Amazon DSP can credit a purchase to an ad up to 14 days after the shopper saw or clicked it, so the sales, ROAS and ACOS for the last 14 days can still change. A line item judged on this week's figures is judged on part of its sales. Judge it on days more than 14 days old.
Spend is known within a day. The sales it bought keep arriving for two weeks.
Why last week's numbers are still moving
Amazon credits a DSP purchase to the last ad before it, and looks back up to 14 days for that ad. Its 2026 attribution update keeps the 14-day window in the "all views" metrics, leaves click attribution unchanged, and shortens the window only for some views, below. Under a 14-day window, a shopper who clicks your ad on a Monday and buys on the Friday of the following week, eleven days later, is still credited to Monday's ad.
Cost does not wait for a purchase. Amazon can still adjust a day's cost, for example to remove invalid clicks or impressions, but most of it is known within a day, while sales keep arriving for two weeks. Every ratio built from the two (ROAS, ACOS, cost per purchase) looks worst on the newest days and improves as late purchases are credited.
The same week, read three times
A new line item runs for seven days and costs $700. Here is that same week, read the day after it ended (day 8), on day 14, and on day 22, when its last day is more than 14 days old:
| Read on | Total cost | Product sales | ROAS | ACOS |
|---|---|---|---|---|
| Day 8 | $700 | $980 | 1.40 | 71.4% |
| Day 14 | $700 | $1,680 | 2.40 | 41.7% |
| Day 22, window passed | $700 | $2,100 | 3.00 | 33.3% |
Nothing about the week changed between the three reads. Its cost was $700 each time; only the credited sales grew, from $980 to $2,100. By day 22, more than 14 days have passed since the week's last day, so no new purchases can be credited to it. Amazon can still correct the figures later, for invalid activity, but the window itself has passed.
What the first read would have cost
Suppose the product's break-even ACOS, the ACOS at which its margin before advertising is used up, is 40%. The break-even ACOS calculator works yours out from price, cost and fees. Against 40%:
- On day 8, 71.4% says the line item loses money on its attributed sales. It gets paused.
- On day 14, 41.7% says it is slightly over. It gets its max bid trimmed.
- On day 22, 33.3% says it was under break-even the whole time.
The day-8 decision stops a line item that was returning $3 of sales for every $1 of cost, on evidence that would have reversed itself within two weeks. The day-14 decision is smaller and still wrong. The week itself still settles to $2,100 in the report; what the pause stops is every week after it. And a seller who has moved on rarely goes back to re-read a paused line item's first week.
Divide by total cost, fee included
Amazon defines total advertiser cost as what an advertiser pays: supply cost, any demand-side fees, and other amounts payable to Amazon to serve the ad. When you rebuild a ROAS or ACOS from an export to test it against your break-even, use total cost too. A line item with $700 of total cost that includes, say, $40 of fees has $660 of spend before fees. $660 ÷ $2,100 is an ACOS of 31.4% instead of 33.3%: the fee left out flatters the result.
Spend without the fee has its place: it is what a flight budget is paced against, which is a different question, worked through in how to read an Amazon DSP order's delivery rate.
Views count, and since January 2026 they count differently
A DSP purchase can be credited to an ad the shopper only saw. Since 1 January 2026, for DSP ads that serve in Amazon's own store, the standard Purchases, Sales and ROAS credit views with a shorter, shopping-signal model, while "Purchases (all views)" keeps every view inside 14 days. Click attribution did not change, and nor did other inventory.
Two consequences. A DSP ROAS from December 2025 and one from February 2026 may not have been counted the same way, so part of a drop across that line can be a change in counting. And a report can now show two purchase figures for the same ad; compare like with like.
Where this still needs a person
Waiting is not a rule for everything. A line item with no impressions, or an order whose budget is running out early, has a problem that settled sales will not fix. The window only concerns the return, and only the newest 14 days of it.
Settled is not the same as caused. Attribution credits a purchase to the last ad before it; it does not prove the ad made the sale, and a remarketing line item shown to people who were already going to buy can look good on any window. Whether a line item earns its cost beyond the credit it receives is a judgement attribution cannot make for you.
Vendlen's DSP Manager puts the 14-day caution under its table and re-reads the last 14 days of DSP data every day, so its recent days move as late purchases are credited. The opposite mistake, averaging a long window until a recent break disappears, is in ninety days will hide the fortnight that broke.
What to do with this
- When you review DSP results, end the date range 15 days before today. Treat anything newer as provisional.
- For a line item that started this week, write down the date its first full week will be more than 14 days old, and make the keep-or-cut call then.
- Compare ACOS with the product's break-even ACOS on total cost, fee included.
- Do not compare a DSP ROAS from before 1 January 2026 with one after it without checking which purchase figure each one uses.
- Check delivery problems (no impressions, a rejected creative, a budget running out) straight away. Only the return has to wait.
Questions sellers ask
How long is the Amazon DSP attribution window?
Up to 14 days. Amazon credits a DSP purchase to the last ad before it, clicked or viewed, within a 14-day look-back. Amazon's 2026 update shortened that window for views of DSP ads in its own store, in the standard purchase metrics only; the "all views" metrics keep 14 days.
Why did my Amazon DSP sales change for days that already passed?
A purchase is credited to the last ad the shopper clicked or viewed, up to 14 days after it. A date range that ends within the last 14 days is still collecting those purchases, so its sales, ROAS and ACOS keep changing while its cost stays where it was.
Does Amazon DSP count view-through conversions?
Yes. A purchase can be credited to a DSP ad the shopper saw without clicking. Since 1 January 2026, views of DSP ads in Amazon's store are credited by a shorter, signal-based model in the standard metrics, so check which purchase figure a report shows.
How long should I wait before judging a new DSP line item?
Until the days you are judging are more than 14 days old, so their attribution window has passed. For a line item launched this week, that means its first full week can be judged about three weeks after launch. Delivery problems, such as no impressions, can be checked at once.