The key point is not that Amazon’s advertising stopped growing, but that its growth rate slowed. Moving Prime Day to June could cause promotional search, product browsing and advertising demand to be recognized earlier; the absence of NBA programming on Prime Video in Q3 also changes the available media context compared with Q2 and Q4. These are changes in timing and content mix, not proof that consumers have broadly abandoned retail media.

For Taiwanese businesses, the impact reaches beyond media buying to campaign planning, inventory, product pages and content production. In a hypothetical example, if a home-appliance brand concentrates most annual performance in one shopping festival, a platform or media partner changing the date could move exposure earlier while inventory, customer-service scripts and delivery promises remain on the old schedule. The result would be difficult to interpret: weaker demand or a misaligned operating plan. This is a hypothetical scenario, not a reported case.

Millennium’s editorial analysis is that Taiwanese businesses should not treat a platform’s quarterly growth rate as a direct forecast of their own return on investment. Amazon’s result reflects a combination of Prime Day timing, Prime Video programming mix and a difficult comparison with last year’s search-marketing changes. If a brand evaluates only exposure or sales during one event window, it can mistake media timing for consumer preference.

AI shopping agents should also be treated as a behavior variable to validate, not an assumed market reality. The report says Amazon does not yet see a meaningful impact from external agents and is advancing its own Ads Agent, including full-funnel capabilities across sponsored ads, display, video and streaming TV; these are Amazon’s direction and statements, not evidence of broad availability in Taiwan. Millennium’s judgment is that businesses should first improve product data, need segmentation and attribution, then run small tests to observe whether agent-assisted shopping changes entry points.

Millennium's view：Millennium editorial judgment: Taiwanese businesses should neither withdraw broadly from retail media because Amazon’s ad growth slowed nor assume AI shopping agents have already replaced traffic. The more defensible approach is to validate timing, content, product data and conversion results separately. The reported slowdown reflects Prime Day timing, the lack of NBA programming in Q3 and a difficult search-marketing comparison, so it cannot be reduced to one demand decline. Amazon also says external AI agents have limited impact so far while continuing to advance its own Ads Agent. Businesses therefore need comparable period-based data and small experiments, not a reaction to one platform narrative.

The first practical step is to separate data before, during and after an event, while recording the campaign dates, media placements, product-page visits, add-to-cart actions and purchases, together with supply and content readiness. This does not require replicating Amazon’s advertising system; it creates a minimum basis for comparison. If an event moves earlier, the business can observe whether traffic merely shifted into an earlier period instead of attributing every change to higher spending.

The second step is to reserve a test design for the possibility that AI agents change search entry points. In a hypothetical example, a skincare brand could present the same products through use-case, ingredient and gifting versions, then compare qualified product-page visits and purchase intent during a small, fixed-period test. This is a proposed experiment, not evidence of an existing agent platform or result. The third step is human review: product information must be accurate, inventory must be able to fulfill demand and customer service must answer needs before automation is expanded.

Suggested action 1：Create pre-event, event-period and post-event data views, consistently recording exposure source, product-page visits, add-to-cart actions, purchases, inventory and customer-service issues, with date changes flagged separately. For every event, report completeness and comparable indicators across the three periods, showing where traffic moved rather than reporting only one event total.

Suggested action 2：Create two or three information versions for the same products, such as use case, ingredients or gifting, and test different need-based entry points with a fixed small budget and period without presenting the hypothesis as an AI-agent result. Compare qualified product-page visit rate, add-to-cart rate and customer-service question types for each version, recording the sample period and stopping conditions.

Suggested action 3：Before scaling automated media buying, use a human checklist for product data, supply, delivery promises and customer-service responses, starting with one product group or one event period. Use incorrect product-information cases, stockout or delay cases, first-contact resolution and post-test conversion differences to decide whether scaling conditions are met.