The Amazon Operator's Field Guide

Start here

Strip away the jargon and an Amazon business is one equation:

Amazon calls those sessions, conversion rate and price. Every lever you have pulls one of the three. Advertising buys sessions. Your listing and reviews drive conversion. Pricing is price. That's the whole game.

Then Amazon takes its cut and you pay for the goods:

That middle line matters more than almost anything else in this guide. It has a name - contribution margin - and it is the pot of money advertising spends out of. If you know it, every advertising decision has a right answer. If you don't, you are guessing, and Amazon is very good at letting you guess expensively for a long time.

The three numbers people confuse

Start here Do this before you touch a single bid.

Break-even ACoS is the ACoS at which an extra advertised sale makes you exactly zero. Above it you're paying for volume; below it you're making money. And it is simply your contribution margin expressed as a percentage.

Do this once per product family, not once per account. It is the difference between "our target ACoS is 30%" - a sentence that is wrong for most of your catalog - and knowing that one line breaks even at 42.6% while a heavier, lower-priced line in the same catalog breaks even at 19%.

Strategy, posture, and the working ceiling

Break-even is the floor of the conversation, not the answer. What you're trying to do with a product changes how far above break-even you're willing to run. Two separate things set that.

Strategy - the long-run intent, set per product

Where the product sits in its life: launch, grow, dominate or maintain. This changes slowly and deliberately, and it is a field you set on the product, not a mood.

Posture - the temporary state it's in right now

On top of the strategy, a product can be in a state that overrides it for a while:

• Hold - deliberately not growing. Winding down, capacity-constrained, or under review.
• Recover - climbing back from a stockout, a rank loss or a listing problem. Judge it on rank returning, not on this month's efficiency.
• Inventory override - stock is the binding constraint, so efficiency rules give way to protecting cover. Chapter 21.

Postures are states, not goals. A dominate product sitting in an inventory override is still a dominate product; it just isn't allowed to scale this week.

The ceiling

Multiply break-even by a factor for the strategy, or for the posture when one is active, and you get the working ceiling. A reasonable starting calibration:

The ceiling cuts both ways. If a keyword on a launch product is running at 48% ACoS, that's inside the 59.6% ceiling - leave it alone, it's doing its job. If a keyword on a Hold-posture product is at 48%, it's well above the 36.2% ceiling and needs cutting. Same number, opposite action, and the difference is written down before you look at the report rather than argued about after.

Start here Twenty minutes of admin that saves you from hundreds of bad decisions.

Keep one simple sheet listing every SKU with three columns beyond the basics. Everything downstream reads it.

Status - is this thing alive?

Which products are heroes

You need some way of saying "this one carries the business." It does not much matter what the scale is - a simple top-quintile flag, an A/B/C grade, or a 1-to-5 rank all work, as long as you are consistent. What matters is what the flag changes.

On a hero product: move in smaller increments, want more evidence before you scale up, and look at the change yourself before it goes live. On a tail product, let the routine run.

Strategy - launch, grow, dominate or maintain

From chapter 2, plus any temporary posture (Hold, Recover, Inventory override). A blank strategy on an Active product reads as maintain. Set launch the day a new product goes Active, not before.

Launches get judged differently

For roughly the first 90 days after a product's first receipt, profitability is the wrong test. The right test is whether organic sales are growing.

One more launch diagnostic worth knowing. If lots of people click but few add to cart, your listing loses them on arrival - that's images and title. If they add to cart but don't buy, that's a price or value objection. And if cart-to-purchase is weak, do not raise your price. Change coupons before list price: a coupon is reversible in a day, a list price change is sticky and messes with your price history.

Start here If you already know match types and placements, skip to chapter 5.

You bid on what a shopper types. Amazon runs an auction, and the winner's ad shows. You pay only on a click. Three things decide whether you show and what you pay: your bid, how relevant Amazon thinks you are, and where on the page you're willing to appear.

Match types - how loosely your keyword catches queries

Placements - where on Amazon your ad lands

The same campaign shows in different places, and they perform very differently. You control the mix with percentage modifiers that raise your bid for a given placement.

Attribution - why yesterday's numbers lie

Once you're running How to lay out campaigns so the account is still legible in two years.

Most accounts get messy the same way: someone adds campaigns as needs arise, nobody can remember which is which, the same keyword ends up in four places bidding against itself, and eventually the only honest option is a rebuild. The fix is to decide the shape up front and never deviate.

Name everything the same way

Unglamorous and enormously valuable. You can sort, filter and search an account like this, and you can hand it to someone else without a translation document.

