Google’s test of AI-generated descriptions in Shopping ads changes an important part of the merchant workflow: product data may no longer appear only in the form submitted. It can also become source material for copy generated by the advertising platform. That increases the value of complete product information, but it also expands the surface area for inaccurate claims, outdated promotions and off-brand language.

The correct response is not to disable automation by default or treat it as a hands-off feature. E-commerce teams need an operating framework that controls inputs, defines acceptable outputs and measures business consequences. Because the reported feature is a test, merchants should not assume that it is universally available or that every Shopping placement will behave identically.

Understand what is changing—and what is not

The Shopping-specific development is Google’s reported testing of AI-generated product descriptions in ads. For merchants, the practical implication is that feed attributes, landing-page content and other product information may influence generated wording. A technically valid feed is therefore not necessarily a safe feed. Content can satisfy formatting requirements while remaining ambiguous, unsupported or unsuitable for advertising.

A separate Google update introduced Structured Data Files v10.1 for Display & Video 360, including expanded campaign-management capabilities and support connected to AI-related disclosures. This is useful evidence of a broader operational direction: Google is adding both automation-related controls and tools for managing campaigns at scale.

However, these developments should not be conflated. DV360 structured data files do not constitute a control panel for Shopping descriptions, and the Shopping test does not establish that DV360 features apply to product ads. Neither report confirms universal availability. Teams should verify the capabilities visible in their own accounts and document observed behavior rather than designing controls around assumptions.

Make the product feed a controlled publishing source

Start by treating every feed field as potential public copy. Assign an owner to each important attribute, including title, description, price, sale price, availability, condition, material, size, shipping information and custom labels. Define where each value originates, how frequently it refreshes and who can approve changes.

Run automated checks before submission. Flag contradictions between the feed and landing page, expired sale dates, placeholder text, excessive capitalization, missing variants, malformed units and descriptions copied from suppliers without review. Add freshness thresholds for volatile attributes such as price and availability. A promotion that is accurate when entered can become misleading if inventory or eligibility changes before the feed refreshes.

Build a claim register alongside the feed. For each factual or comparative claim, record the approved wording, evidence owner, applicable markets and expiry date. This is especially important for environmental, health, performance, origin, certification, warranty and “best” claims. If evidence is absent or market-specific, exclude the claim from product data that could be repurposed.

What most people miss

The highest-risk input is often not the primary description. It may be a specification, image annotation, promotional field or landing-page sentence that is accurate only under narrow conditions. An AI system can remove the qualifying context while preserving the attractive claim. Review the complete product-information environment, not just the field that merchants expect an ad to display.

Define output rules before automation scales

Create a concise language policy that both marketing and compliance teams can apply. Divide terms into three groups: approved, conditional and prohibited. Approved terms can appear without escalation. Conditional terms require evidence or mandatory qualifiers. Prohibited terms must not appear, even if present in supplier material.

The policy should cover superlatives, competitor comparisons, urgency, scarcity, guarantees, sustainability language, regulated benefits, audience references and brand tone. Add rules for price presentation: whether “from” pricing is acceptable, how subscriptions are described, which taxes or fees require disclosure and when promotional language must include eligibility conditions.

Translate these rules into detection patterns rather than leaving them in a policy document. Scan feed changes, landing pages and captured ad outputs for prohibited phrases, unsupported numbers and removed qualifiers. Automated screening will not replace legal judgment, but it can make review proportional to risk.

Use tiered review and a clear escalation path

Not every product needs manual pre-approval. Segment the catalogue by consequence. Low-risk products with stable prices and straightforward specifications can use sampled review. Products with regulated claims, frequent promotions, high return costs, contractual brand restrictions or sensitive audiences should receive stricter review and monitoring.

Assign decision rights in advance. Merchandising should own product accuracy; performance marketing should own experiment configuration and delivery monitoring; brand should own voice; legal or compliance should approve sensitive claims; engineering or feed operations should own suppression and rollback mechanisms. Name an incident lead who can coordinate these functions.

Review should include rendered evidence wherever possible: screenshots, query context, date, device, placement, product identifier and the source data active at the time. Store approved examples and failures in a shared log. This creates an audit trail and helps teams identify recurring input patterns rather than correcting one ad at a time.

Test for business impact, not just click-through rate

Do not expose the entire catalogue at once. Begin with a product segment that has reliable data, sufficient traffic and manageable compliance risk. Keep a comparison group wherever account controls and delivery conditions permit. Avoid mixing the test with major bid, budget, landing-page or promotional changes.

Define guardrails before launch. Conversion rate matters, but it can hide costly outcomes. Monitor return and cancellation rates, contribution margin, average order value, customer-service contacts, policy disapprovals, price complaints and claim-related incidents. Examine results by product category, device, market, new versus returning customer and promotional status.

Use a predetermined evaluation window and minimum evidence standard appropriate to the account. Do not call a winner after a brief fluctuation. If generated descriptions lift conversion while increasing returns or shifting demand toward low-margin products, the apparent improvement may destroy profit. Likewise, a small conversion decline may be acceptable if qualified wording materially reduces complaints and disapprovals.

Prepare an incident response process before launch. Define triggers such as prohibited claims, incorrect prices, missing conditions or a spike in disapprovals. The response should specify how to preserve evidence, pause or exclude affected products, correct source data, notify stakeholders and verify that the issue has stopped. Conduct a short post-incident review and update the feed rule or policy that allowed the failure.

Pre-launch control checklist

• Confirm whether AI-generated Shopping descriptions are actually appearing in the relevant accounts, markets and placements.
• Document that DV360 bulk-management and AI-labeling capabilities are separate from Shopping-specific functionality.
• Assign owners and refresh schedules to price, promotion, availability and core descriptive attributes.
• Check feeds against landing pages for contradictions, expired offers, placeholders and variant errors.
• Create an evidence-backed register for environmental, health, performance, certification and comparative claims.
• List approved, conditional and prohibited language, including required qualifiers.
• Scan all potential source content—not only the main product description—for risky wording.
• Classify products by compliance, margin, returns and promotional risk.
• Define human-review thresholds, escalation contacts and authority to pause affected products.
• Launch with a bounded segment and maintain a credible comparison group where possible.
• Track conversion, returns, cancellations, disapprovals, complaints and contribution margin.
• Set stop conditions, preserve rendered evidence and complete a post-incident review after any failure.