The report looked fine. It was wrong.

A report that will not parse is an inconvenience. You notice immediately, you fix it, you move on. The expensive case is the export that loads perfectly, produces numbers that look entirely reasonable, and is describing something other than what you asked.

We have built parsers for thirty-five marketplace report shapes across Amazon, Flipkart and Myntra. What follows is the set of traps that recur — the ones that produce a confident wrong answer rather than an error.

1. Attribution windows that do not match

Sponsored Products reports default to a seven-day attribution window. Sponsored Brands defaults to fourteen. This is the single most common source of a wrong conclusion in Amazon advertising, and it is invisible unless you go looking.

Fourteen days catches sales that seven days never sees. So identical underlying performance produces a materially better-looking ACOS on the Sponsored Brands report. Sellers regularly conclude their brand campaigns outperform their product campaigns, shift budget accordingly, and have measured nothing except how long each report waited.

2. Attribution that has not finished arriving

Related, and just as costly. Attributed sales keep landing for days after the click that produced them. Spend, meanwhile, is recorded immediately.

That means any recent period looks worse than it will turn out to be. A campaign's first week always shows a worse ACOS than it eventually earns. Pull a report today for the last seven days and you are reading a number that will improve on its own, without you doing anything.

The practical consequence: never compare a fresh pull against a settled one, and give a new campaign at least a fortnight before judging it. Plenty of campaigns get paused in week one for a number that would have been fine in week three.

3. Blank is not zero

On Flipkart's PLA reports, the Expected ROI column is often empty. The natural reading — a campaign with a target of zero, or a campaign failing badly — is wrong in a way that inverts the conclusion.

A blank means the campaign is a CPC campaign rather than an ROI-targeted one. It has a cost-per-click bid and no return target by design. Reading it as a missed target produces the wrong verdict on every CPC campaign in the account.

There is a second layer to it. On campaigns that genuinely are ROI-targeted, the target lives at ad-group level rather than campaign level, so it can look absent at campaign level even when one is set.

The general principle is worth carrying to every export you handle: an empty cell means "not applicable" at least as often as it means zero, and aggregating the two together is how averages quietly stop meaning anything.

4. The same column on two different scales

Flipkart publishes conversion rate as a fraction in some report types and as a percentage in others. Nothing in the file announces which you have.

A parser that assumes one will be wrong by a factor of a hundred on the other, and — this is the problem — the output still renders. Numbers appear. They are just describing a different universe.

The check takes two seconds: look at where the column tops out. Near 1, it is a fraction. Near 100, it is a percentage. Do it before drawing any conclusion from a Flipkart export you have not handled before.

5. The header is not always on row one

Several marketplace exports put preamble above the header — account name, date range, report type — and on Flipkart the number of those rows varies by report type.

Open one with the wrong header row and you get column names sitting in the data and data sitting in the column names. Sometimes that is obvious. Often, when the preamble happens to be the same width as the table, it is not, and the spreadsheet looks close enough to plausible to act on.

6. Targeting is not the search term

In the Amazon search term report, two columns sit next to each other and are easy to conflate. The targeting is what you bid on. The customer search term is what somebody actually typed.

On exact-match campaigns they are usually the same, which is exactly what makes the habit dangerous — you get used to them matching. On automatic and broad campaigns they are routinely nothing like each other, and the gap between them is the entire reason the report exists.

This also decides where a negative goes. Negatives apply to search terms, but the spend is often being driven by a broad targeting matching dozens of them. Negating one term stops one leak; the word-level view finds the one leaking across forty.

7. Settlement is not profit

Not a parsing trap so much as a reading one, and the most expensive on this list because it is a pricing decision rather than a reporting one.

A settlement figure is what the marketplace transfers to your bank. It is net of their fees and nothing else — not your cost of goods, not your advertising, not the tax you owe, not unrecovered returns.

8. Blended figures that average things that disagree

The last one is not a quirk of any particular file — it is what happens whenever a report aggregates. A blended ACOS averages brand defence at 8% with discovery at 60% and describes neither. A catalogue margin of 27% can be one strong product carrying a dozen loss-makers.

Averages are not wrong, but they are answers to a question you probably did not ask. The useful move is almost always to break the number apart along the dimension that matters — by campaign objective, by product, by placement, by day — and see what it was hiding.

A short checklist

• Check the attribution window before comparing any two advertising reports.
• Give a recent period time to settle before judging it, and never compare fresh against settled.
• Treat blank cells as absent, not zero, until you know which the file means.
• Look at where a percentage column tops out — near 1 or near 100.
• Confirm the header row, especially on Flipkart, where the preamble varies by report type.
• Keep targeting and customer search term straight, and negate at the level the spend is actually happening.
• Never price against a settlement figure.
• Break blended numbers apart before trusting them.

None of this requires a tool. It requires knowing which questions to ask of a file before you believe it — which is most of what separates an hour of useful analysis from an hour of confident, well-formatted wrongness.