A $3,000 ChatGPT Ads test costs $4,499 with management, before internal work or excluded extras. In the scenario below, rejecting 45 of 60 leads turns a $74.98 raw lead into a $299.93 sales-accepted lead without spending another dollar.

That is the budget problem. Not the click price.

The first-test media range cited in the draft’s pricing guide is $3,000 to $5,000 per month. Add $1,499 per month for groas management, then account for anything excluded from the engagement. The $4,499 example uses the bottom of that media range. Its lead counts are scenario inputs, not a performance promise.

I will not call a ChatGPT Ads budget defensible until media, management, and rejected leads appear in the same calculation. The announced Amazon Ads pilot in ChatGPT adds a buying route to investigate. It does not establish universal access or cheaper acquisition.

The listed costs: what gets billed, and when

Before debating CPC, I want the invoice structure on one page. Media and management are separate bills; conditional extras are not automatically part of either.

The table separates the test budget from alternative management models and possible additions. Do not add every row together. A percentage-of-spend manager and a flat retainer are alternatives, not two charges for the same work.

Line itemListed cost or rangeTimingSource and cost drivers
Test media$3,000 to $5,000 per monthRecurring during the testChatGPT first-test sizing guide; actual spend depends on the approved cap and buying route
groas management$1,499 per month; media excludedRecurringgroas management breakdown; confirm the engagement’s channel scope
Percentage-of-spend management alternative10% to 20% of mediaRecurringAgency pricing breakdown; the invoice rises with spend
Flat-retainer alternative$1,000 to $5,000 per monthRecurringAgency pricing breakdown; account complexity affects the quote
Agency setup, if charged$500 to $3,000One-offAgency pricing breakdown; may be billed separately from management
Call tracking, if excluded$30 to $150 per monthRecurringAgency cost breakdown; numbers, recording, and integrations affect cost
Launch creative, if excluded$3,000 to $8,000 upfront, plus $1,000 to $2,000 per monthOne-off and recurringChatGPT pilot sizing guide; confirm what the quoted package actually requires

Those are listed costs, not a universal price card. Access, scope, and exclusions still need written confirmation. A vendor who puts the attractive management number on page one and the required extras on page three has not made the test cheaper. Just harder to price.

Approve the complete quote, not the headline fee.

Cost line 1: Confirm the route before pricing the traffic

I price ChatGPT access by route because there is no single front door. For direct access, confirm account eligibility before building a budget around it.

The other route is Amazon’s announced pilot with OpenAI, starting Sept. 10, 2026, for select U.S. advertisers buying through Amazon DSP. Delta Vacations is named as the first tester, with OpenAI retaining control of delivery decisions. The pilot is described as invite-based managed service, so the next step is contacting an Amazon Ads representative, not assuming a campaign can go live.

For ecommerce, units can auto-generate from an existing Amazon catalog. That makes catalog accuracy part of launch preparation. An automated ad does not repair a bad product entry.

The pilot inventory is described as U.S.-only at launch, sold on CPC and CPM, with labeled text and image units beneath answers for Free and cheaper Go tiers. Those boundaries belong in the plan. They are not evidence of the audience coverage your business needs.

On the direct side, the published pricing guide lists a $25 daily U.S. minimum per campaign and a $3 to $5 starting maximum CPC. A maximum bid is not an observed acquisition cost. Neither is a minimum budget a recommendation for how much learning you can buy.

Before modeling a lead price, get four things in writing:

  • Route and eligibility: how the account gains access.
  • Pricing basis: what triggers a charge.
  • Delivery limits: where and to whom ads can appear.
  • Measurement: what the test can attribute and report.

The Amazon announcement gives me another door to knock on. It does not give me a cheaper customer.

A toll booth before a chat window, with separate entry-price signs

Cost line 2: Keep management visible, even when it covers other channels

Every test has a media bill and someone responsible for managing it. groas lists management at $1,499 per month, with media excluded. That fee can also cover Google and Bing in the same engagement, so the approval sheet needs to show how the shared fee is treated.

Do not let ChatGPT look free because Google absorbs the entire retainer. Equally, do not charge the same retainer in full to every channel and then call the total your actual cash exposure. Show the invoice once and make any channel allocation explicit.

I prefer the flat structure because percentage-of-spend pricing rewards a bigger media budget. At 10% to 20% of spend, raising the budget raises the management invoice. Cutting waste can lower it. That is an awkward incentive for someone hired to make spending more efficient.

Do not pay twice for work already in scope

The easiest proposal padding to miss is double counting. Tracking setup, landing-page support, or call recording appears as an extra even though the statement of work already covers it.

The fix is not to assume every service is included in management. Mark an item “included” only when the written scope says so. Otherwise, enter the vendor’s quote or resolve the exclusion before approval.

Use the same discipline when comparing groas with an agency that bills setup, landing pages, and call tracking separately. A low management quote is not a low total if the work required to launch sits outside it.

One management line. Clear inclusions. Quoted exclusions. No invoice archaeology after launch.

Cost line 3: Cap media instead of pretending to forecast it

I do not approve a test from a quoted market-average CPM. The campaign buys your audience, your prompts, and your test window. An average does not promise what that combination returns.

