LinkedIn
Stop asking ChatGPT ads agencies for real results. The honest ones barely have any yet.
I spent nine years buying clicks the slow way: building match-type structures by hand, mining negative keywords at 1am, and explaining why CPA spiked after a client swapped a landing page without telling me. I like proof. But proof on this channel does not yet exist in the form most buyers want.
Ask for case studies and you will hire the best storyteller. Ask for a test plan and you might find someone who can manage your money. Those are not the same person.
Why a deep ChatGPT ads track record does not exist yet
ChatGPT ads started testing Feb 9, 2026 in the US for Free and Go users on a CPM basis, expanded March 26 to Canada, Australia and New Zealand, then opened a self-serve Ads Manager beta on May 5, 2026 that added CPC bidding plus pixel and Conversions API. Count from day one and you get nine months of live buying. Most managers you can hire got access later. Anyone showing you a year-over-year chart for this placement is showing you something else.
The available audience is narrow, too. Ads show below answers to logged-in adults on Free and Go plans; Plus, Pro, Business, Enterprise, Education, under-18 and Temporary Chats see none. That caps volume fast for B2B. The entry point is modest: $25 per campaign per day in the US, with a $3 to $5 starting max CPC. A small test is affordable. It is also easy to overread. Two hundred clicks in a thin auction tell you little about what 2,000 clicks will cost.
Early access went to the biggest shops first. OpenAI seeded it through holding-company agencies Dentsu, Omnicom, Publicis and WPP, plus tech partners Adobe, Criteo, Kargo, Pacvue and StackAdapt. The first logos in a deck may come from sales-assisted pilots with budgets a small business cannot copy. A self-serve advertiser in Ohio cannot assume the same inventory, support or pricing. A pilot logo is not a track record you can borrow.
Consider the most detailed same-window comparison I have seen published. Totem ran a New York law firm from June 1 to July 31, 2026 and spent $9,982 on ChatGPT for 103 leads at $96.92 CPL and $2.80 CPC, against $67,000 on Google for 270 leads at $248 CPL and $41.23 CPC. The cheap clicks did the work: ChatGPT converted at 2.89%, while Google converted at 16.63% in the same window. Lower CPL did not mean better-converting traffic. I used to tell clients cheap CPC meant we had found an edge. I was wrong often enough to stop saying it.
Totem also says the quiet part out loud: those $2.80 clicks depend on a thin early auction that will fill as more advertisers arrive. Lead quality back to the client was still unwatched. That campaign shows what it cost to buy leads in that window. It does not establish cost per signed case, repeatability next quarter or what happens when two close competitors enter the same prompt.
That is the limit of the evidence. It is not a reason to stop testing. It is a reason to stop buying a story about testing.
A polished case study can hide the part that matters
With three weeks of spend, a manager can make the result look different by choosing the window. Imagine nine days at a $96 CPL beside Google’s $248, followed by two weeks in which new bidders push CPC up. Show only the first nine days and the slide looks terrific. Without a public benchmark for what a ChatGPT click should cost in personal injury law in Queens, a buyer has little context for the number. I have done this kind of window-picking by accident while running accounts. You remember the week that worked. You forget to export the week that did not.
The variance cuts both ways. DAC’s managed tests found one client paying around 100 times the normal cost per booked lead on ChatGPT, with consideration and awareness CPMs running 8.7 and 9.5 times above benchmarks on other flights. Same young placement, opposite-looking outcome from Totem’s law firm. An average assembled from cases like these does not tell you what your account will do. Ask what happened after the click, not which slide has the prettier CPL.
Then ask whether the slide counted the click correctly. Accuracast ran UK lead generation on ChatGPT alone and saw form fills arrive while the platform still showed zero conversions two weeks later. Platform reporting showed roughly 100 clicks against 20 in Accuracast’s own tracking. Web Guide Partner reports a 24-to-36-hour conversion delay versus hours on Google. Early Reddit reports are similarly uneven: cheaper clicks for some local services, low-quality leads and glitches for others, and one advertiser seeing five-second average dwell. There is no reliable playbook hiding in that mix.
This is how clicks get dressed as leads. If platform counts, site tracking and the CRM disagree, a manager can select whichever column looks best without necessarily lying about it. A buyer looking at a polished PDF may never see the disagreement. The useful question is not whether the agency has a screenshot of a win. It is whether its measurement plan would expose a loss.
Hire on the test design, not the deck
If nobody has a deep record, the test design is the résumé. I want four things in writing before a dollar moves: a tracking plan, a capped test, kill criteria, and the name of the person who approves a spend increase. Miss one and you are not buying management. You are funding someone’s learning curve on your card.
