Should you spend more or less on every Amazon product you sell?

Two sellers run the same ad on the same keyword. One buys a brand-new customer. The other rings up a regular who was going to reorder next week anyway. Their ROAS is identical. One of them is quietly burning money.

The number they both steer by cannot tell those two dollars apart. ACOS, ROAS, a blended MER, they all measure ad spend against sales, and none of them asks the one question that decides whether a dollar was worth it: did it bring someone new? Answer that, keyword by keyword, and the right move often flips. The keyword your ROAS dashboard loves can be the one bleeding cash.

This is a full working plan. It reads long because it is complete: the idea, the method, one real account worked end to end, and the exact moves that come out the other side. Use the map below to jump to any part.

1. What you’re optimizing: two questions

Start with the goal, because it is not the one most sellers chase. The aim is not a good ROAS. It is the most 12-month contribution profit, in dollars. That is what a customer leaves you over their first year after everything it took to win and serve them: the product, the fees, the shipping, and the ads that acquired them. Only fixed overhead is left out.

The spend curve builds that number, and shows how fast it fades. Each extra ad dollar buys a few more customers, then a few less. So the rule is simple. Keep spending while the next dollar brings back more than a dollar of profit, and stop the moment it brings back less. Both questions below serve that one number.

Question one, the customer: is this one worth winning? The spend curve answers it with LTP:CAC, the lifetime profit a customer brings over a year for every dollar you spent to win them, measured from real profit, not a guess. It also shows how fast returns fade as you spend more.

Question two, the keyword: is it winning anyone new? The new-to-brand rate answers it, the share of ad-driven orders that came from someone who had never bought the brand before. Above 50 percent, the keyword brings new customers. Below 50 percent, it mostly rings up the ones you already had.

One number without the other misleads. A valuable customer won on a returning-customer keyword was never really acquired. A cheap new customer on a product nobody reorders is barely worth the click. Read the two together, and the next dollar knows exactly where to go.

2. The method, step by step

Run it on every product, not just the top sellers. Start at the top for quick wins. Then keep going. The long tail is where wasted spend usually hides.

1. Rank every product, top first. Start where the acquisition volume is, then keep going. No product should escape the two questions.
2. Read each product’s spend curve. LTP:CAC, the acquisition cost, and how fast the next dollar loses efficiency (with a confidence score, so you know when to trust the curve).
3. Sort into spend-more and spend-less. Scale, grow, hold, or reduce, based on LTP:CAC, using the diminishing-returns curve only where it is statistically reliable.
4. Pull the keywords behind those products. Top search terms per product, branded versus non-branded, ranked by purchase share, with a funnel read that flags listing problems masquerading as spend problems.
5. Score each keyword’s new-to-brand rate. Keyword by keyword: is this term bringing in new customers, or re-selling to existing ones?
6. Measure the split, then set each keyword’s ACOS ceiling. For each product, work out what share of its ad spend goes to acquiring keywords versus returning ones. Then set a spend limit per keyword. A keyword that acquires new customers earns a higher limit, because the customer pays back over a year.
7. Cross the two and reallocate. Spend more on high-NTB keywords for valuable customers. Spend less on low-NTB returning keywords. For a valuable product stuck on those keywords, open new acquisition routes rather than bidding up the same terms.

3. The worked example: the spend curve

This account sells health supplements on Amazon. The table shows the top products, but the read runs on every SKU. The bands are simple: Scale at 4x or more, Grow from 2.5x to 4x, Hold from 1.5x to 2.5x, Reduce below 1.5x. The last column shows whether the response curve was reliable enough to trust.

Two facts drive the rest. The flagship brings in most of the new customers, so it sets the budget. And nothing clears the 4x Scale line. This is a solid, competitive catalog. The gains come from moving money around, not from one breakout product.

4. The keywords

Next come the search terms behind each product, ranked by purchase share. Purchase share is your slice of every purchase on a keyword across the whole market, not just the sales from your ads. Branded terms in this account convert at 60 to 98 percent share but drive only about 10 percent of purchases. The other 90 percent happen on generic category terms. That is where the market is.

