One store owner, $1,500 in media, 30 days, and no more than three hero SKUs. I would not load the catalogue into ChatGPT ads: at roughly $50 a day, spreading spend across dozens of products buys anecdotes, not a decision.

I learned that lesson triaging small ecommerce Google accounts. One audit of a 150,000-SKU shop found 99% of SKUs burned 75% of budget for 5% of revenue. A new channel does not cure that habit. ChatGPT ads appear as sponsored cards below the answer, never inside it, so the job is to put a relevant product beneath a question a buyer is already asking. With US minimums at $25 per day per campaign and early CPCs running $2 to $4, three products have a chance to collect readable signals. Thirty do not.

The plan below treats $1,500 as a ceiling, not a dare. Some stores should spend none of it. Some should stop after the first tranche. The point of day 30 is to know what earned another dollar.

Step 0: Pass the skip check ($0 in media; buys you a reason to test)

This test is wrong for plenty of stores that ask me about it. If your contribution margin after product, shipping, fees and refunds is under 30%, skip ChatGPT ads for now. Break-even ROAS is 1 divided by contribution margin. At 40% margin, you need 2.5x just to break even on the first order; at 25%, you need 4x. One breakdown puts the small-basket problem bluntly: at 40% margin and a 4% conversion rate, a $4 click only pays back if the basket is around $250.

WordPress vendor Barn2 learned how unforgiving those clicks can be, spending $477 over six weeks for 19,154 impressions, 370 clicks and zero conversions. I do not cite that to scare you. Its budget was spread across four products with retail prices too low to absorb a $4 to $9 click. That looks a lot like the first tests owners propose to me.

Before you build anything, check three basics:

  • Is there a product over $150 with room for acquisition cost?
  • Is a purchase event tracked, or can you get one working before launch?
  • Does that product already sell without discounts?

If the answer is no, keep the $1,500. A sponsored card cannot repair the unit economics, and clicks without purchase tracking will not tell you whether the test worked. The tighter 35% margin filter for individual SKUs comes next; 30% is the point where I would skip the channel test altogether.

Step 1, days 1–3: Choose three comparison-worthy SKUs ($0 in media; buys you a readable test)

Rank the catalogue by contribution dollars per unit, not revenue. Then ask of each candidate: Would someone ask ChatGPT to compare this before buying? Supplements versus competitors, mattresses by firmness, espresso machines by grinder type. That comparison question gives a sponsored card beneath the answer a job. A commodity SKU with no comparison question belongs in Google Shopping, not in this test.

Do the break-even math for each candidate before choosing your three. Subtract product, shipping, fees and refunds from the selling price to find contribution dollars per order. Multiply that by your store conversion rate to estimate the allowable cost per click. Say a $280 item has a 40% margin: that is $112 in contribution. At a 3% conversion rate, $112 × 0.03 gives you $3.36 per click. A $4 to $9 click fails that test before you write an ad.

Compare that with the $420 basket in one DTC account that produced a $93 CPA and 4.5x ROAS. A larger basket has more room to absorb acquisition cost. That is a reason to check your own numbers, not a promise that your store will get the same result.

My entry floor for each of the three SKUs is simple:

  1. At least $150 in selling price.
  2. At least 35% contribution margin after the costs above.
  3. At least 20 sales in the last 60 days.
  4. A comparison question you can answer clearly on its product page.

The sales floor keeps a product with no existing demand from borrowing the test budget. The margin floor keeps a popular but thin-margin item from winning on clicks while losing on orders. If only two products qualify, test two. Do not promote a weak third SKU to fill a slot; no more than three is the constraint.

Step 2, days 4–7: Fix tracking and pages before launch ($0 in media; buys you attribution)

Install tracking before you build the campaign. Shopify is OpenAI’s first ecommerce integration, and the free Shopify app syncs the catalogue to Ads Manager and wires up purchase tracking. Use it, then add the server-side layer: the browser OpenAI pixel and Conversions API should share one event ID so retries do not double-count.

For a standard Shopify setup, check the whole path, not just the final sale: page view, product view, add to cart, checkout started, purchase. Server-side delivery matters when a browser blocks the pixel. More importantly, a purchase number you cannot reconcile is a poor basis for spending the second tranche. Before launch, make sure the events arrive and that the same purchase is not counted twice. This is setup work, not media spend; $0 here does not mean it takes no time.

Then fix one page per SKU. I mean one useful product page, not a redesign project. Each page should answer the comparison question that could put its ad beneath a ChatGPT answer. Put price, what the product replaces, who should not buy it, shipping and returns above the fold. Give the buyer at least three photos and reviews with dates.

I used to tell clients a good feed could carry a weak page. I was wrong. A buyer who arrives from a comparison answer has already begun weighing alternatives. If the page makes that buyer hunt for the price or the tradeoff, the click has nowhere useful to go. Do not save setup time by paying to send people to a thin page.

Three boxed products lit on a small shop shelf beneath a $1,500 budget label; a large warehouse of boxes remains dark behind them

Step 3, days 8–21: Buy the first signal ($700 in media; preserves up to $800)

Run one campaign with three ad groups, one per SKU. Set the campaign to $50 a day for these 14 days. That is $700 in the first tranche and leaves up to $800 under the cap. One campaign respects the $25-a-day minimum per campaign; three separate campaigns would starve on this budget.

Open at a $3 to $5 max CPC, where OpenAI tells new advertisers to start, and expect $25 to $60 CPMs while the auction learns. Self-serve has had no minimum commitment since May 5, 2026. For this plan, the constraint is the media budget, not a contract.

