Most ChatGPT ads agency pages ask you to trust their track record before you can see a dated campaign. I built a composite page from five live ones to take apart that pitch, not to pick a fight with one shop. The recurring pieces are familiar: a proven results headline, certified-expert badges, an AI-optimized promise, three process steps, a results strip, and a pricing section that may not contain a price.

The problem is the age of the channel. OpenAI confirmed ad plans on Jan 16, 2026, started a US pilot on Feb 9, and opened self-serve Ads Manager on May 5. A reported 100,000-plus advertisers across 89 countries by Oct 21 does not turn a short self-serve history into years of benchmarks. That is the test for this page: which claims can a buyer verify before paying, and which just borrow confidence from an older channel?

The composite page, from top to bottom

At the top sits Proven ChatGPT Ads Management That Drives ROI. Below it come badges for certified specialists, AI-optimized bidding, full-service management, and any budget. The page then offers three steps: audit, launch, optimize. A strip of three client logos stands in for results. Pricing appears last, though the number may not.

I recognize the skeleton. I have seen it sell Google Ads for years. One live version prices one campaign and three ad groups at RM2,000 a month, rising to RM5,000 for five campaigns and 15 ad groups, with media paid separately through the client’s OpenAI account. Its listed work includes strategy, intent research, context hints, ad copy, landing-page matching, and monthly optimization. Those are recognizable services. The question is what the page lets you check.

I will use three verdicts. Holds up means you can inspect a setting, a deliverable, or an action log. Unverifiable means the claim might be true, but the page gives you no way to check before you pay. Red flag means the wording obscures who does the work or who carries the risk. Now start where a buyer starts: the headline.

The hero: “Proven results with ChatGPT ads”

Verdict: red flag. I used to tell clients I had proven Google Ads playbooks. I was wrong to lean so hard on proven even there. On ChatGPT, it has less room to breathe. A seller might have worked with early partner inventory, where one published account of the buying terms cites $60 CPM and a $200,000 minimum. That is not the same proposition as running a small self-serve test. Nor does experience across changing buying interfaces produce a stable ROI benchmark overnight.

The headline could make a narrower, useful promise: Here is a dated ChatGPT campaign, what we spent, and what we counted as a qualified lead. Instead, proven ROI asks you to accept the conclusion before you see the inputs.

Ask for the campaign dates, spend, conversion definition, and evidence that the conversion was tracked. A logo without those details does not establish results; it establishes that somebody approved use of a logo. No dates, no spend, no conversion definition: no proof. Hold that thought when we reach the results strip.

The badge row: “Certified specialists”

Verdict: unverifiable as written. Certified by whom, for what, and when? A badge can tell you that someone completed a credential. It cannot, on its own, tell you who will work on your account or whether that person can set up a ChatGPT campaign and measure it.

I would ask the seller to name the credential and the person responsible for the account. Then I would move straight to work samples: a tracking plan, a dated change log, and a decision made after launch. If the answer circles back to the badge, the badge is doing too much work.

This is not an argument against training. I have spent enough time fixing accounts to prefer a trained operator over an enthusiastic improviser. But a credential is not evidence of hands-on management. The next badge makes that distinction more expensive.

The next badge: “Full-service management”

Verdict: unverifiable until you see the calendar and the account. Full-service can describe three different arrangements: a course and a Slack channel, a monthly audit deck you implement yourself, or somebody who actually changes the campaigns. What the deck calls synergy, I call two people doing the same job.

Ask what happens in week one, whose account holds the campaigns, who checks them afterward, and what you approve before spend rises. One done-for-you description offers a checkable sequence: confirm account and market fit; agree on a spend cap, success measure, and review date; set up in the client’s account; check conversion tracking before launch; and send weekly updates on spend, results, and changes. It also sets out a 30-day trial followed by a 90-day engagement. You can decide whether those terms suit you. At least you know what to inspect.

Then request a sample weekly action log with dates, changes, and reasons. A promise to monitor performance is not the same thing as a record of what changed when performance moved. No log, no evidence of management.

The technical badge: “AI-optimized bidding”

Verdict: unverifiable until the seller names a lever. The phrase sounds precise because it contains bidding. It tells you almost nothing about the work. ChatGPT targeting is described as live-conversation matching through advertiser-written context hints, alongside geo and custom audiences, rather than keywords, retargeting, or lookalikes. The auction is described as relevance-weighted second-price, with CPC or CPM buying and a suggested $3 to $5 CPC maximum.

So which part does the agency change? The bid, the geo, or the context hints? Better-written hints and tighter geo can put an ad beside more relevant conversations. The bid governs what the advertiser is willing to pay for that opportunity. Neither becomes a management service because someone adds AI-optimized to a page.

I want to see the previous hint, the rewritten hint, the reason for the change, and what happened next. If the seller cannot distinguish a bidding change from a targeting change, I would not pay for either. Ask what got rewritten last week, not what “the model” did.

The last badge: “Works with any budget”

Verdict: red flag without an approval gate. A low entry point makes this line easier to print, not easier to defend. The listed US minimum is $25 a day, roughly $750 a month per campaign; GBP and EUR minimums are 15 a day. That can buy delivery. It does not guarantee enough conversions to judge unit economics.

One published live test spent CA$198.58 for 4,561 impressions and 47 clicks: a 1.03% click-through rate, about CA$4.23 per click, and one conversion. With one conversion, the apparent cost per acquisition changes sharply if the next test produces zero or two. Small tests can check whether tracking fires and whether conversation matching looks plausible. They cannot, by themselves, settle the ROI question from the hero.

