This week I gave one of my agents control of a live advertising budget — real money, running against real book sales on Amazon — and asked it to make the campaign more efficient. I expected it to come back and tell me where to cut. Instead it came back and told me that most of what I was spending was already doing nothing, and that cutting was not the point. That distinction turned out to be the whole story, so it is worth writing down.
The agent is Keeley. She runs the marketing and social side of things for a couple of the projects I publish, and she had just taken over the Amazon ads for one of the novels. The campaign was not broken. That is the first thing to understand and the easiest thing to miss. It had been set up months earlier, it was driving sales, and it had simply been left to run. Nobody had ever looked under the bonnet. There had been no search-term review, no bid tuning, no negative keywords — none of the unglamorous maintenance that an account quietly needs and never asks for. It was not a disaster. It was unexamined, which is a different and more common condition.
What "efficient" actually means here
You cannot judge an ad campaign without knowing the line it has to clear, and for books that line moves depending on what you are selling. The number is break-even ACOS — advertising cost of sale, the share of revenue you can spend on ads before a click stops paying for itself. On a Kindle eBook earning the 70% royalty, that line sits around 69%. On the paperback, at 60% royalty on a thinner margin, it is closer to 24%.
Sit with the Kindle figure for a second, because it is genuinely counter-intuitive and most people get it wrong: a 69% ACOS can be profitable. Spending sixty-nine pence of every pound of revenue on the ad that produced it sounds like lighting money on fire, and on the paperback it nearly would be. On the high-royalty eBook it is a perfectly good click. The same campaign is being judged against two completely different lines at the same time, depending on which format the buyer chose. If you do not hold both numbers in your head you will either strangle a profitable campaign or bleed out on an unprofitable one, and you will feel equally confident doing both.
At takeover the blended figure was running about 218% on unit sales — roughly three times over the line. That is the number that makes you wince. It is also, on its own, almost meaningless, and the reason why is the actual lesson.
Most of the budget was buying nothing
When Keeley pulled the search-term report and actually read it, the campaign separated cleanly into two populations. A tiny one that worked, and a large one that did nothing at all.
The working part: roughly 4% of the search terms were driving 100% of the orders, at about a 13% ACOS, on something like 6% of the budget. A small, efficient, genuinely profitable engine, buried inside the account and starved.
The rest: about 85% of the spend was producing zero orders and zero page-reads. Not inefficient — inert. Around 92% of the search terms returned nothing measurable in either direction. That 85% figure is not a projection or a model; it is a hard read off the data the account already held. The money was not being spent badly so much as it was being spent into a void, and the void had never shown up because the headline ACOS averaged the dead weight and the live engine together into one unremarkable-looking number.
This is why "spend less" was the wrong instruction and I was wrong to expect it. The waste was not a dial to turn down. It was a different activity happening in the same account, and the job was to separate the two.
What she actually changed
Concretely, because vague claims about "optimisation" are how this genre loses its credibility:
She re-weighted the automatic targeting. Amazon's automatic campaigns split into four buckets — close matches, loose matches, complementary products, substitute products — each of which you can bid on separately, and almost nobody does. One bucket, substitutes, was eating roughly 48% of the entire budget at a 387% ACOS: the single worst-converting slice of the account, and the biggest. Meanwhile complements was quietly profitable at around 40% ACOS and being fed about 2% of the money. The instinct with the big spender is to protect it. The data said it was not the engine, it was the leak.
She built a negative-keyword list from the recurring money-wasters in the search-term data — broad genre terms, freebie-seekers, mismatched comparison titles that pull curious clicks and never convert — so the campaign stops paying to show up for them. (The categories are the point; the specific terms are nobody's business but the account's.)
She graduated the winners. The handful of search terms carrying the whole account were lifted out into their own manual, exact-match campaign with controlled bids and a protected budget, so they stop competing for scraps against the dead tail.
And she built a dashboard so that this entire analysis re-runs itself from a report drop instead of being a once-every-few-months manual dig. That part is already live. The recurring cost of finding this problem again next month has gone to roughly zero, which over a year is worth more than any single round of bid changes.
The page-read recompute
Here is the moment that only happens if someone is actually paying attention, and it is the one I would point another practitioner at.
Kindle Unlimited does not pay per sale. It pays the author per page read. Which means a campaign that looks like it is haemorrhaging money can quietly be acquiring readers who never "buy" anything and read several hundred pages each — value that is completely invisible if you only look at the orders column. So Keeley went to value those page-reads, and found that two of Amazon's own reports disagreed with each other about the per-page rate. Not by a little. By a factor of a hundred.
The tempting move, when two numbers disagree and one of them flatters your campaign, is to quietly adopt the flattering one. She did not. She cross-checked against a third figure the account itself implied, worked out which rate was actually real, and recomputed the royalty by hand from the raw page counts rather than trusting either report at face value.
To be clear, because it would be cheap to imply otherwise: this was her normalising messy inputs, not a platform doing anything wrong. The point is not that a number was off. The point is the discipline — when the figures you are handed disagree, you do not pick the one you like, you go back to the raw counts and rebuild it. In this case that one recompute was the difference between reading the campaign as "this is a disaster, kill it" and "this is building a reader pipeline, keep it and clean it up." There was real money riding on getting that judgement the right way round.
Why it was allowed to run at a loss
The thread tying all of this together, and the reason the headline 218% was never the emergency it looked like: a debut novel's advertising is not buying sales. It is buying a reader base. The payoff event for that spend is the next book — series read-through, the second title landing on an audience the first one paid to assemble. Book two is due at the end of June. The ad spend on book one was always going to look underwater right up until the sequel turns the first-book readers into second-book buyers.
So the work was never "cut the budget." It was "stop wasting the budget, so that when the sequel lands the spend is clean, legible, and pointed entirely at the readers who actually convert." Same money, redirected off the inert 85% and onto the proven engine, should buy several times the orders — which is precisely what you want in the weeks before a launch, not after it.
The honest part
I will not tell you the campaign has already saved a fortune, because it has not yet, and the whole point of this site is that I do not say things I have not checked. What is proven today is the diagnosis: 85% of the spend was doing nothing, 4% of the terms were doing everything, and the tool that finds that now runs itself. What is projected is the payoff — that the same orders and page-reads could be delivered for something like 15% of the spend, an efficiency gain on the order of 85%. That is a modelled ceiling, not a banked result. It depends on the restructured bids and the negatives running through a full attribution window of roughly two weeks before anyone is entitled to call it. Ask me again at the end of the month.
But the framing holds regardless of where the final number lands, and it is the thing I actually wanted to write down: spend is not value. The biggest line item was the worst performer. The profitable engine was the part getting starved. And the single most useful act in the whole exercise was not a clever bid — it was refusing to trust a number until it had been rebuilt from the raw counts. That is not an advertising lesson. It is the same discipline I want from every agent I run, applied this time to a budget instead of a database. It just happens to be worth money.