The first-test guide’s $3,000 to $5,000 monthly media range is a planning reference, not proof that a particular budget produces enough qualified leads. It also excludes management and creative.

For this illustration, I use a $3,000 media cap over 30 days, with written approval required to raise it. That turns the budget into a limit rather than a prediction. The test then has one job: show what that fixed exposure returns after qualification.

A cheap CPC can help. It cannot rescue traffic that never becomes a sales-accepted lead. Approve a cap the business can afford to lose, not a forecast dressed up as certainty.

The hidden work: approvals, tracking, and sales review

The vendor’s invoice is not the entire cost of running the test. Someone grants access. Someone approves copy. Someone checks tracking. Someone works through the names and decides which ones sales will accept.

I record actual incremental internal cost: hours spent multiplied by loaded pay, plus tools bought specifically for the test. I do not substitute agency rates for employee time or invent a large shadow salary to make the spreadsheet look sophisticated.

Separate launch work from recurring work:

  • Launch: account access, tracking checks, initial approvals, and any test-specific tool setup.
  • Recurring: lead review, follow-up, ongoing approvals, and qualification reporting.

Then name the owners. Sales owns lead acceptance and follow-up. The business owns access and approvals. Any tools or services outside the management scope need their own line.

Some of this work may fit inside existing capacity. Some may create a real extra expense. Record what happens rather than inserting a reassuring allowance. The $4,499 example remains media plus management; internal work sits beside it until measured.

Rejected leads change the denominator, not the invoice

Before buying traffic, I want sales to agree on three definitions:

  1. Valid lead: real, reachable contact information.
  2. Qualified lead: a lead sales accepts against the agreed budget, fit, and timeline rules.
  3. Customer: money in the bank.

These are different outcomes. Counting all three as leads is how a performance report becomes an argument.

A broad lead-quality source reports that only 27% of leads sent to sales are qualified. That is context, not a ChatGPT forecast. Your acceptance rate must come from your own rules and results.

Here is the standalone illustration I would put in front of finance:

MeasureScenario input or calculation
Media$3,000
Management$1,499
Media plus management$4,499
Raw leads60
Sales-accepted leads15
Cost per raw lead$4,499 ÷ 60 = $74.98
Cost per qualified lead$4,499 ÷ 15 = $299.93

The 60 raw leads and 15 accepted leads are assumptions, not observed ChatGPT benchmarks. Replace them with actual lead volume and acceptance counts. Add internal costs and excluded expenses when calculating the fuller cost.

Rejections do not create a second media charge. They leave fewer useful outcomes to divide the original bill by. That distinction matters: adding a made-up “rejection cost” would double-count the damage.

Put raw-lead cost and qualified-lead cost next to each other. Otherwise, marketing celebrates the first while sales has to live with the second.

Cheap clicks are not evidence of cheap acquisition

One self-funded B2B test reported $2,319 spent on 739 clicks at $3.14 CPC, producing zero MQLs. The same advertiser’s Google Search activity produced 2 MQLs at $6.76 CPC.

That is one test, not a universal verdict. It still exposes the problem with celebrating click price: the cheaper click did not produce the outcome being counted.

The same discipline applies to the buying route. Amazon DSP access may change how an eligible advertiser buys inventory. The announcement alone says nothing about that advertiser’s eventual qualified-lead cost.

Once the acceptance rate is known, the next calculation is whether customers cover the outlay. The groas breakdown of a $2,499-style test tackles that question. Do not skip qualification on the way to a customer forecast.

The one-page approval memo: spend, acceptance, and a stopping rule

If the test cannot fit on one page, the budget is not ready. I want the owner, sales, and finance working from the same sheet before media runs. No promised CPL. Just what gets spent, what counts, and what happens next.

  • Route: direct access or the Amazon DSP pilot, with eligibility confirmed and the relevant contact attached.
  • Access dependency: account approval and working measurement before any spend.
  • Cash exposure: $3,000 media cap plus $1,499 management equals $4,499, with excluded extras and internal work shown separately.
  • Qualification rule: sales accepts or rejects each raw lead within 48 hours against the agreed budget, fit, and timeline criteria.
  • Baseline: current-channel cost per qualified lead, using the same cost scope and qualification rule.
  • Decision rule: continue only if qualified-lead cost beats the baseline; pause for review if there are zero accepted leads after 30 raw leads have been assessed; never raise the cap without written approval.

Compare like with like or the test lies to you. Cost per qualified lead divides the relevant costs by accepted leads. Cost per customer divides them by customers. They are not interchangeable, and neither should be compared casually with a media-only lead benchmark.

Run current Google or Meta results through the same sieve: the same included costs, acceptance window, and lag to closed-won. A ChatGPT lead that closes in 45 days does not beat a Google lead that closes in 14 days merely because the form was cheaper.

The cheapest mistake here is spending a week arguing about a $25 daily minimum. The most expensive mistake is approving $3,000 in media and $1,499 in management without writing down that, in this scenario, 60 names at $74.98 each become 15 keepers at $299.93 each.

Put the rejection rate in the memo and the test can teach you something. Leave it out and every report becomes a negotiation.