Purchase share does a second job, and this is the part most sellers skip. Hold 10 percent of a keyword and 90 percent of its buyers are still up for grabs. Hold 1 percent and 99 percent are. That gap is your room to grow. It also predicts new-to-brand before you even open Amazon Marketing Cloud. A keyword you barely hold is full of shoppers who do not know you, so ads there tend to bring new customers. A keyword you already own is mostly your own buyers, so ads there mostly reach people who would have come back anyway. Low share points to acquisition, high share to returning buyers, and the new-to-brand pull confirms it.

The funnel shape is the last check. When a product’s share rises from impressions to clicks to cart to purchase, it wins once shoppers see it. The only limit is how often it shows, and spend fixes that. The flagship is the clear case: about 5 percent impression share, but 19 to 31 percent purchase share on generic terms. When share falls through the funnel, the page is the problem, not the bid. The weakest SKU drops from cart to purchase, so its page comes before any budget.

5. New-to-brand, keyword by keyword

Now the second axis, pulled from Amazon Marketing Cloud at the keyword level. It separates a term that merely ran ads from a term that actually brought in new customers.

The flagship earns its acquisition budget. Its generic terms run 60 to 91 percent new-to-brand, and it even wins new customers on a competitor’s brand name. Put its extra spend on those generic terms, not on the branded ones it already owns.

6. The cost of a new customer, on one scale

Amazon gives three numbers for every keyword: the click cost, the conversion rate (orders per click), and the new-to-brand rate (new customers per order). Chain them and acquisition becomes arithmetic.

In plain words: start with what a click costs. If only some clicks turn into orders, divide by that. If only some of those orders are new customers, divide again. What is left is the price of one new customer.

Conversion rate is the quiet lever. Two keywords can share a click cost and a new-to-brand rate, yet one buys a new customer for far less because it converts better. A keyword that turns a third of clicks into orders buys customers cheaply. A keyword that turns one click in twenty into an order does not, even at a high new-to-brand rate.

The dollar figure is directional, so read it as a ratio, not a price. Set the product’s average cost per new customer to 100 and index every keyword against it. Below 100 is cheaper than the average. Above 100 is dearer. Because the average customer here returns three times their cost, break-even sits at an index of 300, and any keyword’s own return is 300 divided by its index.

In plain words: 100 is normal for this product. Lower is a bargain, higher is pricey, and 300 is the most you can pay for a customer and still break even.

The scale is a ratio, so it does not depend on trusting the exact dollars. It sorts the keywords cleanly. Fund the low-index terms first. Ease the high-index ones. Anything under 300 still turns a profit, so the index says which keywords return the most per dollar, not which ones lose money.

7. Before you trust the numbers

Two cautions keep this honest. The keyword cost and the new-to-brand rate both come from Amazon Marketing Cloud, and each needs a careful read before it drives a dollar.

And read new-to-brand as a signal, not a headcount. Amazon flags an order as new-to-brand when the shopper has not bought the brand on Amazon in the past 12 months. So it resets after a year, it works at the brand level and not the product, and it sees only Amazon, never your own site or a shop shelf. It says the buyer was new. It does not prove the ad caused the sale. Read a high rate as a reason to lean in, not as proof of a customer the ad created.

8. Where your dollars actually go, and the ACOS ceiling

Two numbers turn the read into a spend decision.

First, the split. Add up each product’s ad spend. Split it in two: spend on acquiring keywords (new-to-brand over 50 percent) and spend on returning keywords (under 50 percent). One number per product, and most sellers never look at it.

The flagship is already right. 93 percent of its spend acquires. Leave it. The smaller products run the other way. On three of them, 60 to 70 percent of the spend goes to returning buyers the brand would have kept anyway. That is the leak, and ROAS kept it hidden because every product looked fine.

Second, the ceiling. ACOS is ad spend divided by ad sales on a keyword. Contribution margin here is about 58 percent. At 58 percent ACOS, ads eat all the first-order profit. Most sellers stop at that line.

A new customer is worth more than one order. Here, a new flagship customer returns about 1.26 times the first order in profit over a year. So a keyword that brings new customers can cost more than the first order and still pay off. The ceiling climbs with the new-to-brand rate:

“An ACOS above 100 percent can still turn a profit, on purpose, because most of those buyers come back.”

