Audit Amazon Ads AI defaults before they can widen your reach

Amazon Ads is giving advertisers more ways to influence its models without turning those inputs into targeting walls.

In its July 22 announcement, Amazon described ten Brand+ and Performance+ capabilities that had launched or were rolling out over the next few weeks to Open Beta or general availability. Audience Signals are an Open Beta for Brand+ Prospecting and Performance+ Customer Acquisition and Unified Consideration, with access varying by feature and advertiser. In those tactics, first-party audiences added through Ads Data Manager are optimization signals, not hard targeting constraints.

Some optimization and supply features can be enabled by default in defined campaign types. An operator therefore needs to know which settings express intent, which settings suppress delivery, and which settings Amazon may expand automatically.

Read audience signals correctly

An audience signal tells the model where useful patterns may begin. It does not instruct the campaign to serve only that audience.

That distinction changes how a team should approve an upload. For each advertiser or AMC audience, document the business hypothesis, source system, refresh cadence, eligible marketplace, consent basis, expected behavior, and conversion event. State whether the audience represents high-value customers, recent consideration, category interest, churn risk, or another observable condition.

If the seed contains repeat purchasers, the campaign may learn characteristics that help it find prospects beyond the list. That can be the intended job of prospecting, but it is not targeting only known customers.

Use exclusions as explicit suppression controls. Amazon says they remain available and operate independently of Audience Signals, but operators should still verify their configured behavior and actual delivery. Past purchasers, employees, test accounts, or another applicable suppression group belong there—not in a hope that an optimization signal will avoid them.

Inventory every automatic control

Default-enabled features are still campaign decisions, even when no one clicks an activation box.

Amazon says Automatic Deal Selection is enabled by default on new Brand+ streaming TV prospecting lines in the United States. It can broaden delivery across optimized premium deals on Amazon and the open internet. Amazon also identifies narrower default scopes for Brand Outcome Optimization on reach-goal lines and Prime Video signal integration on Prime Video Ads supply.

Build a configuration ledger at the line-item level. Record marketplace, goal, tactic, format, inventory sources, audience signals, exclusions, conversion event, bid approach, budget, frequency, brand-safety settings, optimization features, and activation time. Mark each value as selected by the team, inherited, or enabled by Amazon.

Save the approved configuration and the live configuration. Without both, a delivery shift can be misdiagnosed as creative fatigue, seasonality, or audience quality when the supply or model changed.

Define the campaign contract

Automation needs a commercial assignment before it needs more data.

Write one sentence describing the campaign’s job: acquire new customers within a contribution threshold, build qualified reach in a defined market, or recover recent consideration without overserving existing buyers. Connect that job to one primary KPI and a small set of guardrails.

The contract should cover maximum spend, geography, inventory boundaries, frequency, excluded audiences, approved creatives, conversion definition, attribution window, product availability, margin floor, and escalation conditions. Audience Signals do not carry an additional feature fee, but broader discovery can still spend the campaign budget in places the previous setup did not reach.

Assign owners separately: media for optimization, brand for placement standards, data for audience and conversion integrity, finance for economics, and one release owner for the live configuration.

Measure expansion, not only efficiency

A model can improve reported efficiency without proving that the advertising created incremental demand.

Amazon’s Performance+ guide explicitly notes that Performance+ does not directly measure incrementality. Plan the comparison before launch: a supported holdout, a matched geography, a staggered rollout, or a stable baseline with clearly documented limitations. Keep price, promotion, inventory, and major creative changes out of the test window when practical.

Review more than the headline KPI. Track spend, unique reach, frequency, supply delivery, new-to-brand behavior where available, conversion lag, returns, and contribution after advertising.

Do not interrupt normal learning because one early day looks weak. Do intervene when spend breaches the approved limit, an exclusion fails, inventory becomes constrained, a placement violates policy, or measurement inputs break. Learning flexibility is not permission to ignore a control failure.

Roll out by line, not by account

Start with one product family, market, or campaign job whose economics are already understood.

Capture the prior delivery mix and outcome baseline. Launch with a dated settings record, confirm that exclusions and conversion events are functioning, and review where the line actually delivered. At the end of the declared window, choose one action: keep, narrow, expand, or reverse. Record why.

Expand only after the release works technically and commercially. A clean setup confirms that Amazon received the controls. A sound result may be consistent with improvement; causal confidence depends on the comparison design.

This staged approach also creates a useful failure boundary. If an automatic feature behaves differently than expected, the team can isolate the line, preserve evidence, and correct the contract without disturbing the entire media plan.

The Operator Read

The new Amazon Ads controls are not a choice between human judgment and AI optimization. They divide the work.

The advertiser supplies a goal, credible signals, explicit exclusions, economic limits, and a measurement design. The model searches for delivery opportunities inside the controls Amazon exposes. Operators then reconcile what was configured, where the campaign delivered, and whether the result added profitable demand.

Treat signals as guidance, exclusions as suppression controls to verify, and defaults as decisions that require an owner. That is how a brand gains the reach of automation without losing the ability to explain—and reverse—what changed.