A shopper asks an AI assistant for a compact espresso machine that fits under a low cabinet, works with pre-ground coffee, and can arrive before the weekend. Three products appear; one is sponsored. Winning that slot is a media outcome. Explaining why the product actually satisfies those constraints is a different job.
Conversational retail media collapses those two jobs into the same screen. The auction may still run on keywords, filters, and CPC bids, but the ad now sits inside an answer that feels like advice. A brand therefore needs more than eligibility. It needs evidence that can support the answer, a visible paid identity, and a way to withdraw or correct the recommendation when product facts change.
The signal is real, but adoption is not universal
On October 9, 2026, THG Commerce said it had integrated Criteo Sponsored Products into its AI Shopping Assistant built with Google Cloud's Gemini Enterprise. The documented flow converts a natural-language request into keywords and filters, sends those signals to the existing Sponsored Products engine, runs an auction, and places a clearly labelled sponsored result beside organic recommendations. Criteo's help center says active campaigns can become eligible through existing keyword targeting and that performance can be isolated with an AI Assistant page-type dimension. It also says advertisers cannot configure a line item to serve exclusively on that surface. THG's announcement and Criteo's implementation guide establish one live deployment and one product mechanism; they do not establish industry-wide adoption.
The announcement also reports large early performance multiples. Because it does not provide enough methodology for an independent buying decision, this article does not use those numbers as evidence of expected return.
Amazon shows a different maturity split. Sponsored Products and Sponsored Brands prompts moved to U.S. general availability on March 25, 2026, with prompt-level reporting and pause controls. Amazon's later Branded Conversations product will let brands contribute deeper product expertise, remains clearly marked as Sponsored, and is scheduled to launch in the United States in January 2027. Those are not the same availability state. The prompt announcement describes what advertisers can use now in the U.S.; the Branded Conversations announcement describes a future launch.
Google's May 20 announcement adds another design signal. It is testing Conversational Discovery ads and Highlighted Answers with an independent AI explainer shown alongside advertiser creative, while retaining a Sponsored label. AI-powered Shopping ads can generate a tailored explanation of why a product may fit a query. These remain tests or staged launches, not proof that every advertiser can buy the same unit today. Google's announcement matters because it makes the separation between paid creative and independent explanation explicit.
The common direction is structural: retail media is moving from buying a position on a page toward buying eligibility inside a product answer. The rules, controls, and availability still differ by retailer and market.
Keep three ledgers separate
The most important governance rule is simple: payment can affect ad eligibility, but it cannot manufacture a product fact.
The organic-relevance ledger
Would the product satisfy the shopper's constraints without a media bid? This ledger holds the actual fit: size, use case, price, fulfillment, compatibility, and disqualifying conditions. The assistant or retailer controls organic recommendation logic; the brand should not present that logic as a paid guarantee.
The auction-eligibility ledger
Which campaign, keyword, bid, retailer, and page type made the SKU eligible? Criteo documents a standard ad request built from the conversation's extracted keywords and filters. The auction decides which sponsored product can appear. It does not certify the claim made around it.
The factual-answer ledger
Which approved source supports every assertion in the answer? A claim such as “safe for this material,” “compatible with this connector,” or “delivers by Friday” needs a SKU, market, source owner, refresh rule, and expiry condition. If those fields do not exist, the bid has won a placement that the brand cannot responsibly explain.
Our earlier article on agent-ready product data addressed identity, variants, availability, and checkout. Conversational retail media introduces an additional governance layer: who entered the answer because of payment, what paid influence may change, and how that influence is disclosed and measured.
The ONCE Sponsored Answer Contract
Before an agency supplies a keyword list, require one contract for each high-value question–product combination. A combination is not ready until five fields are complete.
| Contract field | What the buyer must specify |
|---|---|
| Question boundary | The shopper task, valid language variants, and the conditions that disqualify the product |
| Product eligibility | Exact SKU, market, source of price and stock, fulfillment limits, and promotion expiry |
| Evidence packet | Approved claims, responsible source, required image or video proof, freshness date, and prohibited inferences |
| Paid identity | Where Sponsored appears, how brand-supplied expertise is distinguished from independent explanation, and the destination owner |
| Measurement and correction | Question family, product shown, observable actions, pause authority, correction workflow, and withdrawal trigger |
This contract is not paperwork for its own sake. It gives media, commerce, creative, legal, and fulfillment teams one acceptance object. Keywords can buy consideration. The contract determines whether the resulting answer can remain truthful.
