Amazon Advertising: The Definitive Guide

Amazon advertising is one machine with several surfaces: sponsored ads (the auction for search and detail-page real estate), display and streaming bought through DSP, and AMC, the clean room where exposure meets outcome. Underneath all of it is the retail engine the rest of this library covers: the listing that converts, the reviews that gate, the rank that compounds. Advertising is not adjacent to that engine. It is a way of buying its inputs.

That is easier to hold onto with the specific prompts named, because each one is a real recommendation you will see this week, and each one resolves to the same request.

One framing before the mechanics: it is impossible to be good at advertising in isolation. Ads convert against a listing, sell against a review count, and ship against an inventory position. An advertising decision made without unit economics, creative, and stock in the room is not a decision, it is a guess with a budget attached. That is also the honest warning about hiring ads-only help, which chapter 20 states plainly.

In physical retail you pay a slotting fee for shelf placement: an acknowledged, budgeted cost of buying position. Amazon has slotting fees too. They just arrive dressed as a terrible ACoS, and failing to recognise them kills more launches than any other accounting error.

Two frames make the spend legible. The first is lifetime value: if the average customer buys five units, the first sale is one fifth of the relationship, and losing money on it can be profitable across the whole. The second is the moat: on a launch the binding constraint is usually ratings, and our own page-one tracking puts the median review count on a competitive shelf in the thousands. Bought velocity is how a new ASIN accumulates ratings fast enough to clear the bar, and a rating bought that way is not spend, it is enterprise value accruing to the ASIN.

The instinct that a first order can carry “200% and change” turns out to be arithmetic. If lifetime contribution margin has to cover the acquisition cost, then max first-order ACoS = contribution margin % × expected lifetime units, less whatever margin you want to keep. At a 42% margin and five units, the ceiling is about 210%.

The second question our tracking can answer that the console cannot: what kind of brand is behind a sponsored slot? We compared every sponsored slot against every organic slot sitting on the same captured pages.

The advertiser carries barely a third of the review count of the organic shelf and asks roughly 38% more money, at a rating that is effectively identical, 4.43 against 4.49. That is the challenger's signature: unable to win the shelf on moat, unwilling to win it on price, buying the position instead. It is the market independently confirming chapter 3, and it should be reassuring the first time you run a launch at an ugly ACoS. Everyone in that block is doing some version of the same thing.

The companion finding is who those advertisers are relative to the page they appear on. In the july 2026 slice we classified brand by brand, 69.4% of sponsored slots were held by brands with no organic rank on that keyword at all, renting rather than defending. Renting is legitimate, it is how challengers sample demand, but it prices in forever-rent, and the flywheel case for advertising lives in the other 30%.

Top of search is not a fourth bucket. It is a placement decision layered onto Retention or Acquisition, and keeping that straight prevents two expensive errors.

The first is judging it by the wrong scoreboard. Top-of- search spend serves share of voice, impression share and organic rank improvement, not pure cost per acquisition, so measuring a top-of-search carve-out against the acquisition bucket's efficiency math guarantees you kill it exactly when it is working. The second is forgetting it has a flavour: top of search can be defensive, owning the top of your own brand's page during a siege, or offensive, buying the shelf's premium placement to build rank. Those deserve different targets, and per chapter 4 they are bidding into a block that is already about 71% bought.

The arithmetic underneath is worth seeing, because it is genuinely ambiguous rather than obviously good. The premium placement costs more per click and converts better, and those two facts do not cancel cleanly:

Every permutation exists in real accounts. Sometimes the higher conversion rate more than covers the higher click cost and top of search is simply better on every measure. More often it looks like the table above: worse on ACoS, much better on volume, and better on TACoS once the velocity it buys starts feeding rank. Occasionally the conversion lift never materialises and the premium is wasted. The only way to know is to read the placement report for your own terms, which is why the modifier exists at all.

Audience targeting on Amazon is the most misreported subject in the category, because the answer is genuinely fiddly: the same audience feature behaves differently depending on which ad type you are running it in. Here is the whole thing in one table.

The row that matters most: AMC custom audiences are real targeting in Sponsored Display, but only a bid adjustment in Sponsored Products and Sponsored Brands. In search you still target by keyword or product, and audience membership only changes what you are willing to pay when that shopper shows up. Anyone describing an “audience-targeted Sponsored Products campaign” is describing something that does not exist.

