Refunds and cancellations: the gap no ad platform subtracts

Do Meta and Google subtract refunds and cancellations from the sales they report, and how much does that inflate my ROAS?

Every ad platform learns about a sale at the moment it happens. What it does not learn, unless someone tells it, is that the order was refunded a week later, cancelled before shipping, or never paid at all. So the revenue and ROAS a platform reports are gross, and the gap between gross and what you actually kept is one of the largest and least visible reasons platform numbers run ahead of your store.

What each platform does with a refund

Meta counts the purchase event it received. A later refund or cancellation does not remove it from Ads Manager, so Meta's reported purchases, revenue and ROAS reflect gross sales.

Google Ads can subtract them, but only if you send the adjustment. With conversion adjustments, a conversion tracked with an order ID can be retracted completely or restated at a new value. Google recommends uploading adjustments at least 24 hours after the original conversion and within 7 days so automated bidding can use them, and accepts them within 54 days overall. In practice almost nobody sends them, so Google's numbers are gross too.

GA4 subtracts refunds only when your store sends a refund event, which most implementations do not.

Your store knows every refund, cancellation and unpaid order, because it processed them.

How much it inflates ROAS

The inflation is roughly your refund and cancellation rate on the revenue each platform credits. A worked illustration: a channel is credited with $100,000 of revenue on $25,000 of spend, a reported ROAS of 4.0. If $15,000 of that revenue is later refunded or cancelled, the settled revenue is $85,000 and the settled ROAS is 3.4. The platform still shows 4.0.

That 15% is before any other cause of inflation (self-attribution, view-through credit, overlapping claims), which stack on top. We list the rest in 8 things that inflate your reported ROAS.

Why it is not even across channels

If refunds were spread evenly, every channel would be overstated by the same share and the ranking would survive. It rarely is. Refunds and cancellations concentrate in some places:

• Promotions and sales events. Heavy discounts attract impulse purchases, and the returns arrive after the campaign report has been read.
• First-time buyers. New customers return more than repeat customers in most categories, so prospecting campaigns carry more of the refunds than retargeting.
• Payment methods paid later. Orders paid by bank transfer or cash are created before they are paid, and a share never are.
• Some categories and sizes. Apparel and footwear with size uncertainty return far more than consumables.

The result is a ranking of channels that can change once refunds are subtracted, which is exactly the decision the report was supposed to inform.

How to fix it

• Measure every channel on settled revenue. Take revenue per channel from your store after refunds, cancellations and unpaid orders, and compute ROAS from that. This is the single largest correction you can make to a platform report.
• Use one rule for every channel. Decide whether a sale counts when the order is created or when it is paid, and treat partial refunds as a reduction in value rather than a removed order.
• Send Google its adjustments. Upload retractions for refunded and cancelled orders and restatements for partial refunds, within the 7-day window where bidding can learn from them.
• Wait before judging a promotion. Read results two to four weeks after an event, once returns have settled, not the Monday after.
• Track refunds by campaign. A campaign with a high refund rate is buying the wrong customers, even if its reported ROAS is excellent.

Where this fits

Refunds are one of eight causes of the gap between platforms and store, all collected in why Meta, Google, GA4 and your store never match. For choosing the metric you judge channels on, see ROAS, POAS and MER. Adray applies one rule everywhere: a fully refunded or charged-back order is excluded from every count, on every surface, so a converted order cannot mean two different things in two places.