The facts shoppers never say out loud: build a deal-breaker sheet for every SKU

Every category has questions shoppers do not think to ask until it is too late. Is it scented? Does it have sesame? Will it fit the dispenser I already own? Is this the 12-count or the 24? A good store clerk asks. A good product page answers before anyone asks. An AI shopping assistant, it turns out, often does neither.

That is the most useful finding in Visa's Agentic Commerce Benchmark, which we charted in full in today's data post. Visa's researchers say capable models "seldom misinterpret a fact once it's in their context. Their failures occur earlier." Two of the five failure patterns the benchmark targets are, in plain words, about product data:

• Satisfying every stated requirement while never asking about unstated ones.
• Treating a value missing from a listing as acceptable.

Neither is something a brand can fix inside the model. Both are things a brand can make much less likely by writing its deal-breakers down, completely, in every place an assistant reads. This playbook is how we would do that.

Why the unsaid needs decide the outcome

Visa's one public scenario shows the scale of the problem. A shopper named Sam asks for help replacing a 65-inch TV and says it must be returnable. Four more needs exist, but Sam only reveals them when asked: a budget of up to $1,200 all in, Dolby Vision for the spouse, HDR10+ for the kid's phone videos and native 120 Hz for the console.

Show the data

After this requirement	TVs left
Start	265
65-inch screen	125
30-day returns	65
Budget up to $1,200	44
Dolby Vision	35
HDR10+	30
Native 120 Hz	20

Source: Visa Payments Frontier, Agentic Commerce Benchmark, Episode 1. The first two requirements were stated in Sam's request; the last four came out only when the assistant asked.

The stated needs got the field from 265 to 65. The unsaid ones got it from 65 to 20. An assistant that stops at what it was told picks from a pool where most of the options are wrong. Across the whole benchmark, the five models that uncovered the most hidden buying factors passed 70.2% of their runs; the five that uncovered the fewest passed 44.4%, according to Visa's release.

Packaged goods have the same shape. A shopper asks for "a good granola for my kid's lunchbox." Unsaid: the school is nut-free. A shopper asks for "laundry pods that work in cold water." Unsaid: someone in the house has sensitive skin, and they want it unscented. The assistant may never ask. If your page leaves the answer blank, the assistant may assume the best and recommend you to someone you will disappoint.

Step 1: List the reasons a shopper would reject the product

Start with one SKU, ideally your best seller. Put the brand, product and regulatory people in a room for 45 minutes and write down every reason a real shopper would say no to it. Do not filter yet. Sources for the list:

• Customer service tickets and returns reasons from the last year.
• One-star and two-star reviews, on your site and at retailers.
• The questions in retailer Q&A sections.
• The filters retailers already offer for your category, which are a record of what shoppers sort by.
• For food, the nine major allergens identified by US law, which the FDA lists as milk, eggs, fish, crustacean shellfish, tree nuts, peanuts, wheat, soybeans and sesame.

Then cut the list to the 8 to 12 that actually decide purchases. A long list nobody maintains is worse than a short one that is always right.

Step 2: Give every deal-breaker a definite answer

This is the step that matters, because it is the one that addresses Visa's second failure pattern. Every line on the sheet gets a yes, a no or a number. Not a blank, not "see packaging," not a marketing phrase.

Deal-breaker	Answer	Where it is proven
Contains peanuts or tree nuts	No; made in a facility that also processes tree nuts	Label, allergen statement
Contains sesame	No	Label
Gluten-free	Yes, certified	Certifier listing
Added sugar per serving	6 g	Nutrition facts
Serving size and servings per bag	45 g; 8 servings	Nutrition facts
Bag size	12 oz (340 g)	Label
Count per case or multipack	Single bag; also sold as 4-pack	Pack spec
Suitable for vegan diets	Yes; contains no honey or dairy	Ingredient list
Who it is for	School lunches, breakfast, snacking	Brand copy
Who it is not for	Shoppers avoiding oats	Ingredient list

Two rules make the sheet useful. First, negatives are written as negatives. "Not for septic systems" or "contains coconut" saves a wrong purchase. Second, each answer points to where it is proven, so the sheet can be checked against the label every time the formula or pack changes. Never put a claim on the sheet that is not already approved for the label or your substantiation file.