Once you can target audiences, the harder question is which ones deserve the money, and the answer runs opposite to what the reports will tell you. Audiences sort along a spectrum from cold to warm, and warmth is inversely related to how much credit an ad actually deserves:

This single figure explains most of the disagreements between advertising teams and finance teams. The lines at the bottom of the spectrum, retargeting and branded search, report the best returns in nearly every account and contribute the least new revenue; the lines at the top do the real acquisition work and look worst on a dashboard. None of that makes the warm lines worthless, defending and recovering carts is real work, but it does mean the ranking of campaigns by reported return is close to an inversion of their ranking by contribution.

The practical consequence is a sequencing rule. Audience sophistication belongs where audiences are actually targetable: Sponsored Display for remarketing and Amazon's in-market and lifestyle segments, streaming for reach against interest and life-event audiences, DSP for the full library. In search, audiences are a refinement on top of keyword strategy, not a replacement for it. Sponsored Brands does carry bid adjustments for its own audiences, including shoppers who previously purchased your brand or added it to cart, which is a cheap and underused lever because it costs nothing to configure and only bites when the shopper is already warm.

With bids doing the steering, budgeting becomes what it should have been: a finance instrument rather than a throttle. Bids ration toward your best traffic; caps sit underneath as a fail-safe. Which is why you should never be running out of budget: the dark afternoon in chapter 2 was not overspending, it was over-bidding, and the fix was the dial rather than a bigger cap. TACoS sets the envelope at the P&L level. The envelope cascades to product groups and then to buckets, because Retention, Acquisition and Discovery are different investments with different rules. Pacing tracks the month as a curve rather than a line, weekends and paydays weighted, with an adaptive daily target, remaining budget over remaining days, that absorbs reality as it happens.

Pacing hot has an exchange rate, and it is worth carrying in your head: in a 30-day month, every 3% you run over plan is roughly one day out of budget at the end of it. Six percent hot is two dark days, nine percent is three. That is what makes a pacing number urgent rather than administrative. It is not a tidiness metric, it is the calendar telling you which days in the last week you have already decided to sit out.

Two alerts earn an interruption. A campaign capping out while performing under target is declining profitable demand every capped hour: the most expensive quiet failure in ad operations, and by the rule above it should be nearly impossible in a well-run account. And a bucket underspending its plan by half means the plan and the account have diverged, either because the auction got expensive, which is the margin-versus-underspend decision from chapter 2, or because the structure has a hole.

Amazon DSP is programmatic display, video and audio bought on audiences instead of search terms, across Amazon's own properties and a large and rapidly growing pool of third-party supply. It is the part of the channel where the most money is wasted per hour of attention paid to it, and the reasons are structural rather than moral: the reporting flatters itself, the inventory is invisible unless you go looking, and the people selling you access have their own reasons for how they frame the entry price.

The minimum that is not a minimum

You will be told DSP requires a large minimum spend. Figures between thirty-five and fifty thousand dollars circulate as though they were platform rules. They are not. There is no minimum spend that gates the technology. What those numbers actually describe is the commercial convention of managed service, and, more often, the floor a particular agency requires to make your account worth staffing. That is a reasonable business decision on their part and a terrible piece of information for you, because it gets repeated as if Amazon set it.

The attribution problem, which is the real barrier

The reason DSP deserves caution is not its price. It is that DSP has an enormous tendency toward inflated attribution, and the inflation arrives pre-installed in the default reporting. Search advertising flatters itself too, mostly through branded interception, but display and streaming flatter themselves harder because they run on impressions: a person who never clicked anything can be credited with a purchase they were going to make anyway.

Put those three together and a campaign can report a handsome return while contributing almost nothing, and the failure mode is not that somebody lied. It is that nobody asked the reported number what it was made of. Asking is a mechanical exercise: strip one layer at a time and watch what survives.

Every step down that ladder is a question the default report does not ask on your behalf, and the last bar is the only one that corresponds to money the business would not have made anyway. Run the decomposition before a renewal, not after. This is the surface where the incrementality discipline from chapter 15 stops being an intellectual exercise and becomes the difference between a channel and a subscription.

