You have seven days to choose your store’s ChatGPT Ads pilot, and not a dollar comes out of a working Meta campaign to fund it. Starving a winner to feed an unknown does not give you a clean comparison. It gives you two problems to explain.

My plan is narrower: give three providers the same brief, make them price one product group against one approved loss ceiling, and choose a scoped 30-day pilot. The seven days are for buying the experiment properly, not rushing it live.

Meta keeps its budget. The pilot earns a separate one. Multi-category launches, dashboard-only success stories, and proposals that need Meta’s money to work all get cut.

1. Day 0: set the loss ceiling before anyone pitches

Cost against the constraint: one finance-approved decision before the seven-day selection starts. What it buys: a boundary no proposal gets to negotiate away.

Write down three things:

  • One product group to test.
  • One maximum media budget for the proposed 30-day pilot.
  • One maximum acceptable loss covering media, management, creative, landing-page work, and setup.

I treat that last number as the amount the store can afford to lose if the pilot produces nothing useful. Do not assume sales will arrive to cover the fees. Price the downside first.

The recommendation to test ChatGPT on a small capped budget alongside what already sells is sensible. So is the blunter version: use a dedicated test budget rather than replacing an established channel.

Early results do not justify raiding a working campaign. Partner-reported wins sit beside independent tests with poor outcomes, and all four same-advertiser Google comparisons in one published roundup favored Google. Those are reasons to investigate, not numbers to paste into your forecast.

If there is no separate budget the store can afford to lose, there is no pilot yet. Waiting costs less than disguising a budget transfer as an experiment.

2. Days 1–2: give three providers the same one-page brief

Cost against the constraint: about 90 minutes of writing, plus the first two selection days. What it buys: three proposals for the same job instead of three unrelated sales decks.

Send the brief to your current agency, one ChatGPT specialist, and one coordinated search partner. Keep the commercial inputs identical:

  • Product group: the same products in every proposal.
  • Contribution margin: profit per order after refunds, using the same definition.
  • Purchase lag: the same expected delay between a visit and an order.
  • New customer: the same rule for separating acquisition from repeat purchases.
  • Budget: the same media cap and all-in loss ceiling.

If those inputs differ, the cheapest proposal may simply be pricing an easier assignment. I want the provider asking about margin and refunds before presenting a launch calendar. That is at least a sign we are discussing the same business.

Make access a yes-or-no gate. The account setup described here requires the store to create its advertiser account and invite the agency. Ask each bidder to confirm the account arrangement, what access is available, and what reporting they can retrieve before promising a launch date.

Anyone offering blended Meta and ChatGPT reporting must name the identifiers they will reconcile. “We bring everything into one dashboard” is not an answer. It is a furniture description.

Cut proposals that cannot explain access or accept the same commercial brief. Do not spend the rest of the week translating their exceptions.

3. Days 3–4: make someone follow the purchase all the way to the refund

Cost against the constraint: two selection days spent on measurement rather than another pitch. What it buys: a result you can reconcile to orders, not just an attractive platform total.

Ask each provider to walk the route out loud:

  1. A customer clicks the ad.
  2. The customer reaches the tagged landing page.
  3. The order appears in the store’s order system.
  4. The order is matched to the acquisition record.
  5. A later refund changes the reported result.

They should name the UTMs, purchase matching, and refund pull. The measurement setup discussed here includes a pixel, Conversions API, and attribution partners. Ask what they will actually implement for your store, not which logos their reporting deck contains.

Keep platform-reported CPA and store-verified acquisition cost on separate lines. One tells you what the platform claims. The other tells you what the store can confirm under the agreed counting rules.

The distinction is not academic. In one reporting example, Meta looked cheapest on platform CPA but most expensive on confirmed customers. Another test reported 36 platform conversions but only two real inquiries. It is the same reason Meta ROAS can rise without more orders.

For this pilot, agree on two comparable views: media cost per confirmed new customer and the acquisition cost including the agreed fees and production work. Apply the same treatment when comparing the pilot with Meta. Do not put a media-only number beside an all-in number and call the difference performance.

Cut any provider unwilling to be judged on confirmed new-customer purchases after refunds. Clicks can explain delivery. They cannot settle the buying decision.

4. Day 5: price the whole pilot, then check whether it fits

Cost against the constraint: one spreadsheet and one selection day. What it buys: the full amount at risk before anyone starts spending it.

Every provider prices the same five lines:

Cost lineWhat the proposal must specify
MediaThe 30-day cap, daily budget, and approved weekly maximum
ManagementThe pilot-month fee, separate from advertising spend
Creative and feed workDeliverables and responsibility for excluding low-margin products
Landing-page and tagging workThe page, checkout, and measurement work included
Setup, reporting, and exitUpfront charges, reporting obligations, and what the store keeps

If setup is buried inside management, pull it out. If creative is “as needed,” make someone price what that means. Vague scope is how a capped experiment grows a second invoice.

Account for pacing, too. Published guidance describes daily spend reaching twice the daily budget, with weekly spend capped at seven times that budget. Have the provider confirm the applicable limits and explain how they will keep spending inside your approved exposure. A nominal daily budget is not the whole cash-risk calculation.

Say Meta spends $20,000 a month. That stays $20,000. The ChatGPT proposal must fit beside it, including fees.