Build in phases, one keyword family at a time

The old approach was to stand up a full exact-match architecture on day one and let discovery feed it. That produces a tidy account that cannot reach half the places shoppers now find products. The current model is the opposite order: start broad, prove the family, then graduate the winners into exact control.

Work one keyword family at a time. A family is a head term and the queries that orbit it - not your whole catalog, and not one ASIN's entire keyword list.

The exact campaigns are still where you control bids and build rank - nothing about that has changed. And they do not have to wait for stage 3: when chapter 6's research hands you real evidence - honest relevancy, a page one full of products like yours, an organic rank worth amplifying - launch the single and/or stacked exacts alongside broad and the autos on day one. Evidence is the gate, not the sequence. Either way, broad and auto keep running underneath the exacts permanently rather than being starved once exact exists.

Set bids as a ladder off your main keyword

Broader match types reach more queries per dollar and convert worse, so they should pay less per click. Pick one bid for your main exact keyword and derive the rest:

Managing overlap between your own product lines

In a multi-product account, your broad and auto campaigns will inevitably start buying clicks for queries that belong to a different product of yours. Sometimes that costs you money. Sometimes it makes you money. The mistake is treating all of it as one thing.

Practically: when you launch a new family, pre-load the head noun of directly substitutable lines as negative phrases in that family's broad and auto campaigns - a backpack line and a duffel line are substitutes, and a shopper searching one is not shopping the other. Do not pre-load complements or adjacent items from the same use case. Then audit the overlap list quarterly with the method above and adjust, rather than assuming the launch-day guess is still right a year later. Competitor brand names belong on the same pre-load list, for a different reason: those are a deliberate conquest decision, not something discovery should stumble into.

Once you're running

You need four things about every candidate keyword, and each answers a different question. Most people gather two of them and wonder why their keyword lists are full of terms that never work.

1. Your product relevanceDoes this keyword honestly describe what you sell? Usually a 0-9 score from a keyword tool, but read a few by hand - tools mis-score edge cases constantly.
2. Competitor relevanceHow many products rank on page one for it, and are they products like yours? This is the market's verdict on what the query actually means. A term where page one is fifteen products in your exact category is a real category keyword, whatever a relevance score says. A term where page one is a scattered mix is either a broad informational query or one Amazon hasn't resolved - treat it with caution.
3. Your organic rankWhere you currently show without paying. This does not decide whether to target a keyword - it decides what kind of play it is. See below.
4. Search volumeHow many people actually type it. Search Query Performance (Brand Registry) gives you the real number rather than a third-party estimate, plus click share and purchase share, which are more useful than volume on its own.

Throw things out before you cluster

• Other languages - unless you're deliberately targeting them.
• Other brands' names - hold these out by default. Conquesting is a deliberate decision made on contribution dollars, not something that sneaks into a launch plan.
• Keywords that don't share your product's main noun - if the query doesn't name your product type or a genuine synonym for it, it isn't your keyword. Check the synonym list against page one rather than your own vocabulary: shoppers routinely use a word for your product that you never put on the listing.

Keep the discard list. When something underperforms later, the first useful question is "what did we exclude, and why?"

Group twice: by product family, then by shopper intent

First pass, cluster the survivors around their main noun - that's your product family, and it decides which campaigns they can live in at all.

Second pass, and this is the one most people skip: within a family, group by what the shopper is actually trying to do. The same product noun carries very different intents, and they want different bids, different ad copy and different landing experiences.

Then pick the single best keyword in each family - highest relevance times highest volume. That's your anchor. It gets its own exact campaign, and it's the only keyword in your phrase campaign. If no keyword in a family clears your relevance bar, that family doesn't get a launch plan. That's a real answer, not a failure.

Where each keyword goes

Note what is not in that table: a hard organic-rank gate. Rank changes the expected cost and the time to payback, so it belongs in how you bid and how patiently you judge - not in whether the keyword is allowed to exist. A term you don't rank for, where page one is full of products like yours and volume is real, is exactly the kind of target that grows a brand. Budget it as a test, hold it to a longer horizon, and watch for organic rank appearing as the signal it's working.

So: in broad, collapse aggressively. Lowercase, drop the connector words, singularize, sort the remaining words alphabetically - if two strings match after that, you only need one broad keyword, and running three is bidding against yourself three ways.

In exact and phrase, only collapse plurals and misspellings. Keep re-ordered and re-phrased variants as separate keywords, at least to begin with, and let the data decide. If two orderings show meaningfully different purchase share in Search Query Performance - roughly 30 percentage points apart is a reliable signal - they are genuinely different queries and both deserve their own campaign. If they track each other closely over a few hundred clicks, keep the stronger one and pause the other.

Once you're running

Start from Amazon's own suggested bid for your anchor keyword in exact match. It's an imperfect number but it's built from real auction data, and it beats a guess. Then apply the ladder from chapter 5 for everything else.