Say you spend $20,000 a month on Google and someone proposes a $750 ChatGPT test. I am fine with testing. I am not fine with we’ll see how it goes as the entire budget policy. Without a cap, a success measure, a review date, and a rule for pausing or increasing spend, a small test can drift into several months of invoices and a slide called early signals.

Set those rules in writing before launch. Include what happens if invalid clicks or irrelevant conversations appear. “Any budget” works only when someone can tell you when that budget stops or grows.

The three-step process: “Audit, launch, optimize”

Verdict: unverifiable until the steps name the setup work. The sequence is sensible. It is also broad enough to fit on almost any services page. I would look underneath each verb.

  • Audit: What account, conversion event, landing page, and success measure will the seller inspect?
  • Launch: Who sets the spend cap, approves the ads, and checks that a conversion records before money goes out?
  • Optimize: Who reviews spend, changes context hints or bids, checks lead quality, and records the reason for each change?

Tracking is where a polished three-step graphic can become a very expensive misunderstanding. The setup described for ChatGPT ads uses the OpenAI Pixel and Conversions API with the same event ID, preserves the oppref click reference through redirects, adds UTMs manually because oppref alone does not identify source and medium, and links the campaign to a conversion event. If the campaign is not linked to the event, the report can show zero. If the click reference or event data breaks, the CPA you use to judge the test is unreliable.

The landing page needs the same attention. A context hint can match a conversation, but the click still has to land on a page that answers the visitor’s question. I made that mistake with Google Ads often enough: good intent in the ad, generic page after the click. Do not pay to repeat it here.

There is also a risk the process graphic rarely mentions. A TrafficGuard analysis of 40,021 ChatGPT clicks across 47 advertisers reported invalid-click rates ranging from 0.1% to 34% by advertiser and a 3.6-times-higher hosting-and-proxy rate than Google. TrafficGuard sells verification, so I would read its analysis skeptically. I would still ask who checks click quality each week and what triggers a pause.

Finally, ads run separately from answers. Paying for an ad is not buying a citation. If the process blurs those outcomes, ask the seller to separate them before approving a test. No tracking plan, no landing-page plan, no launch.

The results strip: three logos, no dates

Verdict: red flag when the logos stand in for campaign evidence. This is where the hero should cash its cheque. A buyer needs to know what inventory ran, when it ran, how much the advertiser spent, what was measured, and whether the comparison uses the same advertiser. Three logos answer none of that.

A more inspectable example reports 36% lower CPC for its own ChatGPT ads versus Google, 29% lower median CPC for the same advertiser, and a 28% median CPC drop from the first to the latest month. The same description ties its reporting to Pixel and Conversions API setup, with spend shown beside calls, forms, and sales daily. Those figures do not promise you the same outcome. They give you measures and a setup to question, which is more useful than a logo wall.

I would still ask for the dated campaign and the exact definition of a sale or qualified lead. CPC is a cost measure, not the ROI promised at the top of the page. A result worth buying has a date, a denominator, and a conversion you care about.

The pricing section: a fee, or a request for a call

Verdict: red flag when the fee rises with spend. A missing public price is inconvenient; a fee structure that rewards premature scaling is the larger problem. One live list charges RM2,000 a month for one campaign up to RM5,000 for five, with media paid separately. That tells a buyer something concrete, even if it does not tell them whether the work is good.

A percentage-of-spend fee creates a different incentive. The person recommending a larger ChatGPT test also earns more when you approve it. On a young channel, I want the test to stay small until conversation match and qualified conversions justify more spend. I do not want the management bill pushing in the other direction.

That is why I prefer a flat fee. groas lists Search Ads, including Google, Bing, and ChatGPT, at $1,499 a month per service, with no cut of ad spend, no setup fee, and month-to-month cancellation. You can disagree with the price and still see why the structure fits this test better: the management fee stays put while the operator and buyer decide whether the evidence supports scaling. If the fee grows with spend, treat the scale recommendation as a financial interest, not just advice.

Five questions to ask before you sign

The page has now made its pitch in order. Before giving it access to an account or a budget, I would ask five questions in the same plain language:

  1. Proof: Can you show one dated ChatGPT campaign with spend, clicks, and the exact conversion counted?
  2. Work: Who changes the account each week, and can I see a sample log with dates, changes, and reasons?
  3. Levers: What did you change last: context hints, geo, or a bid? Why?
  4. Budget: What is the test cap, the pause rule, and the date we decide whether to scale or stop?
  5. Measurement: How are the Pixel, Conversions API, oppref, UTMs, and conversion event connected? Who checks landing-page fit and invalid clicks?

If two answers stay vague, I am probably talking to someone selling instruction rather than management. Instruction can have a price. It should not arrive disguised as a person who will run the account. Buy the work you can inspect.

Overall verdict: keep the controls, not the certainty

The composite page sells more certainty than its short channel history can support. Its hero needs dated conversion evidence. Its badges need a named operator and named levers. Its process needs tracking and landing-page checks. Its results need more than logos, and its price needs to leave the scale decision honest.

I would keep three things from it: work in the client’s own account, agree the cap and success measure before launch, and put spend beside calls, forms, and sales in regular updates. Those are useful commitments precisely because a buyer can check whether they happened. Then run the test small, read the conversions, and scale only when the CPA earns it. Until then, keep the cap.