Look at the bottom two rows. When 80 percent of buyers are new, 112 percent ACOS still turns a profit. The ad costs more than the first order returns, on purpose, because most of those buyers come back. ROAS would cut that keyword. The rule runs both ways: on a returning keyword the ceiling drops to the margin, so cap those at 58 percent. The formula: ceiling = margin + new-to-brand rate × (lifetime multiple − margin).

In plain words: the more of a keyword’s buyers are brand new, the more you can afford to pay for them, because they come back and buy again.

This is how you add spend without guessing. Compare each keyword’s ACOS to its ceiling. Below the ceiling, add budget. Above it, even the generous version, pull back. On a product that already acquires well, that comparison is the whole job.

9. Cross the two axes

Place each product on the grid of customer value against keyword new-to-brand. The plan falls out.

One more check before spending. A separate ExpandFi measure of acquisition efficiency, built from different data, splits these products the same way: the flagship acquiring, the weak and crowded ones reaching returning customers. Two methods with no shared inputs land in the same place. That is enough to act on.

10. The other half: what you don’t bid on yet

Everything so far tunes the keywords you already run. That is only half the job. A keyword you do not bid on has no ad orders, so it never appears in your own campaign data. A ranking built from your campaigns can never surface it. A complete plan runs two loops. Exploit optimizes what you know. Explore finds what you do not.

In this account, most ad spend sits on keywords the brand already found: exact and phrase terms. The campaigns whose job is discovery, auto and broad match, run near zero. The map is well tuned. The brand mostly stopped drawing it.

Three inputs reopen it. First, missing keywords: cross your bidded terms against Brand Analytics to find terms you already win without an ad. This account had dozens the flagship wins organically yet never bids, a whole cluster of one product spec plus a set of common misspellings. Second, competitors you do not target: product targeting was a thin slice aimed at a few rivals, while the search data named others taking the same shoppers. Third, what auto and broad surface: run them at a low bid as listening posts, read the search-term report they produce, and graduate the winners.

Winners feed back into the same scoring. Each new term runs through the cost-per-new-customer chain and the index. The good ones become new exact keywords or product targets. The rest cost little to test. Judge this work by a different yardstick than the rest: its job is how many new terms and rivals it finds, not its ACOS on day one. This is where the growth budget goes once the known terms are full.

11. Watch for variations

Before you scale a product, check one thing: is it a standalone listing, or one variation in a family? A size, a count, a scent, a color. On Amazon these sit under one parent, and it is easy to read the whole family as a single product. It is not one. Each variation is its own customer, with its own margin, its own repeat rate, its own share of new buyers. A variation that pays its way does not vouch for the one beside it.

The spend curve reads at the variation level, so pull each one on its own numbers. Two variations of the same item can split apart. One returns five dollars of lifetime profit for every dollar of acquisition cost; the other is newer, cheaper to acquire, and busier, but its twelve-month profit rests on a handful of customers, so trust its six-month figure and underwrite it again once it grows up. Same shelf, different bets.

Then the trap that makes this matter. Many sellers run several variations in one campaign, so a single bid change lifts or cuts all of them at once. You cannot feed one and starve another. Read the campaign roster: what share of the spend sits in campaigns that name two or more variations together, and what share is on single-variation campaigns you can steer alone. The per-keyword plan then works at the campaign level, not the variation level, on all the money the two share.

Two things keep this from being a problem. The new-to-brand read attributes each order to the variation that actually sold, so every variation’s cost and profit are its own, not a family average. And when all the variations point the same way, spend more, raising a shared campaign’s bid helps all of them. The trap only bites when two variations under one campaign disagree: one worth scaling, one worth cutting. Then split the campaign, or move that one variation onto its own, so the bid you raise reaches only the product that earns it.

One note on choosing where to start. For a first, low-risk move, pick a standalone product, one parent, one selling child, its own campaigns. In a catalog built on variations there is often a real trade: the big, data-rich products are the tangled families, and the clean standalone ones are newer, with a thinner history behind their numbers. Name that trade out loud. Do not dress a tangled product up as a simple one.