Budgets are set from the bid, not from a feeling. The reasoning: a campaign needs enough daily budget to buy a meaningful number of clicks, or it runs out by 10am and you learn nothing.

The first eight weeks

1. Week 0 - launchOne broad seed for the family, plus autos on close match and substitutes. Add the single and/or stacked exacts now too when the chapter 6 evidence says the product should convert there - hold them back only when budget is the constraint, because an exact built on nothing but a guess is just discovery at your highest bid.
2. Weeks 1-2 - aggressive cleanupThis is the period where discovery buys the most irrelevant traffic, and where negation pays for itself several times over. Run the negation pass (chapter 11) and expect to block a lot. That is the discovery campaigns doing their job, not failing at it.
3. Week 3-4 - first harvestTerms that are converting graduate (chapter 10): 1,000+ monthly searches gets its own single-keyword exact campaign, under that goes into a stacked exact grouped by intent. Add the phrase campaign on the family anchor at the same time.
4. Week 4+ - widen the autosWith two weeks of negatives in place, turn on loose match and complements deliberately, one at a time, each with a budget you're willing to lose. Measure them on their own terms.
5. Once the exacts have 3+ conversions from harvested termsThe family has proven it converts on controlled traffic. Now expand the ad types - Sponsored Brands video, product targeting against the competitor ASINs your substitutes campaign found.

Throughout all of this the broad and auto campaigns keep running and keep being funded. They are not scaffolding you remove once exact exists - they hold placements the exact campaigns cannot reach at all.

Once you're running The core loop. An hour a week once you've done it a few times.

Ten rules cover almost everything. Run them in this order, because protective cuts should always get first claim on a keyword before a growth rule touches it.

Two things fall out of writing the rules this way. The cut rules run off the ceiling, so a Launch product is automatically allowed to run hotter without you remembering to make an exception - on that same 42.6% break-even, a Launch product isn't cut until 59.6%. The scale rules run off raw break-even, not the ceiling, because you should only be buying more of something that is genuinely profitable, regardless of what you've permitted it to spend up to.

Four things that stop a rule before it fires

1. Low stock - and mind the grain. Two separate gates, and confusing them is expensive in both directions.
   Product level: a product at 14 days of cover or less gets no scale-up. Cuts always still fire.
   Campaign level: a campaign only goes reduce-only when every product advertised in it is warned or out. One low sibling in a multi-product campaign does not freeze it - the in-stock siblings simply absorb the impressions, and throttling the campaign to protect one variation just hands that traffic to a competitor.
2. Discontinuing. Sell-through mode - cuts only.
3. Unlaunched. Out of scope entirely.
4. Goal = Hold. No scaling.

One change per lever, then wait

A campaign has separate levers: budget, bid, and one for each placement. Change one thing per lever per week, then wait at least seven days plus the attribution window before you judge it. If you cut a bid and drop a budget and change two placement modifiers in one sitting, you will never know which one did what - and next week you'll be arguing with yourself about it.

Once you're running

Placement modifiers are percentage uplifts on your bid for a given surface. A 50% Top of Search modifier means a $1.00 bid becomes $1.50 when competing for that slot. They run from 0 to 900% - and note there is no such thing as a negative modifier.

The thresholds, and why they differ

Same principle as chapter 8 - everything is a multiple of that product's break-even ACoS, so these travel correctly across a catalog with mixed margins.

Top of Search is tolerated above break-even because it's where rank gets built and where the best shoppers are - you're buying a position, not just a sale. Rest of Search is held exactly at break-even because it's the least intentional surface and it doesn't buy you rank. Product Pages is allowed the widest band before you cut, because a product-page ad that is working is genuinely valuable.

How far to move a modifier

The thresholds above tell you when. This tells you how much, and it is the part that goes wrong.

As for what the modifier should eventually be: the target isn't a fixed number, it's the level at which your cost per conversion is roughly equal across the placements. If Top of Search costs $9 per order and Product Pages costs $22, the modifiers are wrong regardless of what either ACoS reads. Move toward equalizing that, a clamped step at a time, and re-read after a full attribution window.

You need at least 10 clicks in a placement before any of these mean anything - and at least one order before you raise anything.

Off-Amazon, in plain terms

Amazon shows your ads on other websites. Two things to hold in mind:

• Its reported ACoS is a ceiling, not a truth. Off-Amazon is prospecting. Someone sees your ad on a news site, searches you on Amazon three days later, clicks a Top of Search ad and buys - and Top of Search takes all the credit. Judge off-Amazon on contribution dollars, not on ACoS parity with your search campaigns.
• Lots of off-Amazon spend usually means your bids are too low. When your bid can't clear the on-Amazon auction, the overflow goes to the cheap offsite inventory. The right fix is often a bid increase, not an off-Amazon cap.