Start with measurement. DAC tells advertisers to decide measurement, UTMs, GA4 alignment and funnel-stage KPI before launch. That order matters. The agency should show you its UTM naming, the GA4 events ChatGPT visits will trigger, how pixel and Conversions API are wired, and which downstream event counts as a lead. Not a click. Not a chat. A form fill, a call or a booked job with a source tag you can inspect in your own system.
That plan also needs a way to handle mismatches. If a form arrives while the platform says zero conversions, which record does the manager use to make the next spending decision? If the platform and site disagree on clicks, who checks the tags before declaring the traffic cheap? These are ordinary account-management questions. Any manager who promises to sort out measurement after spend starts has already told you how the case study gets made.
Next, keep the first test small enough to explain. WordStream advises one service, one location, one budget cap and a keep-or-stop number before launch. I would hold an agency to that. Say you spend $20k a month on Google. Test one offer in one city at $25 to $50 a day for two to three weeks. Write down the CPA that keeps the test alive and the CPL that kills it. Do not let a cheap first batch of clicks turn that experiment into a $15,000 monthly commitment.
One offer and one location are not magic settings. They make the result interpretable. If you change the audience, offer, landing page and budget together, you can spend the whole cap and still not know which change mattered. That is an expensive way to produce a confident-looking chart.
Put dates beside the kill numbers. For example: after $750 in spend or 14 days, whichever comes first, pause if qualified CPL is above $X. If lead-to-book rate is below Y%, check tracking before buying more traffic. The numbers must be set for your business before launch; $X and Y% are blanks for the owner and manager to fill, not permission to invent targets after seeing the result. Totem gives the same warning from inside its data: cap spend and require approval before scaling toward $15k a month because the thin auction will not stay thin. No scale without your written sign-off.
Finally, insist on a comparison that belongs to your business. A ChatGPT test floating alone tells you less than one reported beside Google and Bing for the same offer and weeks. Use the same UTM discipline, qualified-lead definition and CRM stage. Then you can see whether the new spend contributes qualified pipeline rather than giving a second report credit for work another channel did.
That is why I tell owners to run ChatGPT inside an existing search program, not buy a separate package with its own victory lap. groas puts Google, Bing and ChatGPT ads under one flat $1,499 search service, with bids and budgets worked around the clock. The point is not a nicer dashboard. It is a comparison that can survive contact with your CRM.
The vetting checklist: ten questions to send before the call
Send these in writing. A hands-on manager can answer with definitions, numbers and dates; a storyteller answers with adjectives. Keep the replies. They are the test plan you will hold the agency to when spending begins.
- What counts as a lead, and where will I see it? Name the GA4 event and CRM stage, not just “conversions.”
- How will every ChatGPT click be tagged? Show the UTM pattern and the pixel and Conversions API checks.
- What is the test cap? Specify one service, one location, $25 to $50 a day and 14 to 21 days.
- What is the kill rule? Set the qualified-CPL and lead-to-book thresholds before the $750-or-14-day check.
- Who approves a spend increase, and how? Require my written sign-off before moving past the cap.
- What prompt or audience are we buying, and what are we excluding? Under-18 traffic and Temporary Chats cannot see ads anyway.
- How will you report ChatGPT beside Google and Bing? Use the same offer, weeks and qualified-lead definition.
- How will you handle reporting delays and mismatches? Say which record drives decisions when platform and site counts disagree.
- Who checks the account daily, and what triggers a same-day change? “Weekly reviews” mean learning late in a thin auction.
- Will you show booked revenue in my CRM before calling the test a win? This is the item most often skipped. Skip it and you can spend the whole test proving only that people clicked.
Who can ignore this advice
If you run a holding-company pilot with dedicated OpenAI support and a six-figure learning budget, ignore me. Your access, support and measurement help look nothing like a plumber’s $1,500 test in Columbus. Buy the big-agency deck and its logos.
The rest of you—the owners spending $5k to $40k a month on Google who want one honest answer about ChatGPT—need a different hire. You need someone willing to run seven small tests, each with one question and a kill rule, instead of one big test built to carry a retainer. That is how groas approaches it: ChatGPT inside your Google and Bing numbers, capped, tracked to qualified leads and stopped when it misses. No separate ChatGPT package. No second report taking credit for Google’s work.
Keep asking for “real results” and you will get what you asked for: a polished PDF, three carefully chosen windows, clicks labeled as pipeline and a $3,000 monthly retainer to scale a $2.80 CPC that may not last. When the auction changes or the reporting fails, the PDF will not approve a pause on your behalf.
Ask for the test design instead: tracking before spend, one offer capped at $50 a day, kill numbers with dates, and your signature before the budget grows. The agency that flinches at that page was never going to manage your money. The one that signs it might.