Limits and freshness

The honest fine print for MixShift Intelligence. The Intelligence page links here for methodology and coverage details, so this is the article that carries the specifics.

Weeks of cover

Weeks of cover is computed differently depending on where the data comes from, because Seller Central and Vendor Central do not expose the same inputs.

Vendor Central uses Amazon forecast demand over a forward forecast window. The default window is 9 weeks; the engine also supports 6 and 12.
Seller Central uses a trailing-sales window. The default window is 30 days; the engine also supports 7 and 14.

Through the service (REST, MCP, the CLI, and Monthly Report Max) the windows run at those defaults and are not yet a request parameter. The Intelligence Workspace is the only surface where you can choose a different window today.

Vendor Central view

Vendor Central Intelligence supports the Manufacturer view, which provides the more robust operational data the analytical model needs. Sourcing is not supported.

Currency

Formatted figures assume USD today. If your account runs in another currency, read the envelope's own currency field rather than assuming the display units.

Vendor Central runs the comparability scan misses

Seller Central runs get the full comparability scan: dark-period, surge-window, and matched-window detection, described in Reading a bridge. Vendor Central runs do not include that scan today, so a Vendor Central bridge carries fewer of those caveats even where the same kind of gap exists in the underlying data.

Grain: what a coarser run cannot tell you

A run grouped at anything coarser than the product level now says so in its own limitations. This matters for one specific reading: whether a change came from selling a different mix of products or from each product converting differently. At a coarser grain, genuine rate movement gets absorbed into the mix figure, so that comparison is not safe to quote. The monthly request groups by item group by default, so most monthly runs carry this note; ask for product-level grouping when the mix-versus-rate question is the one you care about.

Lost sales and wind-downs

Lost sales estimate what stock-outs cost you. A product you are deliberately discontinuing looks identical to one that ran out, so if the service has no End of Life declarations for the account, a planned wind-down is counted as a supply failure and the estimate reads high. Runs now say so explicitly when no declarations were available, and the figure should be treated as an upper bound until they are. When a declaration did reach the engine and removed something, the run says that instead.

Lost sales is read from an account's daily sales and inventory history. When that history cannot be read for a run, because it took too long or is too large to read in one run, the run still completes: lost sales is left out of the whole run, never priced as zero for one period, and the limitations say so in one sentence while every other figure stays complete. A run cut short by time is not cached, so trying again later may bring lost sales back; a size verdict is cached, so asking again will not. Check the limitations before you read a missing lost sales figure as zero.

Large accounts and async

On the retail-backed requests, a large account can answer account_too_large_use_async instead of computing inline. Retry the same request with "async": true and poll for the result, as covered in Run an insight from the command line.

Concurrency

Each account can have two requests computing at a time. A third answers busy (429); retry shortly rather than immediately.

Cache windows

Results are artifacts, not live queries: the same request, for the same merchant, period, and engine, returns the same figures. A closed month is cached for up to 24 hours. A period that is still open (the current month, for instance) refreshes every 15 minutes.

What is retail-only today

The year-over-year leg of the monthly read is retail-only: there is no year-ago advertising bridge yet, so a YoY comparison covers retail performance only.

Lost sales figures run on their own stamped methodology version, separate from the engine version. Every result, of any request, stamps the engine and evidence versions it was computed with, so you always know exactly what produced a given figure.