Know what you are buying: targeting and inventory

Two competences separate DSP that works from DSP that merely spends. The first is knowing your targeting precisely: which audience, built how, refreshed how often, and whether the tactic is prospecting into new demand or harvesting demand you already created. Amazon's goal-based products complicate this in a way worth stating plainly, because it is easy to misread: when you attach your first-party or clean-room audiences to an automated conversion-optimizing campaign, those audiences generally act as optimization signals rather than hard targeting constraints. The system is being pointed, not fenced. That is powerful when the system has enough conversions to learn from, and it is how budgets quietly leak into audiences you never intended to buy when it does not.

The second is knowing your inventory: where the impressions physically ran. Programmatic buys placements you did not individually approve, and the spread inside a single campaign runs from premium streaming to the bottom of an app nobody scrolls to. The supply pool has also changed faster than most published guidance, and it has changed enough that it is worth seeing whole rather than as a list of names in a sentence.

The middle two columns are the story. Every one of those deals was signed inside five months of 2025, which means a plan written from a 2024 understanding of DSP is planning against roughly the Amazon-owned column alone. It also means the honest description of the product changed: this is a general streaming and audio buying tool that happens to have Amazon's retail signal attached, not an Amazon-inventory product with some extras bolted on. Which raises the stakes on the placement report considerably, because “where did it run” now spans four columns instead of one.

When it earns the next dollar, and how to judge it

Sequencing first: sponsored search should be structurally sound before DSP gets budget, because DSP's job is to create demand that search then harvests, and an account that cannot harvest efficiently will leak whatever DSP sends it. Once that holds, DSP buys the things search cannot: shoppers who are not searching yet, people who viewed and did not convert, past purchasers worth reactivating, and reach on screens. Streaming TV specifically is self-service in the console, which makes it the usual first step off the search shelf for a brand that wants screens without a commitment.

One structural note for planning. The wall between DSP and sponsored ads is coming down inside the interface, with unified campaign management and unified reporting, which means the old organizational split, one team for search and another for programmatic, is becoming an artefact of the tooling rather than a fact of the channel. The measurement discipline, however, does not merge automatically. Keep reading the two apart until you have proven they can be read together.

Every dial in chapter 2 asks the same underlying question: would this sale have happened anyway? Attribution cannot answer it, because the last click takes credit whether it caused the purchase or merely stood nearest to it. But the answer is estimable, from a ladder of indicators that starts free and climbs as the budget justifies.

The cheapest rung is the most neglected, and our shelf data shows why it matters at market scale: about seven in ten sponsored slots are held by brands with no organic rank on the keyword. Renting is maximally incremental and maximally temporary; defending your own top-three organic slot is the opposite on both counts. Neither is wrong. Not knowing which one you are doing is.

AMC is Amazon's clean room: ad exposure and conversion events joined at the hashed-customer level, answering questions the console structurally cannot. What it is for, in this doctrine, is better decisions about incrementality, the new-to-brand rate of specific keywords, the lifetime value of specific audience segments, and the precision to optimize against customer acquisition cost and real incrementality instead of blended return.

The access story changed and most published guides have not caught up: AMC became free and self-service for sponsored-ads-only advertisers in September 2025, with no DSP spend and no partner required, and the lookback window has since roughly doubled. If your mental model is “AMC is for enterprise DSP buyers,” it is a year out of date.

The honest adoption rule is a maturity curve. For a new or immature brand the gap between AMC-informed decisions and instinct is small, because instinct is roughly right: someone searching a non-brand term has never heard of the brand that launched yesterday, so new-to-brand rates are high nearly everywhere and the clean room mostly confirms it. For a mature brand the gap is enormous: which keywords still bring genuinely new customers, which audiences actually repeat, where frequency stops buying anything. AMC is always of value; the marginal gain grows with the budget and the brand's maturity, and so should the investment in it. Its limits travel with every read: ad-exposed customers only, so pure-organic buyers are invisible; a bounded lookback; and small cells suppressed by design.

Most Amazon advertising content on the internet describes a 2022 product. That is not a jab, it is an operational warning: several changes in the last year break playbooks that were correct when they were written, and two of them change numbers you are probably reporting to somebody.

The reporting sunset at the end of 2026 is the one with a deadline attached. If your dashboards, exports or agency reporting are built on the legacy separate feeds, that is rebuild work with a fixed date, and it is the kind of thing that gets discovered in January rather than planned in August.