Step 3: Put it on the product page as text and a table

An assistant can only use the sheet if it can read it. Put the deal-breakers on the product page as a short plain-text list or an HTML table, close to the top, in the same words as the sheet. Not in an image, not in a tab that loads with JavaScript after the page renders, and not only on the back-of-pack photo. Tip 22 covers why HTML tables work well, and tip 21 covers how to show certifications with a link to the certifier.

Add a short "Who it's for, who it's not for" line. It reads naturally to people and it gives an assistant a direct answer to the question it should have asked. Tip 20 has more on writing for specific needs.

Step 4: Copy it into every feed, field by field

Most assistants do not read only your page. They read feeds and retailer listings, and they will believe whichever one they see. Map each line of the sheet to a field.

Google Merchant Center. Google's product data specification includes product_detail for technical specifications or additional details, product_highlight for the most relevant highlights, multipack for custom groups of identical products sold together, and unit_pricing_measure for the measure of the product, such as 1.5 kg. Allergen and certification lines fit well in product_detail. The newer conversational attributes, which we filled in for snacks and shampoos in our Merchant Center post, are a natural home for "who it is for."

ChatGPT shopping. OpenAI's product feed specification has fields for description, material, dimensions and more, and it notes plainly that "eligibility does not guarantee display." The point is not the field names; it is that the same facts, in the same words, appear there.

Retailer listings. Amazon, Walmart, Target and grocery retailers each have their own attribute forms. Fill in every attribute that matches a line on your sheet. A blank allergen field at one retailer is the same problem as a blank on your own page.

Your Shopify catalog or other platform data. If you sell direct, check what your platform passes to agents. We walked through Shopify's version in tip 45.

Step 5: Test with the question the shopper would not ask

Once the sheet is live, test whether it works. Ask two or three assistants a shopping question that implies a deal-breaker without stating it, then follow up with the deal-breaker itself:

1. "What's a good granola for a kid's school lunch?"
2. Then: "The school is nut-free. Is that one okay?"

Watch two things. Does the assistant ask about allergies before recommending? (Visa's data says it often will not.) And when you add the requirement, does its answer about your product match your sheet? If it says your product contains nuts when it does not, or says it is fine when it is not, find the page or listing it read and fix it. Tip 25 covers what to do when an assistant gets a product fact wrong.

Run the same questions more than once. In Visa's benchmark, all three runs of a scenario passed only 34% of the time, so a single answer is not a result. Teams that need to do this across many SKUs and assistants sometimes use SearchPanel, a tool from XENA Intelligence that shows the actual answers per product; for one hero SKU, a spreadsheet and a calendar reminder is enough.

Step 6: Keep it current

The sheet is only useful while it is true. Tie it to the change process you already have: any formula change, pack change, new certification or new facility triggers an update to the sheet, and the sheet triggers updates to the page, the feeds and the retailer listings in the same week. Put an owner's name on it. Re-check the whole sheet against the current label every quarter.

Where this fits

Visa's benchmark is about TVs and electronics, not snacks, and it is a simulation in which one AI plays the shopper. We would not lean on its exact percentages for packaged goods. The pattern, though, matches everything else we have seen this fall, from General Machines' simulated store visits in our CartTrace post to the brand-fidelity study we covered on November 9. Assistants handle the facts they are given well. They are poor at noticing what they were not given.

A deal-breaker sheet is the cheapest protection a brand has against that. It takes an afternoon per SKU to write, it uses information you already have on the label, and it makes the wrong recommendation harder to make, whether the reader is a shopper, a retailer's assistant or an agent working on someone's behalf.