Check three things each day: spend pacing, CPC by SKU, and add-to-cart rate. Look at purchases, of course, but do not let a day’s ROAS or CPA swing the whole plan. The practitioner rule to budget for 50 conversions in the learning window shows the problem: $1,500 cannot buy 50 orders if CPA lands near the $93 seen in one DTC account spending $40k+. Google’s Learning Limited warning points at the same volume problem. A small test needs a modest job: find a product worth testing further, not declare a stable long-run ROAS.

Keep a simple daily record by SKU. Did it spend? What did a click cost? Did clicks lead to carts? A 1.5% CTR, a CPC below that SKU’s allowable CPC, and carts coming through are reasons to keep watching. The allowable CPC matters more than an attractive-looking average: a $3.50 click is fine for one product and already too expensive for the $280 example above.

Do not give a SKU a second tranche because its CTR looks good in isolation. If it cannot buy clicks inside its own economics, stop feeding it. The first $700 is there to narrow the field, not to make all three products feel equally supported.

Step 4, days 22–30: Kill, hold or scale (up to $800 in media; buys you a decision)

By day 22, you have spent about $700 if the first tranche ran as planned. Now move money toward the products that earned it. One winner can get $60 a day. If there are two viable SKUs, split daily spend $40 and $20 according to add-to-cart rate. If none qualifies, spend nothing more. The remaining $800 is a cap, not a target you must exhaust. Those daily allocations may leave money unspent, and that is better than buying extra evidence for a product you have already ruled out.

This is the part owners resist. Killing a product after roughly $230 in spend feels like quitting. It is not. It is why you limited the first test to three products instead of inviting the whole catalogue to eat.

  • Kill: After more than 3,000 impressions, CTR remains under 0.8%; or the SKU spends past 1.5× its break-even CPA without a purchase. Cut it and do not revisit it in this test.
  • Hold: Clicks cost less than the SKU’s allowable CPC and carts are coming in, but CPA is still above break-even. Keep it at $20 a day through day 30. Read the cart signal, but do not call this a profitable winner.
  • Scale: CPA is at or under break-even and ROAS is above the 1 divided by margin threshold. Give that SKU the $60-a-day allocation and a shot at month two.

My line for month two is one SKU with at least five purchases at or below its break-even CPA. Fewer than five is a hold, not a win. Five orders do not prove ROAS stability either; I would not pretend they do. They show that a product can turn some of these clicks into purchases without already breaking its first-order economics.

Watch the landing page alongside that decision. Barn2’s ad visitors stayed 37 seconds versus 443 seconds for organic ChatGPT referrals. If your own visitors disappear quickly, inspect the page and the match between ad and comparison question before calling the channel hopeless. Do not use page time to excuse a losing CPA. Use it to decide what to fix next.

Step 5: If the cap falls to $750, cut products first ($750 maximum; buys you one cleaner read)

The $1,500 plan does not require spending $1,500. If the available media budget falls to $750, I would narrow the test rather than run a weaker version of all three ad groups. At $25 a day for 30 days, use one campaign, one ad group, one SKU, one landing page. Do the tracking and page work before that 30-day media run starts.

At a $3 to $4 click, $750 buys roughly 190 to 250 clicks. Split that three ways and you get about 60 to 80 clicks per product: enough to confirm that clicks cost money, not much more. Put those clicks behind the single SKU that best combines margin with comparison intent. Then you can read its CTR, CPC and add-to-cart rate at day 30 without wondering which product diluted the result.

Cut the two-tranche structure too. There is no useful $700 first tranche and $50 finale. Run $25 a day straight through, with a kill-or-keep check on day 15. If the SKU fails its allowable CPC or produces no meaningful cart signal, stopping early is still a decision. It also means you do not have to spend the whole $750.

What I would not cut is the pixel and Conversions API setup, or the comparison page. Skipping either to save four hours can turn $750 into a screenshot of clicks with no trustworthy purchases attached. If you have one weekend, spend Saturday on the event ID match and the page, launch Monday, and live with a single-SKU test. Smaller budget means a narrower test, not thinner setup.

Step 6: When the cap lifts, decide who runs the next pod ($0 from this test; buys you a scaling plan)

If the 30 days produce one winner and you are spending an hour a day babysitting bids, feed syncs and comparison pages, you have outgrown the DIY part. That was my tell in small ecommerce accounts: the test is owner work; scaling is operator work. Feeds break, CPCs drift, and each new SKU needs its own break-even math and page. What took a few hours a week can start taking a few hours a day.

I wrote about when a founder should stop running Google Ads and hire management. The signs transfer here: tracking you no longer trust, a winner you cannot scale without breaking CPA, and no time to build the next comparison pages.

That is where I put groas, and I am blunt about why. The alternatives I usually see are an hourly freelancer learning a new auction on your budget or a percentage-of-spend agency that earns more when you spend more. groas runs Search Ads across Google, Bing and ChatGPT for a $1,499 flat monthly fee, with no cut of media spend. Bids, budgets and landing pages get worked around the clock, and changes are logged with reasoning. That fee is separate from the $1,500 media cap; it belongs in the conversation after this constrained test, not hidden inside it.

When the cap lifts and you move from three SKUs toward ten, you do not need another dashboard. You need the losing products cut without waiting for a weekly call. Take the winner into that model, or keep running it yourself at $50 a day if you have the hours. Just do not hand a 30-SKU catalogue to someone who charges you more when CPCs rise.

Month two is not the month you finally load the catalogue. Run the same test with the next two SKUs that pass the margin and comparison filter, alongside the month-one winner. Keep the three-product pods and let losing products stop consuming budget. Lifting the spending constraint is not a reason to lift the discipline.

You have one job on day 31: name the SKU that earned month two, name the CPA it earned it at, and name what you killed. If you can say all three in one sentence, the test did its job. If you cannot, narrow the next run to one SKU and $750 before you pay anyone to manage it.