Creative production becomes evidence production
Conversational ads do not eliminate photography or video. They reduce the value of generic ambience and increase the value of shots that resolve a specific buying question.
For an appliance brand, reusable evidence might show cabinet clearance, port placement, setup steps, noise context, or the difference between adjacent SKUs. For skincare, it might show texture, order of use, quantity, packaging size, and what the visual cannot prove about medical efficacy. The goal is not to shoot a bespoke commercial for every prompt. It is to build a small, governed library of proof that can be attached to recurring question families.
Each clip or image should carry the SKU, market, supported claim, source, usage period, and prohibited inference. When a design, formula, offer, or policy changes, the brand can withdraw the affected answer rather than pausing the entire campaign. A vendor that delivers only “one master plus three crops” has supplied files, not conversational evidence.
A 30-day pilot that isolates the new risk
Days 1–5: Select questions, not novelty
Choose one market, one category, and a small set of valuable SKUs. Use customer service logs, onsite search, reviews, and return reasons to identify roughly twenty question families that genuinely block purchase. Exclude queries that require diagnosis, individual advice, or facts the business cannot refresh reliably.
Days 6–10: Complete the contract and close proof gaps
Reuse existing product pages, feeds, photographs, and videos when they already prove the answer. Commission new evidence only for missing dimensions, compatibility, process, or usage boundaries. A claim without an owner or source does not go live.
Days 11–20: Cap spend and isolate the assistant surface
Set a bid ceiling and activate only products with completed contracts. Confirm that the retailer or platform exposes the assistant, prompt, or equivalent page type as a separate reporting dimension. Record the available steps—impression, click, continued conversation, PDP visit, add-to-cart, or order—and document the steps that cannot be observed.
Days 21–30: Classify the outcome
Separate three results: paid visibility without a useful answer; a valid answer without sufficient commercial value; and an answer that passes truth, disclosure, and contribution thresholds. The first returns to evidence design, the second to media and offer strategy, and the third qualifies for expansion. Out-of-stock, expired, misleading, or undisclosed combinations leave the denominator immediately.
Measure the full cost of a validated answer
CPC captures media cost. It does not capture what the brand spends to keep the answer accurate. Use this denominator instead:
Full cost per validated sponsored answer = (product-data cleanup + claim substantiation + evidence production + feed or integration work + media clicks + disclosure review + monitoring and correction) ÷ question–product combinations that remain truthful, available, clearly disclosed, and above the brand's contribution threshold for the full observation period.
This is not an industry benchmark and it does not promise ROI. It exposes two costs that media dashboards hide: maintaining the facts behind the answer and removing combinations that fail after launch. A low-priced vendor that requires constant brand-side repair is not necessarily the cheaper option.
When conventional sponsored search is the better buy
Stay with a conventional search placement or product page when any of these conditions apply:
- Price, stock, or delivery changes faster than the business can refresh and withdraw answers.
- The central claim involves health, safety, sustainability, or performance, but current substantiation is missing.
- The experience does not clearly distinguish organic recommendation, sponsored placement, and brand-supplied information.
- Assistant or prompt performance cannot be isolated from blended campaign reporting.
- The purchase is simple enough that price and availability already answer the shopper's need.
- No named owner can pause an answer when the product, policy, or market changes.
The U.S. FTC's native advertising guide says disclosures needed to prevent deception must be clear, prominent, and as close as possible to the native ad. That principle becomes more—not less—important when the ad is embedded in a conversational answer. A Sponsored label is the floor. The brand should still know which product, statement, and source the label covers.
McKinsey's August 5 analysis describes commerce media as moving from page placement toward decision infrastructure and argues that networks need explicit rules for sponsored influence. That is a useful market framing, not proof of platform adoption or advertiser return. The operational decision remains narrower: do not scale a conversational placement until the business can separate relevance, payment, and truth at the question–product level.
Conversational retail media may still be bought with keywords and CPC bids. Its scarce asset is a sponsored answer that does not need to be corrected after the shopper acts on it. If the brand lacks stable product facts, proof assets, visible disclosure, or pause ownership, keep spend in more observable formats. If all five contract fields are complete, run a narrow pilot and let validated answers—not novelty—earn the next budget increment.
