What does a shopper type when the thing they want is not a word? As of November 11, 2026, Walmart's answer is a photograph. Its assistant Sparky can read an uploaded image, turn a handwritten list into a cart, light up a shelf label inside the store and swap ingredients for healthier versions; Walmart says Sparky users spend 40% more per order and that visual search lifted conversion 57% against text search. The interesting part for a brand is not the assistant but what it reads before it recommends anything.
Six features and one shared input
The mechanics are mundane and that is the point. A shopper uploads a picture or a screen grab, Sparky identifies the need, recommends products and can arrange delivery in as little as 30 minutes. A handwritten list becomes a cart, with substitutions offered when an item is out of stock. The Shop to Light tool, folded into Sparky, pins a product's location on a store map and lets the shopper trigger the electronic shelf label on arrival (OKRVV).
Why the photo outranks the chat box
A chat box asks the shopper to describe what they want, which most people do badly. A photo skips the description and goes straight to matching. That shift changes the ranking problem: text search matched keywords in a title, while image matching compares attributes such as colour, material, dimensions and compatibility. A listing built for keyword coverage can be invisible to an assistant that is trying to match a picture of a navy wool throw against a catalogue.
The review summary becomes the shelf
When Sparky launched in June 2025, one of its stated jobs was synthesising reviews. That makes the assistant the first reader of a review corpus, and a summary is only as useful as the reviews it can parse. PYMNTS and OKRVV describe the same product logic: the assistant answers questions, builds carts and, when something is unavailable, offers an alternative (PYMNTS).
| Signal | Where it surfaces | What a brand controls |
|---|---|---|
| Review summary | Assistant answer and comparison | Structured attributes and verified purchase reviews |
| Rating dispersion | Order in a shortlist | Consistency across variants and sizes |
| Stock status | The substitution moment | Which SKU is offered when the first choice fails |
Walmart's own worked example is more concrete than the feature list. Sparky told a shopper that a previously purchased trash bag brand was out of stock at her usual store, recommended alternatives and produced a nine-item cart of about $64. Purchase history set the default, store-level inventory decided availability, and where neither existed the assistant chose. In a cart built from a handwritten list, the shopper never types a brand name, so the selection happens inside the retailer's system.
The third row is the one brands underrate. Substitution is not a customer service event, it is an assortment decision made in a fraction of a second by a rule nobody in the brand team wrote. If the substitute is chosen by availability alone, the sale can move to a competitor inside the same cart, and the lost order never appears in a conversion report because the shopper did convert.
Attributes beat adjectives
Reviews that say the jacket runs a size small, the seal failed after two weeks, or the shade is darker than the photo give a summariser something to extract. Reviews that say the product is amazing give it nothing. The practical consequence is that review generation should ask for specifics, and that review analysis becomes a listing task rather than a marketing report, and listing copy should repeat the specifics a shopper would otherwise have to infer from a photograph.
The pricing worry is a data worry
Walmart's electronic shelf labels have drawn speculation about dynamic pricing, and chief executive John Furner answered it directly in an open letter: engaging with Sparky is an invitation to serve the customer better, not to use personal information to set a personalised price, and the company has never used store relationships to charge more. The more ordinary risk is that shelf price, app price and the price an assistant quotes drift apart, which reads as variable pricing even when nothing moved (AI Best Practices).
What to fix before the next peak
- Complete attributes on the variants that carry volume. Fill material, dimensions, compatibility and care fields for the top twenty SKUs before touching the long tail, because an incomplete variant cannot be matched from an image at all.
- Rewrite review requests around failure modes. Ask buyers what went wrong and at what point, so the corpus contains sentences a summary can extract instead of sentiment it cannot use.
- Check price and stock consistency across shelf, app and assistant. Sample twenty stores in one week and compare the three sources, then treat every mismatch as an availability defect rather than a display bug.
None of this is new work. It is the same catalogue and review discipline that search rewarded, applied to a reader that compares instead of matching strings. The brands that move first will not be the ones with the best assistant. They will be the ones whose product data survives being read by a machine that has never seen the packaging.
FAQ
Does an AI shopping assistant really change how a listing should be written?
It changes what the listing is read for. A keyword-ranked search rewarded phrases that matched a query. An assistant that identifies a product from an image and then compares options rewards structured attributes, because it has to match colour, material, size or compatibility before it can put anything in a cart.
Why do reviews matter more once an assistant summarises them?
Because the summary becomes the first thing a shopper reads and often the only thing. Reviews that describe a specific failure, such as a size that runs small or a seal that failed after two weeks, survive summarisation. Reviews that only praise the product contribute nothing a summary can extract.
Is the dynamic pricing concern a pricing issue?
Mostly it is a consistency issue. Electronic shelf labels and app prices can drift apart, and an assistant that quotes whichever value it can read creates the impression of variable pricing even when the shelf price has not moved.
What should a brand fix first?
Attribute completeness on the top-selling variants, then review structure, then stock accuracy. Attributes decide whether a product can be matched at all, reviews decide whether it is chosen, and stock decides whether it survives the substitution step.