The test: take the product's break-even ACoS and subtract the campaign's on-Amazon-only ACoS (the three on-Amazon placements - never the blended total, which is contaminated by off-Amazon). That gap is your headroom.

Off-Amazon serving can only be changed through a Seller Central bulk file, and the only setting is "Limit off-Amazon spend." There's no bid modifier and no true off switch. Also: make bid and modifier changes first, and off-Amazon limits a few days later, so you can still read which one did what.

Once you're running The growth loop. Run it every two to four weeks.

Your broad, phrase and auto campaigns are constantly discovering search terms you never thought to bid on. Some of them convert. Harvesting is the act of taking those proven terms and giving each one a campaign where you control the bid.

Why bother, if it's already converting? Because in a broad campaign it's sharing a bid and a budget with hundreds of terms you don't care about. Move it to its own exact campaign and you can bid it properly, budget it properly, and set its placements - which is how a discovered term becomes a ranked term.

The one test that matters: conversions per week

Because it's a rate, it doesn't matter whether you pull 30 days or 60 - the bar is the same. Use 30 days as the default so the numbers you work from match what the reporting tools show you by default.

Check stock before you create the destination

Check coverage before you launch anything

Group the small ones by intent, not alphabetically

When you're putting sub-2-per-week terms into stacked campaigns, group them by what the shopper wants: head terms, use-case terms ("backpack for air travel"), audience terms, feature/spec terms, comparison terms. Terms in the same intent cluster want similar bids, which is the entire point of stacking them together.

Once you're running The cleanup loop. Every two to four weeks, paired with harvesting.

A negative keyword tells Amazon "never show my ad for this." It's the most powerful and most over-used tool in the account. Used well it removes real waste. Used carelessly it quietly strangles a campaign and you find out three months later.

Four things you must never negate

1. Anything inside an exact-match campaign. You already chose that keyword deliberately. If it isn't working, lower the bid or pause the keyword. A negative there either does nothing or switches the campaign off entirely.
2. Anything inside a rank campaign. High ACoS there is the point - you're buying a position.
3. Your own brand name, or your product's core noun. Branded traffic is usually the cheapest, highest-converting traffic you have.
4. Anything with fewer than 5 clicks. Not enough information. Full stop.

The click threshold, done properly

"Negate after 10 clicks with no sales" is the advice everyone gives and it's wrong for most products. The right number comes from your own conversion rate:

When a term converts but costs far too much: the converter ladder

This is the case people get wrong in both directions. One school says "it converts, never touch it." The other says "it's above 2× margin, block it." Both are too blunt, and a rule of either kind will cost you real revenue.

A term with orders is neither automatically protected nor blockable at some ACoS multiple. When a converter is running far above break-even - say 2× or more - that is the trigger to run this ladder, not a verdict. Work down it and stop at the first step that resolves the term.

1. Is it self-exact?The term is, or close-variants to, its own exact keyword in that campaign. Then this is a bid question and nothing else - a negative there either does nothing or switches the campaign off. Lower the bid. Stop.
2. Is it significant?Does the term carry a meaningful share of its ad group's or its family's conversions? If so it is protected from negation outright. Fix the economics through the bid or the harvest path. You do not repair a margin problem by deleting the traffic that is producing your orders.
3. Is the bid actually tunable for this term?Ask whether a bid change reaches only this term. If it effectively has its own bid - a single-keyword exact campaign - bid it down and stop. If it sits under a shared bid (broad, phrase, auto, or a stacked campaign), check the collateral first: look at what else that keyword is bringing in. If its other search terms are performing at or under target, cutting the shared bid punishes the good traffic to fix the bad. The bid is the wrong lever here.
4. Then harvest, then negate - as one move.A converter that loses money at the shared bid but works at its own price gets graduated: give it its own exact campaign, bid it to its own economics (its revenue per click × that product's target), and only then negate it in the discovery source. The two ship together, deliberately. The traffic moves to a bid you control instead of disappearing.
5. Abandon it - and show your working.Negate a converter outright only when it cannot be made profitable at any realistic bid and its conversion volume is insignificant. The test is arithmetic: its maximum affordable CPC (revenue per click × the product's ACoS ceiling) sits below any plausible auction floor. Write that number down when you do it. If you can't produce it, you haven't finished the ladder.

Check 90 days before you pull the trigger

Where to put the negative

Default to negative exact at ad-group level. It's the narrowest, most reversible thing you can do. Negative phrase blocks everything containing that word and should never be automatic - one careless negative phrase can take out half a campaign's traffic.