12. What to do

• Leave the flagship’s split alone. 93 percent acquiring is already right. Tune it by the ceiling: move budget off the few head terms running above their limit, onto the high-new-to-brand terms sitting below it.
• Cut the returning-customer spend on the smaller products. That is where most of the waste sits. Where a product is worth keeping, find higher-new-to-brand terms to acquire on instead of dropping it.
• Fix the weakest product’s page before adding spend. Its funnel leaks at the cart. Ads cannot fix a page.
• Treat branded search as defense, not growth. It converts well, but it is a small slice and mostly reaches buyers who already know the brand.
• Scale a variation on its own numbers, not its family’s. If its keywords share a campaign with a sibling, the bid moves both, so split the campaign when the two disagree.

13. Run it yourself, and where it lives

Here is the whole plan as one prompt. It returns the same output shown above: the spend-curve read, the keyword index, the ACOS ceilings, the per-keyword spend-per-day plan, and the discovery list. It runs through the ExpandFi MCP, which is the only place that holds all three data sources together: your spend curve and true lifetime profit, your Brand Analytics search share, and Amazon Marketing Cloud new-to-brand.

# Run this across my ENTIRE catalog for [MARKETPLACE], top products first.
# The goal is the most 12-month contribution profit in dollars, not ROAS.

1. Pull ALL my products from a spend-curve / CAC report, ranked by
   acquisition volume. Start with the top for the fastest wins, then work
   down through every product. For each, get LTP:CAC, effective CAC, the
   spend-response factor, and the model's confidence (r-squared).

2. Classify each product: SCALE (LTP:CAC 4x+), GROW (2.5-4x),
   HOLD (1.5-2.5x), REDUCE (under 1.5x). Only trust the response factor
   where the model is reliable (r-squared >= 0.35); otherwise decide on
   the LTP:CAC ratio alone.

3. For each SCALE/GROW product, check variations first: is it standalone
   or one child in a family? Pull each variation's spend curve on its own
   numbers (do not inherit a sibling's verdict), and read the campaign
   roster for the share of spend in shared vs single-variation campaigns.
   On a shared campaign, a bid change moves every variation at once.

4. For each product (start with SCALE and GROW, then the rest), pull my
   search keywords from Brand Analytics / Search Query Performance. Split
   branded vs non-branded, rank by purchase share, and flag any product
   whose funnel LEAKS (share falls from click to cart to purchase) as a
   listing fix, not a spend play.

5. Pull keyword-level cost, conversion rate, and new-to-brand (NTB) from
   Amazon Marketing Cloud. Per keyword, compute cost per new customer =
   click cost / conversion rate / NTB rate. Index each keyword to the
   product's average (100), and read its return as (LTP:CAC x 100) / index
   (break-even at index = LTP:CAC x 100). Compare ratios, not dollars; the
   clean room reports spend low, so the real CAC is the spend-curve number.
   If a product is too small and the clean room nulls its keyword rows for
   privacy, drop to match-type or campaign grain, and say you did.

6. Set each keyword's ACOS ceiling = margin + NTB x (lifetime multiple -
   margin), and its spend split (share on acquiring vs returning keywords).

7. Cross it all and hand back a per-keyword plan, in spend-per-day:
   - Spend MORE on low-index, high-NTB keywords under their ceiling.
   - Spend LESS on high-index or over-ceiling keywords; pause the drains.
   - For a valuable product stuck on returning keywords, open NEW routes.

8. Explore, not just exploit: cross my bidded terms against Brand
   Analytics to find high-share terms I DON'T bid, competitors I don't
   target, and under-used auto/broad. Graduate winners into step 5.

9. Before executing: confirm each scale candidate is standalone or note
   the variations that share its campaigns, verify any CAC that does not
   reconcile with spend and volume, confirm more than two weeks of
   inventory, and re-pull NTB on a window ending about 14 days ago.

Figures come from one real Amazon account, shown as ratios and percentages, as a worked example rather than a benchmark. New-to-brand is a Sponsored Products measure from Amazon Marketing Cloud, on a roughly two-week lag. The clean room reports spend low, so the spend split and the ceilings hold, since they do not depend on it, while exact ACOS stays directional. This is a planning method, not a guarantee. Seasonality, competitors, and launches are not modeled. Read your own numbers.