Creating campaigns

Everything here is Sponsored Products. It is the only program that can host product ads, and the only one the builder creates.

Two doors, one discipline

Quick advertise — the Auto — quick advertise entry on a product's + New campaign menu, or Advertise this product on a product that has no ads yet. Pick the ads account if you have more than one, and you get a one-product discovery campaign with royalty-aware bids in a few clicks.

The full builder lives on the Advertising page and opens as a slide-over. Three targeting modes: Auto — discovery, Manual — keywords, and Manual — product targets. The product-page slide-over offers the two manual modes; Auto from a product is the quick route above.

The three rules the builder enforces

1. One match type per campaign. Auto, Broad, Phrase and Exact are control levels, and the builder gives a campaign exactly one of them — every keyword in it inherits it, and the default name says which. Adding keywords to a campaign that is already mixed keeps it mixed; the app has a name for that state and an opinion about it, but it will not rewrite a campaign you already have.

2. Bids come from the royalty, not from habit. Every bid field is anchored to the product's ceiling — its average royalty × its conversion rate, the most a click can pay before the sale loses money. Both come from your own last 90 days in that marketplace.

Auto prefills at 85% of the ceiling. Manual prefills by how tight the targeting is: 95% for Exact, 85% for Phrase, 70% for Broad. The invariant is not the percentage — it is that the prefill is never above the ceiling, so on a product with a very low ceiling the prefill sits right at it.

Where the history is thin the ceiling degrades in steps, and it is worth knowing which step you are on: your royalty with the marketplace's conversion rate, then the marketplace's royalty with the marketplace's conversion rate, and finally — for a product with no usable history at all — no ceiling and a flat default bid. That last case is why a brand-new product can prefill at a number that looks unrelated to its economics. It is not a calculation you should trust; it is the app declining to invent one.

A multi-product campaign anchors to the lowest ceiling among the products that have one, so no product in it can be overbid.

3. Graduate on evidence, negate separately. Harvest the search terms that already converted in your Auto campaigns into a manual campaign. Negating the source term is its own opt-in step, never a side effect.

Names that read back

Campaigns are named from a template:

Marketplace | SP | Targeting | Role | ASIN | YYYY-MM
US | SP | Auto | Discovery | B08WLMFS6G | 2026-07

A product type slots in before the subject when there is one, so a name can carry seven segments rather than six.

Role is the part people ask about. Six exist — Discovery, Scaling, Defensive, Seasonal, Second Chance and Cleanup — and the builder assigns two of them on its own: Exact and product targeting become Scaling, everything else Discovery. The rest are yours to type.

The name is a default, not a lock: it is a text field, and it re-derives when you change the targeting until you overwrite it.

Six months later the name still says what the campaign is, which product it feeds, and when it started.

Nothing goes live on its own

A built campaign is a queued campaign. It lands in the Publish queue with its budget, bids and targets attached, and reaches Amazon when you publish.

You can point a harvested keyword at a campaign that is still queued: the queue amends the pending campaign rather than sprawling into duplicates. Two caveats worth knowing — the destination has to be a queued manual keyword campaign for the same product and marketplace, so you cannot graduate a term into a queued Auto campaign; and the amendment is a replace rather than an edit in place, so a crash at exactly the wrong moment leaves you looking at two rows rather than none.

That is the deliberate direction to fail in. Creating a campaign is the one action that cannot be safely retried — running it twice makes two campaigns — so when a run is interrupted mid-send, the row goes to Failed and tells you to check Amazon before queueing it again, rather than quietly trying once more.