Ask Claude What to Change in Your Amazon PPC This Week

✨ AI-assisted research, automated editorial review by Avanta Global EOOD. Learn more

"What should I change in my ads this week?" is the question every Amazon seller asks on Monday, and the one the ads console answers worst. It shows ACoS, which is not the same as losing money, and it suggests bids without knowing what your product earns after Amazon's fees.

This post shows the question asked in Claude, connected read-only to AgentXray's public demo store, and what came back. Every number below is from that demo store, a synthetic US seller, on 23 October 2026. None of it is a real account.

What came back

Claude called AgentXray's recommendation tool, which reads the store's Sponsored Products reports and prices every suggested change against the store's own break-even. The answer was a short list, in order:

1. Add 5 negative keywords. Three search terms had heavy clicks and no orders, costing $531 over 60 days: "plastic cutting board toy" (189 clicks), "cheap knife set" (207 clicks) and "cutting board for kids" (50 clicks). Each one had passed the evidence bar (enough clicks or enough spend with no order) before it was suggested.
2. Cut three keyword bids to break-even, from $1.35: coffee scale 0.1g to $1.11, bamboo cutting board to $1.06, knife block set to $1.13. All three convert, at about 15%, but cost more per order than the product earns.
3. Cut the auto-campaign targets harder than suggested. This is where Claude disagreed with the tool, and said why.
4. Skip a budget cut. Moving one campaign's budget from $16.07 to $16.00 a day changes nothing; the bid cuts are the lever.

Where the AI disagreed with the tool

AgentXray suggested cutting two auto targets (close-match and loose-match) from $1.35 to $1.22. Claude read the same output and noticed the break-even cost per click on those targets was about $0.58, so $1.22 still loses money on every order. Together they took about $1,110 of spend in 30 days, at about 26% ACoS.

Its call: take them to about $0.95 now, and step down again after the next read, because cutting straight to $0.58 risks losing most of that campaign's volume in one move.

That is the useful shape of this tool. The recommendation is a priced starting point with its evidence attached, not a verdict, and the AI on top of it is free to argue.

Why the break-even was 15.9% and not 20%

The answer ended with two caveats. The first is the one most PPC tools get wrong. Break-even ACoS is a property of the price, not of the account. This week the demo store sold at $46.21; its settled price over the last 30 days was $49.13. At this week's price the break-even ACoS is 15.9%; at the settled price it is 20%. Claude said both, and added that if the lower price is temporary, the deeper bid cuts should be eased.

A tool that quotes one break-even for the whole month would have told you every cut above was too aggressive, or not aggressive enough, depending on which week you asked.

Nothing changes on Amazon by itself

This conversation used the public demo, which is read-only, so Claude could only recommend. On your own account, a connection that you allow to draft turns each recommendation into a draft: the exact bid, negative or budget change, filed into your AgentXray approval queue. Nothing reaches Amazon until you open it in your browser and approve it. The token your AI holds cannot approve its own drafts.

Each approved change is then followed for 14 days and judged on net contribution after fees and ad spend, not on ACoS; that is the "re-check on Nov 6" in Claude's answer. Once a few verdicts exist, you can ask which kinds of change have actually worked on my account and get the tally.

Asking it on your own store

The tools behind this answer are get_ads_recommendations and get_change_scorecard, with ppc_audit and harvest_keywords as ready-made prompts. All are listed with sample answers in the MCP tool reference.

• Connect Claude or connect ChatGPT to your Amazon Ads and Seller Central accounts.
• Or try it on the read-only demo store first: the consent screen offers "connect the demo store instead".
• Why profit per day beats ACoS as the judge: ACoS vs profit.
• How the approval step works: AI drafts, you approve.

FAQ

Can Claude manage my Amazon PPC?

Claude can read your campaigns through an MCP connector and recommend changes. With AgentXray it can also file those changes as drafts, but it cannot apply them: every change waits for you to approve it in your browser.

How is the break-even bid calculated?

From what one unit actually earns after Amazon's fees, taken from your settlement lines, divided across the clicks it takes to make a sale. A bid above that loses money on every order the click brings in.

Why does the recommended break-even change from week to week?

Because it depends on price. When your selling price drops, for a deal or a coupon, the margin left for advertising shrinks and so does the break-even ACoS.

Does it work with ChatGPT?

Yes. The same tools work in ChatGPT, Claude and any other assistant that supports MCP.