A Price for One, a Promise to Everyone

At 9:13 on a Tuesday morning, two screenshots arrive in the same group chat.

The first shows a $180 skin-care set at $180.

The second shows the same set, from the same brand, with a Sponsored deal for 20% off. The difference is $36—large enough to fund lunch, small enough to feel almost petty to challenge, and perfectly calibrated to make the first shopper type the sentence every pricing team eventually meets:

Wait. Why do you get that?

This is a hypothetical scene. But it is no longer a hypothetical mechanism.

Google introduced Direct Offers as an AI Mode pilot in January 2026. In that first description, retailers set up relevant offers in their campaign settings and Google used AI to decide when an offer was relevant to display. The pilot began with discounts, including Google’s own example of a special 20% off. In May, Google announced a more generative extension: brands will be able to upload promotions such as discounts, giveaways, and local coupons; supply eligible products and guardrails; and let Gemini construct a customized deal for a particular search, such as a product bundle. Google also says AI Max and Performance Max advertisers can create Direct Offers in the Google Ads interface and set value limits. The surface is still described as a pilot, with parts of the expansion “coming soon.” (Google’s January introduction; May announcement; Direct Offers product page)

That distinction matters. A system that chooses when to display a finished coupon creates one kind of risk. A system that constructs a shopper-facing deal from products, promotion inputs, and limits creates another.

The advertiser has not necessarily written the exact promise the shopper sees. It has approved the grammar from which that promise may be made.

The reassuring answer that does not answer the question

An experienced operator knows what to do when a screenshot like this appears. Confirm the SKU. Check the geography, audience, dates, stackability, inventory, margin floor, and checkout path. Make sure the offer was actually served, not edited. Find out whether the second shopper can still redeem it. Brief support.

Suppose all of that works.

The discounted shopper is eligible. The deal sits inside the approved value limit. Checkout produces $144. The order can be fulfilled. The margin is acceptable. The offer has an ID, a timestamp, and a clean audit trail. No model escaped its guardrails. No coupon leaked. No employee has to improvise a refund.

The group chat is still not resolved.

The first shopper did not ask whether the second shopper’s deal was technically valid. She asked what her own $180 price now means.

That is the consequence an operational-readiness model can miss. Readiness asks whether the company can recognize and honor the promise delivered to the recipient. The screenshot introduces a witness. The witness is not trying to redeem the offer. She is using it to interpret the company.

She now knows the brand was willing to sell the same set for $144 to somebody who, from her vantage point, looks exactly like her. Perhaps the system had a sound reason. Perhaps her friend’s query showed stronger purchase intent. Perhaps the offer applied to a particular combination, location, or moment. Perhaps the two shoppers were not similarly situated in some important way.

But none of those reasons travel automatically with the screenshot.

The portable fragment contains the brand, the product, and the better price. It tends to shed the query, the eligibility logic, the expiration state, and the obscure little distinctions that made the offer safe inside the ad system. The part that survives is precisely the part that looks most like a public promise.

A private offer can create a public price

“Discount leakage” is the familiar name for this problem, but it is too small.

Leakage assumes that the valuable object escaping is the discount. The response is therefore to prevent sharing, constrain redemption, or accept a controlled amount of cannibalization.

The screenshot can do commercial work even when the offer itself cannot travel.

It can reset the other shopper’s reference price. It can turn a full-price purchase into evidence that she was the person who failed to ask the machine correctly. It can make a loyal customer feel less recognized than an anonymous high-intent query. It can teach a group of friends that $180 is the price for people who do not receive the interesting version of the brand.

No one has to redeem the 20% offer for any of that to happen.

This is why traceability, though necessary, is not enough. A company can trace a promise perfectly and still be unable to explain it without sounding arbitrary. The support agent can locate the event, confirm the guardrails, and say, with complete accuracy, “That offer was only available to eligible shoppers in AI Mode.” To the person who paid $180, this is not an explanation. It is a restatement of the mystery in company language.

Personalization happens one shopper at a time. Reputation happens when shoppers compare notes.

The second audience

Every constructed deal has at least two audiences.

The first is the person for whom the system constructs or selects it. For that shopper, the governing questions are concrete: Is the offer truthful? Is it relevant? Can it be redeemed as shown? Will the cart, store, and support team honor it?

The second is everyone who learns that the offer existed. They encounter it without the private context that made it relevant. Their question is different: What rule about this company does this difference reveal?

Guardrails are usually designed for the first audience. They control eligible products, promotion types, values, audiences, and other inputs close to the transaction. The second audience requires something less machine-shaped: a reason for the difference that the company is willing to say aloud.

Consider three possible offer families.

A local coupon tied to excess stock at one store can survive comparison: “That location had inventory to clear.” A bundle discount can survive: “Buy the cleanser and moisturizer together and the set costs less.” A lower price shown only to the shopper an opaque model predicts is most persuadable may be profitable, valid, and impossible to explain to the full-price friend without revealing the uncomfortable policy underneath.

Those are not identical promotion strategies with different implementation difficulty. They make different claims about what the brand believes a price is for.

The mistake is to assume that any deal inside the approved financial range is therefore inside the approved brand range. Money has a ceiling in the interface. Meaning does not.

Authorizing explanations, not just outputs

This does not mean a marketer must approve every generated sentence one at a time. That would erase much of the point of the system. Nor does it mean every shopper must receive the same offer. Brands have always used geography, timing, inventory, bundles, loyalty, and acquisition offers to vary terms.

It means the authorized range needs a shared reason, not merely shared limits.

Before enabling a family of constructed deals, the decisive work is to complete the sentence the ineligible shopper will force the company to finish:

Some people may receive a different offer because…

If the honest ending is grounded in a commercial condition the brand can defend—what is being bought, where it is available, when the promotion runs, or what reciprocal commitment the shopper makes—the range may remain coherent even when individual outputs differ.

If the honest ending is “because the system estimated we needed to give them more,” the company has discovered its real policy. It may still choose that policy. But it should choose it in daylight, accepting that the conversion gained from one shopper can alter the trust of another.

That is the cost the most permissive guardrail cannot price for you.

The temptation is to say that Gemini made the offer. Yet the advertiser supplied the promotions, eligible products, guardrails, and value limits. Google describes Gemini as constructing the customized deal, not adopting the commercial obligation. And where Google’s UCP checkout is involved, Google says the retailer remains the merchant of record. (Google’s UCP update)

Automation can assemble the sentence. It cannot volunteer to be the company that meant it.

Tuesday morning, again

Back in the group chat, the discounted shopper may buy the set for $144. The full-price shopper may receive a price match, wait for a sale, or move on. Those outcomes matter, but they are not the whole event.

The brand has learned something about the object it authorized.

It thought it had approved a range of private incentives. What it actually approved was a range of possible explanations for treating comparable shoppers differently. The screenshot merely made one of those explanations due.

The useful standard is therefore not “Would we honor every offer Gemini can construct?” Of course the answer must be yes.

It is: “Would we still recognize the reason for every offer when the shopper it was not meant for is the one holding the screenshot?”

Gemini can construct a deal for one person. The company has to construct a reason that survives the room.