The word-level view Advanced

Sometimes the problem isn't a term, it's a word. Break every search term into its individual words, pairs and triples, total up the clicks and orders on each, and compare its conversion rate against your account average. Words converting well under half the account rate, with real spend behind them, are usually worth a phrase negative - "refill," say, when you don't sell refills, appearing across forty different search terms.

It works in reverse too: words converting comfortably above your average are a shortlist of keywords worth launching.

Advanced A monthly audit. Not everything that looks wasted is.

Run this monthly against a 30-day window ending at least a week ago - you want everything fully attributed. Then sort every spending search term into one of three buckets, and never add them together:

Separating these is the whole discipline. Lumping a rank campaign's intentional overspend in with genuine waste produces a scary number that nobody can act on, and it teaches your team to distrust the report.

Seven kinds of "waste," and only one is safe to negate

Reading the total

Advanced Quarterly, or when growth stalls.

Harvesting finds winners inside your own account. This finds the opportunity outside it: keywords where competitors rank above you organically and you aren't advertising at all, or aren't advertising properly.

Pull three things and join them: your organic rank per keyword, the same for the competitors above you, and your current campaign coverage. Then flag every row:

Gates, so the list is short enough to act on

The expand verdict

For keywords you already run, compare your conversion rate on them against the market's. If yours is at or above market, expand - you convert this traffic at least as well as everyone else, so buy more. If yours is below market, hold: the constraint is your listing, not your bid, and spending more will just lose money faster. Launch those paused or at a reduced bid, and put the listing on the fix list.

Advanced Genuinely optional. Don't do this until the rest is running well.

Dayparting means bidding more at the hours that convert and less at the hours that don't. It's real, and it's smaller than people expect - and if you set it up carelessly it interacts badly with everything else you've already tuned.

What the data can and can't tell you

Work in six four-hour blocks across seven days, on four weeks of data. For each block, compare its revenue per click against your overall revenue per click - that single number folds conversion rate and order value together, which is what you actually care about.

Amazon has no native time-of-day control, so applying a schedule needs a rules tool sitting on top - in AMZ Vault the dayparting analytics derive the blocks from click-floor-gated hourly data and the schedule applies them for you, so the method below is what to ask for rather than what to compute by hand. Re-check in three to four weeks: did the blocks you boosted actually perform? If the weekday tilt didn't hold up, drop it and keep the hourly pattern only. And turn dayparting off entirely during peak events - Black Friday hour patterns look nothing like a normal week - running manual control with roughly a 25% overspend allowance instead.

Start here

It's worth being blunt about the order of operations. If your conversion rate is poor, every keyword in the account costs more than it should, because you're paying for clicks that don't turn into orders. Lifting conversion from 8% to 10% is mathematically identical to cutting every CPC in the account by 20% - and it's usually easier.

Which means: when advertising looks broken, check the listing before you touch a bid. The tell is in the funnel:

Start here The rules changed on 27 July 2026. If you haven't updated, this is urgent.

Amazon split the title into two fields, and both are searchable:

• Item Name - 75 characters. This is your identity: what the thing is.
• Item Highlights - 125 characters. This is relevance: the phrases that earn you traffic.

How to write the Item Name

1. The defining attribute and materialWhat makes it this product and not the generic one.
2. The product noun, inside the first 3-5 wordsNon-negotiable. If a shopper can't tell what it is from the first few words, the rest doesn't matter.
3. The audience, if it's a real filter"For Women" earns its space if it genuinely narrows. "For Everyone" does not.
4. One qualifying featureThe single strongest one.
5. One style word, optionalOnly if it doesn't push the product noun past word five.

Do not pad to the limit. Brand, product noun and one differentiator landing at 48 characters is a good title, not an unfinished one.

How to write Item Highlights

Three to five short phrases separated by semicolons, most decision-driving first. Rough priority: key functional benefit, then material or safety, then use case, then gift or occasion.

Truncation - where titles go embarrassingly wrong

Things that will get you rejected or throttled

No "best" or "#1." No "eco-friendly" without a certification. No medical claims. No competitor brands. No promotional language ("Sale," "Free Shipping"). No ! $ ? _ { } ^. No ALL CAPS. No em-dashes - use a spaced hyphen. Pipes are technically allowed but waste a character and read like a 2019 keyword-stuffed title.

Once you're running

Five bullets, each doing one job

Write exactly five. Each one should answer a real objection or question - something from your reviews, your returns, or your Q&A - not restate a feature. Lead each with a short capitalised phrase so it's scannable, then one line of plain explanation.

