If an ad system can choose among discounts, gifts, local coupons, and bundles at the moment a shopper is ready to buy, should a brand upload more offers—or first constrain what the system is allowed to use?
Start with the constraints. Once a promotion can be called by a model according to shopping context, it is no longer just copy. It is an operating permission that consumes margin, inventory, promo codes, and fulfillment capacity. It can also collide with member pricing, affiliate codes, or another live campaign.
Google’s 2026 materials describe Direct Offers as a pilot and testing program. Advertisers can supply promotions, eligible products, guardrails, and an AI Brief; Gemini can then select or construct a relevant deal, including a bundle. Google is expanding the format into AI Mode answers, adding native checkout for eligible UCP merchants, and says AI Max and Performance Max advertisers can create Direct Offers in Google Ads with value limits.
Those statements establish what Google is building. They do not establish universal account availability or guaranteed profitability.
The central risk is not low visibility—it is a miscalled offer
Traditional promotions are commonly scheduled by date, channel, or customer list. Context-selected offers add another variable: the system can decide which value to surface in a particular search or conversation.
That creates four costly failure modes:
- a low-margin SKU enters an expensive gift bundle;
- a clearance incentive keeps serving after inventory tightens;
- an acquisition offer stacks with member pricing, a site coupon, or a creator code;
- the ad describes a valid deal that the landing page or checkout cannot honor.
A “20% off” image is therefore not a complete deliverable. The real unit is a verified path from eligibility through redemption.
The ONCE Offer Guardrail Matrix
Before uploading a Direct Offer, create one row with eight fields. Treat each row as bounded commercial authority, not a piece of promotional copy.
| Field | Required question | Typical owner |
|---|---|---|
| Product eligibility | Which SKUs, variants, and product groups can participate? | Merchandising |
| Margin floor | After incentive, gift, shipping, and platform costs, what contribution is acceptable? | Finance/ecommerce |
| Inventory and fulfillment | At which stock or delivery threshold does the offer stop? | Supply chain |
| Market and identity | Which country, currency, acquisition state, membership tier, or channel qualifies? | Marketing/CRM |
| Stacking | Can it combine with loyalty pricing, coupons, affiliate codes, or free shipping? | Ecommerce operations |
| Content and proof | Which descriptions and claims are allowed, and where does the page substantiate them? | Content/legal |
| Code and term | Is the code single-use or shared? What are the start, end, claim, and redemption limits? | Growth/engineering |
| Owner and rollback | Who approves, monitors, and shuts it down under which trigger? | Program owner |
The most important field is often the stop condition. An offer without an expiry or rollback rule is a continuing right to spend value without a defined way to withdraw it.
Four authorization levels beat one automation switch
Not every incentive should enter the same pool.
Level 1: AI may select
Use for offers with stable margin, healthy inventory, simple eligibility, and a checkout path that has already been tested—for example, a fixed gift that cannot stack.
Level 2: AI may combine pre-approved components
Use for bundles, threshold gifts, or local coupons. The system can choose within an approved component set, but it cannot invent a higher total value.
Level 3: Human release only
Use for deep discounts, constrained SKUs, member-only pricing, or markets with complex tax and fulfillment conditions.
Level 4: Never automate
Use for products constrained by minimum advertised price, distributor agreements, regulation, or exceptional service promises. The product can remain discoverable without joining the dynamic offer pool.
Authorization should change when inventory, margin, return rate, or the campaign calendar changes. Approval is versioned, not permanent.
Pilot one redemption path, not the entire catalog
A low-risk Direct Offers pilot can be intentionally narrow:
- Choose one product group with stable stock, simple returns, and a controlled landing page.
- Allow one base incentive and one approved bundle; block all other stacking.
- Use traceable, limited-quantity, limited-duration codes or a hard value cap.
- Validate the same eligibility and price across ad, landing page, cart, and order system.
- Predefine stop triggers for inventory, claims, contribution margin, code errors, and service incidents.
- Review eligibility accuracy, redemption success, stacking errors, refunds, and make-good costs—not only clicks and conversions.
If the order system cannot tell whether the correct offer was honored, the brand is not ready to expand automated selection.
What agencies and production partners should deliver
Direct Offers turns promotion content into smaller components with tighter operating dependencies. A supplier brief should add four deliverables:
- an offer component library with names, short descriptions, product scope, allowed combinations, and prohibited language;
- proof assets that show bundle contents, gifts, size differences, and material restrictions clearly;
- a handoff map connecting each offer to the landing page, cart behavior, terms page, and code;
- a change log showing who changed value, eligibility, inventory thresholds, or copy and which old versions must stop serving.
This is not a request for more sale graphics. Creative must prove the transaction and remain synchronized with commerce rules. As checkout moves closer to AI interfaces, product facts, offer eligibility, and fulfillment promises need to travel as one versioned package. See When Checkout Leaves the Product Page, Content Must Become Transaction-Ready for the broader content architecture.
Price the work per qualified redemption path
Compare proposals using a denominator tied to operational readiness:
Cost per qualified redemption path = (rule mapping + margin and inventory checks + content production + integration + code operations + testing + monitoring and review) ÷ offer paths that pass eligibility, value, stacking, fulfillment, and rollback checks.
Ten promotional headlines are not ten deployable offers. If only two can be redeemed by the right shopper, for the right SKU, in the right market, with the right margin, the denominator is two.
The decision now
A brand is a reasonable pilot candidate when its product data is stable, contribution logic is explicit, inventory is visible, codes are traceable, and the ecommerce team can roll back quickly. It should wait when prices are synchronized by spreadsheet, distributor inventory is opaque, discounts stack ad hoc, or customer-service credits routinely repair broken fulfillment.
The structural change is not that AI can discount products. It is that a promotion is becoming a machine-callable commercial object. Define eligibility, value, and revocation before letting a model decide which offer to show.
Sources and limits
- Google, “Direct Offers”: placement expansion, bundling, eligible products and guardrails, AI Brief, native UCP checkout, and value limits.
- Google Ads & Commerce Blog, “A new generation of ads for the AI era of Search,” May 20, 2026: pilot expansion, more offer types, and AI Mode placement.
- Google Ads & Commerce Blog, “New tech and tools for retailers to succeed in an agentic shopping era,” January 2026: initial pilot mechanics and AI relevance selection.
- Search Engine Land, “Google Ads tests new Direct Offer asset for AI Mode promotions,” October 7, 2026: observed account and campaign controls, codes, terms, dates, and daily limits. This report does not establish universal availability.
This article separates Google’s confirmed product direction, a trade publication’s observation, and ONCE’s operating framework. It uses no ONCE client performance data and makes no profit or media-performance guarantee.
