Run AMC queries from Claude

Amazon Marketing Cloud (AMC) is a clean room from Amazon: a place where you can analyze detailed advertising signals that never appear in normal reports, with results returned only as aggregates, so no individual shopper is ever identifiable. Working with AMC normally means writing SQL against its tables yourself. With the MixShift plugin, you ask in plain English and Claude handles the SQL.

Before you start

Your advertising account needs AMC access (an AMC instance). If you are not sure you have one, ask; the plugin can check.
The MixShift plugin installed and signed in. See Install the plugin.

How it works

Ask a question in a Claude chat, for example "What AMC instances do I have?" or "Using AMC, how many of last month's purchasers saw both our display ads and our sponsored ads?"
Claude finds your AMC instance, looks at the data sources available in it, and writes the query.
The query runs on the Amazon side. AMC queries are not instant; expect minutes, not seconds. Claude checks progress and tells you when it is done.
Results come back as a downloadable CSV you can open in a spreadsheet or hand to another tool.

Asking what a customer is worth, not just what an ad did

The tables AMC gives every advertiser only see purchases Amazon could tie to an ad, over roughly the last thirteen months. That is enough to compare ad paths and not enough to ask what a customer is worth over time.

Amazon sells a separate dataset, Amazon Retail Purchases, that covers every purchase of your products in the Amazon store across a much longer window, whether or not an ad was involved. If your account subscribes to it, you can ask the questions that need it:

"What is a customer worth after their first order?"
"Do people who buy the small pack come back for the large pack?"
"Group my buyers by the month they first bought and show me what each group spent afterwards."
"Which product earns the most per customer it wins?"

Two things worth knowing before you start.

It is a paid subscription and it can lapse. Claude checks whether the dataset is switched on for your instance before writing any SQL. That check exists because Amazon's own error for an unsubscribed dataset never mentions the subscription: the query simply fails to compile, which reads like broken SQL and sends you off correcting something that was never wrong. If a query that worked last month stops working, an expired subscription is the first thing to check.

Ask for the period you actually mean. An AMC query carries no dates of its own, so the period comes from the request, and Amazon's default is a single day. A question about a year would quietly answer for one day. MixShift now refuses a query that does not state its period rather than returning that answer, so you will get a clear message instead of a wrong number. Just say the window in your question ("for the last 18 months", "for 2026 so far") and Claude sets it.

Good first questions

"List my AMC instances and what data sources they have."
"How much did display and sponsored ads overlap for purchasers in the last 30 days?"
"Build me a path-to-purchase view: which ad types did buyers touch, in what order?"

Notes

Everything here is read-only. Running a query changes nothing in your ad account.
Results are aggregated by the AMC clean-room rules. Very small groups are suppressed by Amazon, so tiny segments can come back empty; that is AMC working as designed.
Shopper-level identifiers can be used inside a query to group and join, but never returned. A result that comes back empty may be that privacy floor rather than an absence of sales.
If you ask for a period in your own time zone, the boundary is converted to UTC, which can pull a few hours of the next day into the window. On anything grouped by month or by day that shows up as a small extra period sitting beside the full ones. Asking in UTC avoids it.