Specifics beat adjectives every time. "Fits laptops up to 17 inches" is worth ten instances of "premium quality." And whatever you claim, make sure it's true and verifiable - a number you can't support in a bullet is a return waiting to happen.

Description and A+

1,200-2,000 characters of description, in your brand's voice, with an actual argument - not the bullets pasted again. Then three to five A+ modules with a story arc: main benefit, then how it works and what it's made of, then the outcome in real life, then a comparison, then what's included or the questions people always ask.

Three practical constraints: A+ text is not keyword-indexed, so write it for humans. A+ should complement your image stack, not repeat it. And keep everything important inside the middle 60% of the module width - the outer thirds get cropped on mobile, which is where most of your traffic is. If three modules say everything, ship three. Padding to five is visible.

Backend search terms

249 bytes, invisible to shoppers, indexed by Amazon. Use all of it. Rules: no repeats of anything already in your title or bullets (already indexed - wasted), no competitor brands (against policy and it doesn't work), no punctuation, no repeated words.

Fill it in five buckets: primary terms you haven't used elsewhere, semantic variations, long-tail phrases, the conversational way people ask for this ("what do I use to keep a laptop dry when cycling"), and problem-solution phrasings.

Once you're running

The main image is a click instrument

It has one job: earn the click in a grid of competitors. Amazon's rules are strict - pure white background, product filling 85%+ of the frame, photographs only, no added text or graphics, no props you don't sell, 1600px minimum on the long side.

Within those rules there's more room than most sellers use. Five distinct approaches, worth testing as five separate hypotheses rather than five versions of one idea:

The rest of the gallery, slot by slot

Front-load. Attention decays fast down the gallery, so slots 2 and 3 are where you answer the questions that actually stop a purchase. If a comparison image names competitors, don't - and make sure every claim in every cell is something you can back up.

Video

Video sits around the second position in the gallery and is widely reported as a 2-3× conversion lever. If you have no video showing the product in use, at scale, or being set up, that's usually the highest-value single thing on your listing to-do list - and it matters more now than it used to, because AI shopping assistants will answer "how big is it actually" from somebody's content, and if it isn't yours it's a competitor's.

Advanced New, real, and most sellers haven't adjusted.

Amazon search is no longer only keyword matching. There's now an intent layer that asks whether your product fits the need behind the query, and a shopping assistant that reads whole listings and compresses fifty results down to about five recommendations.

The practical consequence: about 15% of one-word searches trigger an AI answer, versus roughly 68% of ten-word searches. So the longer and more conversational the query, the more this decides whether you're seen at all. And there appear to be hard filters - being in stock, and a rating somewhere around 4.0.

The five relations

Interview the assistant about your own product

Open the shopping assistant on your own listing and ask five questions, in a shopper's voice, and write down the answers verbatim:

1. What is this product for?
2. What do people like about it?
3. What don't people like about it?
4. What are people buying instead?
5. Why should I choose this over those?

Question one is the important one. If the assistant's answer to "what is this for" is off-base, everything after it is built on a wrong premise - and that's your listing's fault, not the AI's. Fix positioning before anything else.

Question three is the most useful. If the complaint it surfaces is something your copy could have prevented, that's a free fix. If it's a real product problem, that's a supplier conversation. And if the assistant says something simply untrue, that's a hallucination - note it, don't let it change your listing.

Your Q&A section is training material

Alt text

Your A+ images have alt-text fields and most sellers leave them empty or stuff them with keywords. Write a real sentence under 100 characters describing what's in the image: product, feature, use case. That text is machine-readable when the image isn't. Don't repeat the same alt text across images.

Once you're running The order matters more than the copy.

A full listing rebuild touches title, bullets, images, A+, backend terms and probably your targeting. Change them all at once and you'll never know what worked - and if the listing gets worse, you won't know what to undo.

The order to work in

1. Know who you're selling toWrite down five to ten real customer profiles, including at least one who would not buy and at least one skeptic. If they all agree with each other, you've written flattery, not research. The useful moment is when two of your best-customer profiles split on the same question - that's an A/B test, not an opinion.
2. Compare what the listing says against what you're paying forPull the search terms you spend on and check whether the listing actually contains those concepts. Terms you spend heavily on that the listing never mentions are either a copy gap or a targeting error - and knowing which is the entire question.
3. Interview the assistant and read your reviewsForty reviews, balanced: twenty most-helpful and twenty critical. Critical reviews matter more than their number suggests, because that's what gets quoted back at shoppers.
4. Plan the targeting changes but don't make them yetWrite them down and park them. You're about to change what the page says; the targeting should follow the new positioning, not the old one.
5. Write the title first, then the imagesThis order specifically, so text on your images can't contradict the title you just wrote.
6. Ship one change at a timeTitle and highlights first. Then images, one swap at a time. Then A+. Then, about two weeks after the listing settles, unpark the PPC changes.

