Case Study 07 · Advertising Efficiency · Amazon India
+350% Return On Ad Spend By
No Longer Bidding Against Itself
Darkins, an Indian bean to bar chocolate brand, had rising Amazon ad spend, steady sales and falling profit — the combination that hides a structural problem.

We had the budget — what we didn’t have was an account that stopped bidding against itself. Harpy rebuilt it in four weeks: 78% less spend, the same sales, and 2.4x the profit after ads. I still can’t believe how much of our old spend bought nothing at all.
Richa ChaudharyFounder & Chief Chocolate Innovator, Darkins01 · Project context
What was happening in the account.
Darkins makes artisanal, bean to bar dark chocolate in India. Its range runs on 70% and 63% bars, with a sugar free line and gift boxes alongside. Amazon.in is one of its established channels, next to its own website and the quick commerce apps.
The challenge
Sales there were stable month to month, which is exactly what made the problem easy to miss. Ad spend was climbing, cost per sale was climbing with it, and profit was quietly eroding underneath a flat revenue line. Darkins’ own diagnosis was that Amazon had simply become more expensive — and that reading had been accepted across the business for most of a year. It was never tested, because the revenue number at the top of the report kept looking healthy. That is how expensive assumptions survive: not because anyone believes them strongly, but because nobody has a reason to challenge them.
02 · The strategy
Set the floor first: break-even per SKU, calculated from real economics, before a single bid moved.
Three moves, run in order. Nothing was cut until the previous step had proved itself. The spend did not go up because advertising got more expensive — it went up because the structure had drifted: campaigns added over years, each sensible on the day it was created, none of them revisited as a whole.
Establish the true break-even first
Break-even ACoS was calculated for every product from real economics — sale price, category referral fee, fulfilment, landed cost and returns allowance. That produced a floor per SKU rather than a single target for the account, and it immediately reframed the conversation: several campaigns that looked underperforming were profitable; several that looked comfortable were quietly losing money on every order.
Find the spend that buys nothing new
A search-term review across the whole account showed where the same customer journey was being paid for more than once — broad and automatic campaigns harvesting the exact matches sitting in their own campaigns, brand terms bought by auto targeting, and duplicate keywords running in parallel. That overlap was peeled out first, with negatives and consolidation, before a single bid was lowered.
Rebuild by job, then trim
Campaigns were reorganised so each had one job: hold the known converters, discover new terms under a capped budget, defend the brand cheaply, or take share where the margin allowed it. Bids were set against the SKU-level break-even floor — not an account-wide ROAS target — and the discovery budget was capped so it could not quietly become the largest line in the account.
03 · The execution
Peel the overlap out, rebuild by job, then trim — in that order.
Two failure modes were designed out before any bid was lowered.
Waste that was not in the ads at all
The last sweep found money outside the campaign structure entirely: bars that were out of stock but still receiving impressions, listings with suppressed Buy Boxes collecting clicks Darkins was paying for, and placements where the cost per sale was higher than the whole margin on a bar. Fixing these is unglamorous, and it moved the number more than any bid change did.
Guard against the fix creating a new problem
Nothing was removed until the negatives and the consolidation had run for a full week and the account was spending less on the same sales. Only then were bids trimmed, and only against the SKU floor. Two terms were restored during the month because the data did not support cutting them — that is not a failure of the process, it is the process.
A report that cannot hide the problem
Campaign-by-campaign reporting was replaced with a single account view: spend, sales and cost per acquisition for the account as a whole, with the SKU-level break-even floor beside it. The question it answers is not how any one campaign performed — it is whether the account is buying customers more cheaply this month than last. That is the number that would have caught this a year earlier.
04 · The results
Lower spend, same sales, in four weeks.
The rebuild took four weeks. Over the 90 days that followed, Darkins spent 78% less on Amazon ads than in the 90 days before, total sales rose 3%, and return on ad spend was four and a half times what it had been. Because the changes were structural, the savings held rather than showing as a one month dip. The account had stopped paying twice for the same sale, and it had a floor to test every future bid against.
Amazon India · 90 days after vs 90 days before
Return on ad spend
TACoS
Ad spend · account total
Profit after ad spend
What the rebuild removed
Account structure
Nothing about the chocolate, the price or the range changed. The same Darkins bars sold to the same shoppers. The only difference was that the account no longer competed with itself to reach them. Most of the money that had gone on duplication was simply no longer spent, and the rest went to the bars that could carry it.
05 · Proof · from darkins.in
The brand behind the account.
Product, factory and gifting imagery served from the Darkins website — the pages shoppers actually landed on.



06 · The takeaway
It didn’t need a bigger budget. It needed to stop bidding against itself.
Darkins had enough Amazon budget. What it needed was an account that stopped bidding against itself. One month of structural work recovered more margin than a year of bid tinkering, and the floor it produced keeps every future bid honest. The account has not needed a second cleanup: the floor is recalculated whenever costs move, the negatives list is maintained as part of the weekly routine, and the overlap that built up over years cannot rebuild itself unnoticed — the reporting would now show it as money spent twice.