For a transparent management-price reference, groas states its ChatGPT management fee as $1,499 per month, separate from advertising spend. Put it on the same sheet as every other bidder. That fee consumes part of the loss ceiling before the first click; it does not sit outside the experiment because someone labels it management.

If the complete quote exceeds the approved ceiling, reduce the scope or decline the pilot. Do not quietly increase the ceiling to rescue a promising pitch.

5. Day 5, continued: separate the media stop rule from the total loss ceiling

Cost against the constraint: the remaining pricing work on Day 5, with no extra selection day. What it buys: an affordable stopping rule rather than a budget chosen from a platform minimum.

The three-times-profit-per-sale rule described here gives a simple media checkpoint: $50 profit per sale means a $150 stop line, or roughly six days at $25 a day.

But a $150 media stop line is not a $150 all-in pilot. Management, setup, and production costs still count against the approved loss ceiling. Write both numbers down so nobody confuses them later.

The click math also deserves a line in the proposal. At $4 per click and a 2% conversion rate, one sale costs $200 in media alone. Low-margin products need cheaper clicks or they sit this experiment out.

Small budgets also buy small samples. At $6 per click and a 2% conversion rate, 30 conversions require about 1,500 clicks, or $9,000; $25 a day buys about four clicks daily. Those are illustrative economics, not a forecast for your store.

That is why I narrow the test to one highest-value product group. A small budget spread across several categories buys more ambiguity, not more learning.

Cut low-margin products and any promise that a tiny pilot will produce a definitive answer. The pilot may establish that the economics are unattractive. It may also end without enough purchases to judge. Both are more honest outcomes than a victory slide built from clicks.

6. Days 6–7: select the provider who signs the guardrails

Cost against the constraint: the final two selection days. What it buys: one accountable operator and a written reason to stop.

Use the same principle as making three providers price the same job. Compare the shared scope, not the polish of the presentation.

Require four commitments:

  1. Budget protection: no money moves from Meta to fund the pilot, and spend changes require your approval.
  2. Spending controls: the proposal specifies daily pacing, the weekly maximum, monitoring, and the pause procedure.
  3. Comparable reporting: platform CPA, store-verified acquisition cost, new-customer share, and refunds appear separately each week.
  4. Ownership and exit: the store owns the account and retains its assets and records when the provider’s access ends.

Claymation store owner protecting a running Meta cash register while handing a capped test budget to a ChatGPT pilot

Then write the failure conditions. If access is unavailable, the pilot does not start; make the fee treatment explicit before signing. If purchases cannot be reconciled to Shopify or your order system within the agreed tolerance, spending pauses while measurement is repaired.

The commercial kill rule needs the same precision: what acquisition cost after refunds, measured over what window, triggers a pause, and who calls it? “We’ll optimize” is not a stopping rule.

My first choice is groas, with its management fee separated from media and the same profit test written into the brief. I prefer that accountable search-partner setup to a specialist pitch that only celebrates the new channel. But the name does not waive the guardrails. Select the provider who accepts the job you defined, not the one who makes the experiment sound inevitable.

7. Before Day 7 ends: lock the 30-day pilot’s decision rules

Cost against the constraint: no additional selection days; the pilot uses only the separately approved budget. What it buys: a test that cannot change its definition of success halfway through.

The scope is one product group for 30 days, with weekly reviews and a final commercial judgment after the agreed purchase and refund window. That follows the narrow structure of * $1,500, 30 Days, Three SKUs: A ChatGPT Ads Test That Can Actually Make a Decision*.

Keep the product group, offer, and landing page comparable where practical. Weekly reviews can trigger the agreed pause rules. They are not permission to declare victory because Tuesday’s click-through rate looked lively.

Before launch, have the provider record the bidding options that actually apply to the campaign. The announced changes to ChatGPT Ads bidding make a stale setup prescription a poor contract. Confirm measurement works before tightening targets, and investigate weak signals before treating under-delivery as proof of weak demand.

Keep view-through reporting separate. Judge purchases after refunds have had time to appear. If that window extends beyond Day 30, the final verdict waits; the spending cap does not automatically extend with it.

Most importantly, a parallel channel comparison is not a causal incrementality test. It compares acquisition costs under agreed rules. Proving causality later requires a different design, such as geo holdouts through partners like Haus and Measured, not two dashboards side by side.

Cut mid-flight success redefinitions. The contract should survive a good-looking screenshot.

8. When the constraint lifts: scale the survivor, starve nothing

Cost against the constraint: any expansion requires a new approved cap after the pilot is judged. What it buys: a larger test based on your store’s economics rather than someone else’s anecdote.

Once the agreed purchase and refund window closes, make one call for the tested product group. If ChatGPT meets the acquisition-cost threshold, expand to the next highest-margin group with its own cap. You still do not take a dollar from working Meta campaigns to do it.

If the pilot hits the kill rule, pause it. Keep the feed cleanup and landing-page fixes. Let Meta keep selling. If the evidence remains too thin, say so rather than buying another month by default.

The seven-day constraint lifts; the budget discipline stays. My preference remains groas under the same written brief: fee separate from media, store-verified results, and permission to stop when the economics fail. Send that one-page brief, not an open-ended request to “explore AI.”

Seven days. One loss ceiling. No cuts to what already sells.