Start here Check this daily. It takes two minutes.

"Everything you can sell" means FBA on hand, plus FBA inbound, plus anything sitting in AWD that's ready to go. Then:

Low stock isn't the alarm. Low stock with nothing coming is.

The wall - and how it flips every decision

Your wall is the earliest date new stock could possibly arrive: today plus order lag, plus production, plus shipping, plus a safety buffer. For most importers that's somewhere between 60 and 120 days.

Every SKU is then either short of the wall (it will run out before new stock can land) or long past it. And which side it's on completely inverts what you should do:

Advanced Three months out is reliable. Twelve is a planning exercise.

A forecast is four things separated from each other: a baseline level, a seasonal shape, one-off events, and damage from stockouts. Mix them up and you'll forecast last year's disaster as this year's plan.

1. Separate the seasonal shape - per productWork out each month's typical index against the year, for that product family specifically. You need 18-24 months of continuous history for this to mean anything - with less, borrow from a product with the same demand driver, or hold the curve flat and say so. Honesty about a flat baseline beats a confident fake seasonal curve every time.
2. Strip out event spikesPrime Day and equivalents. Take the lift out of history, then put it back on the month it'll fall in this year. Watch October specifically - if Amazon moves its autumn event, an October forecast that assumes it will be overstated by roughly 8-10%.
3. Find and repair stockout monthsA month whose de-seasonalized sales fall below about 35% of its neighbours was almost certainly a stockout, not a demand collapse. Exclude it from your baseline and substitute a healthy figure. Don't tighten that 35% - genuine demand dips land at 40-60%; real stockouts crater well below.
4. Set the baselineBlend the last few complete months with the current month-to-date, weighted roughly two-thirds to one-third. And anchor the current month to what actually sold so far plus the remaining days at your run rate - never model the month you're standing in purely from seasonality.
5. Gate recovery on inventoryRecovering rank after a stockout takes four to eight weeks. But only credit recovery you can actually stock: under 45 days of cover, hold the forecast at current velocity and flag it. Forecasting a recovery you have no units for is how a plan quietly becomes fiction.

Two forecasts, side by side, always

Produce a "fully stocked demand" number and an "actually available" number. The gap between them is the headline. That's the sales you're leaving on the table for want of inventory, and it's usually the most persuasive number in any purchasing conversation.

Once you're running Monthly, without exception.

Amazon loses inventory, damages it, receives less than you shipped, and refunds customers for items that never come back. Some of that it reimburses automatically. A meaningful amount it doesn't, and it will not chase you about it.

What to look for

Returns-not-restocked is typically the biggest pot. Inbound shortages are the one Amazon has no report for at all - you have to reconcile shipped against received yourself, which is exactly why most sellers never claim them.

Filing

• Always check the reimbursements report first. Amazon auto-reimburses a lot now. Claiming something you were already paid for wastes everyone's time and gets your cases looked at more sceptically.
• Five to ten cases a day. Amazon throttles, and a flood gets handled worse than a trickle.
• State units, never dollar amounts. Let Amazon value it.
• Follow up after seven business days.
• Never re-file a denied case. Reply on the existing case and ask for escalation - re-filing puts you at the back of the queue with a fresh, less-informed agent.
• Inbound shortages go through the shipment's own research request, not a support case, and need a supplier invoice as a PDF or photo. Amazon rejects Excel and Word.

Once you're running A sequence, not a hunt. Follow it in order.

First, pick an honest benchmark

Then work the sequence

1. Split revenue into its three partsSessions, conversion, price. Sessions down with conversion flat is a demand or ranking problem. Sessions flat with conversion down is a listing, price or competitor problem. Both down usually means you ran out of stock. Sessions up with conversion down is the worst one - you're buying traffic your listing can't convert.
2. Check whether the market movedLook at search volume on your category's head terms year over year, not at your own impressions. If category volume fell 20% and your sessions fell 20%, you're tracking the market and that's not recoverable by working harder. If the category is flat and you're down 20%, that's share loss and it is recoverable. Check a competitor's brand-name search volume too - if theirs is down as well, the whole category is soft.
3. Decompose before you believe any account-level numberTwo things corrupt account rates. New launches convert at roughly a third of a mature catalog's rate and drag the average down while nothing is actually wrong - so report the mature catalog separately and lead with that. And re-parenting moves revenue between rows without a unit changing hands - build your product map from the current period and apply it to both sides.
4. Check fees and returnsFulfillment fee per unit above about 25% of price means go check dimensions and weight. On return rate, read the reasons, not the number - they tell you whether it's your listing, your product or ordinary remorse.
5. Then, and only then, look at the adsIn this order of payoff: placements first (Product Pages is usually taking more spend than it deserves), then match types (exact with your lowest conversion and highest CPC means you're overbidding head terms), then search terms by intent, then whether your branded terms are starved, then budget-capped winners.
6. Price lastBecause it's the stickiest. And check whether a past price cut actually bought volume - if it didn't, it's a standing loss you should reverse. A price that's been "on sale" for thirteen months isn't a promotion, it's your price.

