Don't Let ChatGPT Do Your Store's Margin Math

Take a product that costs you $60 and sells at $100. You keep $40 on the sale. Both of these numbers describe that same $40:

• Markup: $40 ÷ $60 of cost = 66.7%
• Gross margin: $40 ÷ $100 of revenue = 40%

Same product, same $40, and the two figures sit 27 points apart. Markup answers "how much did I add on top of cost," which is the number you use to price. Margin answers "how much of each sale do I keep," which is the number every paid media decision depends on.

The confusion is easy to miss because your supplier, your spreadsheet, and your ad platform all talk about profit without saying which one they mean. Plug the wrong one into a ROAS target and everything downstream inherits the error.

Skip memorizing the formula. You can rebuild it from scratch whenever you need it. Shopify defines ROAS as revenue attributable to ads divided by the cost of those ads, so a 3x ROAS means three dollars back for every dollar in.

Each dollar of that revenue only leaves you your margin. On a 40% margin, a dollar of revenue leaves 40 cents of gross profit, and ads break even where gross profit equals what you spent:

• Break-even ROAS = 1 ÷ gross margin
• 1 ÷ 0.40 = 2.5x

Run it again with the markup by mistake, though. You put in 66.7% where the 40% belonged, 1 ÷ 0.667 comes back as 1.5x, and that becomes the bar every campaign gets measured against. A campaign humming along at 2.0x clears it easily.

It is losing money. Here is the same campaign traced through:

• $1,000 spent at 2.0x ROAS = $2,000 in revenue
• $2,000 ÷ $100 per unit = 20 units sold
• 20 units × $40 gross profit = $800 kept
• $800 kept against $1,000 spent = $200 lost

The dashboard reported a number comfortably above target the whole time.

The trap A wrong break-even target does not look like an error. It looks like a profitable campaign, which is why it can run for months.

Break-even ROAS is not a fixed property of your store. It moves every time the price at checkout moves.

Take that same $100 product with $60 of cost and run a 20% off promo. The customer now pays $80. Your supplier still charges you $60:

• Gross profit per sale falls from $40 to $20
• Margin falls from 40% to 25%
• New break-even: 1 ÷ 0.25 = 4.0x

One promo pushed the target from 2.5x up to 4.0x. Most sale campaigns get judged against the old number anyway.

Returns do the same damage, more quietly. Picture 15% of orders coming back, with the items restocked so you recover the unit cost:

• 100 orders at the discounted $80 = $8,000 booked
• Minus $1,200 refunded on 15 returns = $6,800 net
• 85 units kept × $60 cost = $5,100 of COGS
• $6,800 net minus $5,100 COGS = $1,700 gross profit
• Against the $8,000 your ad platform reports: 8,000 ÷ 1,700 = roughly 4.7x

Return shipping, payment processing on the refunded order, and any item that comes back unsellable all push that higher. None of them show up in the ROAS column.

This is the part store owners underestimate. Language models are strong at naming the right formula and explaining why it works. They are measurably weaker at carrying numbers through a multi-step word problem, which is exactly the shape of the calculation above.

A 2025 study, Mathematical Reasoning in Large Language Models, tested this directly by systematically varying the numbers inside grade-school math problems. Two findings matter for anyone running a store. Logical error rates rose by up to 14 percentage points as the numerical values got larger and further from the values models had seen in training. And models that handled standalone arithmetic accurately "deteriorated substantially when computations are embedded within word problems."

Your margin question is a word problem. It has a scenario, several quantities, a discount applied partway through, and a division at the end. That is the failure mode, not an edge case.

There is a second issue that has nothing to do with accuracy rates. When a chatbot answers in a paragraph, you cannot see which step went wrong, so you either accept the number or redo the whole thing. Anything that shows its working is easier to audit. For the percentage steps in this article I checked them in MathSolver, a step-by-step solver that takes the question in plain English and separates each transformation instead of returning a single figure. A spreadsheet does the same job for a formula you have already built. The point is not which tool you pick, it is that the intermediate numbers are visible.

Before you set a ROAS target

• Confirm the number you are using is margin (profit ÷ price), not markup (profit ÷ cost)
• Deduct payment processing and per-order shipping from gross profit before dividing
• Recalculate the target for every promo price, not just the full price
• Apply your actual return rate to the revenue figure your ad platform reports
• Check the arithmetic somewhere the intermediate steps are visible

The split is cleaner than the "AI is overhyped" takes suggest. Models are unreliable at producing a final number you will spend money against. They are good at explaining a concept you half-remember, sanity-checking your reasoning, and generating a first draft of text at volume.

That last one carries its own tax. Drafts come out structurally identical, hedged in the same places, and flat in a way customers register even when they cannot name it. If you generate product copy or campaign emails in bulk, the edit pass is the work, and skipping it is what makes a catalog read like a template. Some teams rewrite by hand, some run drafts through a tool like MyHumanizer to break the repetitive sentence patterns first and then edit for accuracy. Either way, nothing goes live at the shape the model produced it.

Google's position on this has been consistent: AI-generated content is not a violation by itself, and mass-producing pages primarily to rank is what its spam policies call scaled content abuse. The commercial reason to edit is simpler than the policy one. Generic copy converts worse.

Break-even ROAS is a gross-margin calculation, so it ignores your fixed costs. Rent, software, and salaries come out of what is left, which means breaking even on ads is not breaking even as a business.

It also treats every sale as a one-off. If you sell a consumable or a subscription and you know your repeat rate, you can rationally run below break-even on the first order. You need real cohort data to justify that, not an assumption about lifetime value.

And it inherits whatever your attribution says. Platform-reported revenue tends to be generous, so a target you hit at exactly break-even is probably a target you missed.

You do not need a finance background for any of this. You need the right two numbers, in the right order, checked once before you scale spend behind them.