performance promise.
New money or a paused loser can fund a capped test with a holdout and a preset scale rule. Before launch, write down:
- The qualified lead volume and customer cost that justify continuing.
- The time needed to evaluate those outcomes against your sales cycle.
- The spending limit and the person responsible for stopping the test.
Pay for learning on purpose, not by accident. That is the opposite of shifting 30% because ChatGPT promises a new audience. One is a budget. The other is a hope with an invoice attached.
This is why fee structure matters. A percentage-of-spend model rewards managing more spend; it does not reward telling you to leave the money alone. groas runs on a fixed monthly fee, so its fee does not rise just because the test budget does. That removes the incentive to expand the test for a larger management bill.
If your current proposal cannot name the pipeline at risk, separate fees from media, and put the kill rule in writing, reject it. Fund the test from new money or losers, judge it on customers in your CRM, and hire the team that will tell you no. Ignore that discipline and you can end up with a starved Meta account, a thin ChatGPT sample, and an agency billing you for both.