Fix things in order of how free the money is

1. Broken listings and stockouts - pure lost revenue, costs nothing to fix
2. Spend that's losing money outright
3. Efficiency work - bids, placements, negatives
4. Price
5. Fees, reimbursements, and everything else leaking at the edges

Put a dollar figure on each. A plan where every line has a number attached gets executed; a plan of adjectives doesn't.

Start here

Every day - ten minutes

• Check days of cover. Anything at 14 days or under with nothing meaningful inbound gets ordered or shipped from AWD today.
• Glance at yesterday's spend for anything that looks nothing like the day before.
• Check for suppressed or inactive listings on products that have inventory or ad spend behind them.

Every week - one to two hours

1. Read the week against the one beforeEnding three days ago, both windows the same length. Look for: spend growing faster than sales by 5+ points, any placement or ad type whose ACoS worsened by 5+ points on real spend, and campaigns running above twice their product's break-even ACoS on $100+ of spend.
2. Run the bid rulesChapter 8. One change per lever, log every one.
3. Look at where the money wentTop twenty campaigns by spend. Anything surprising gets opened.
4. Write down what you changedDate, entity, from what to what, why. This takes two minutes and is the single highest-return habit in the whole guide.

Every two to four weeks

• Harvest (chapter 10) and negate (chapter 11). Always as a pair - you're reading the same report.
• Re-check any change you made last month: did it help? Compare the 14 days after against the 14 days before.

Every month

• Refresh your margins from the last fully settled month. Flag any family whose margin moved more than five points.
• File reimbursements. Non-negotiable - the 60-day clock doesn't care that you were busy.
• Listing health sweep - anything broken, suppressed, unpriced, or with a rating sliding toward 4.0.
• Waste audit (chapter 12).

Every quarter

• Gap analysis (chapter 13) - where competitors are beating you.
• Re-forecast and plan purchasing against the wall.
• Pick one or two products for a full listing rebuild (chapter 20).
• Review which of your rules are actually working, and change the ones that aren't.

Start here The short list. Every one of these has cost somebody real money.

Reading the data

• Judging performance on a window ending yesterday. Sales take up to 7 days to be credited. You will revert working changes.
• Comparing ACoS against TACoS, or an ad conversion rate against a business-report one.
• Believing page one of a report. If it's paginated, page all the way through before you total anything.
• Reading a campaign report as a product report. A campaign's spend includes every product in it, which can double what you think one product costs.
• Comparing a parent ASIN's numbers against a child's.

Advertising

• Negating a keyword inside its own exact campaign. Pause it or lower the bid instead.
• Negating something that converted two months ago because it hasn't lately. That's a diagnosis, not waste.
• Negating your own brand or category term.
• Scaling up spend on a product with under 14 days of cover.
• Freezing a whole campaign because one variation is low. The in-stock siblings absorb that traffic. A campaign only goes reduce-only when everything in it is short.
• Launching a new campaign into a product with under 30 days of runway with no confirmed inbound.
• Cutting a placement modifier by 25 points when you meant 25 percent. On a modifier of 33 that is a 76% cut, and it will take a rank campaign's position with it.
• Negating a converting search term on an ACoS multiple alone instead of running the converter ladder.
• Chasing placement modifiers before fixing the base bid.
• Running two broad campaigns in the same product family.
• Turning on loose match and complements before the family has two weeks of negation behind it.
• Changing four things at once and then trying to work out which one did it.
• Using one ACoS target for the whole account.

Listings

• Leaving a title over 75 characters. The SKU gets rejected and Item Highlights is suppressed with it.
• Uploading a flat file as "Update" instead of "Edit (Partial Update)." It blanks every field you left empty.
• Swapping the live main image to test it instead of testing from the last gallery slot.
• Writing a claim into a listing because a search term wanted it.
• Changing the listing and the targeting in the same week.

Stock and money

• Reading cover at the parent while the best-selling child is at zero.
• Discounting into a stockout. Short of the wall you should be raising price, not cutting it.
• Reordering a discontinued SKU because the low-stock alert didn't know it was discontinued.
• Auditing reimbursements quarterly. The window is 60 days.
• Sending Amazon your cost data to argue a reimbursement value.

Start here