What It Takes to Get Recommended by Walmart Sparky: Over 70% of Its Citations Sit Outside the Retailer Listing
Walmart disclosed that Sparky users are up 70% year over year and spend 40% more per order. Here is Azoma's measured breakdown of what Sparky cites, and the order in which e-commerce operators should fix their AI visibility.
Key Takeaways
- On its August 20, 2026 earnings call, Walmart disclosed that the number of customers using its AI shopping assistant Sparky is up 70% year over year, and that Sparky users spend 40% more per order than non-users.
- According to Azoma's measurement, Sparky assembles its recommendations from earned media (36%), brand.com (30%), retailer content (27%), UGC (5%) and affiliate sites (2%). The retailer product listing accounts for roughly a quarter of the citation surface.
- Polishing the listing alone leaves most of that surface untouched. That said, reading the 40% figure as causal is premature, and paid placements on the Sparky surface have already started to open up.
Walmart Put a Dollar Figure on Its AI Assistant
Azoma reveals how brands can improve Walmart Sparky product visibility, boost AI shopping agent discovery, and drive retail media revenue.
www.martechcube.comThe claim that AI assistants are reshaping how people shop has been repeated for two years. What changed is that a number is now attached to it.
On Walmart's second-quarter earnings call on August 20, 2026, CEO John Furner said that the number of customers using Sparky is up 70% from last year, and that customers who use Sparky for shopping spend 40% more per order than those who do not. Global e-commerce net sales grew 23% in the same quarter, and global advertising grew 38% (earnings call transcript).
The growth rates matter less than the granularity of the disclosure. Until now, retailers described their AI assistants in qualitative terms: usage is growing, engagement is strong. Framing it as order value, a metric any operator can hold against their own numbers, is a meaningful shift.
And for a long time this surface has had no explicit auction the way paid search does. Whether you appear is decided by the state of your information rather than the size of your budget, which raises the obvious question of what exactly needs to be in order.
What Sparky Actually Reads Before It Recommends
Azoma, which positions itself in agentic commerce optimisation, answered that question with measurement. The breakdown below is the source mix the company reports from an analysis of millions of shopping agent citations.
| Citation source | Share | Nature of the work | Time to take effect |
|---|---|---|---|
| Earned media (third-party articles, reviews, comparisons) | 36% | PR, press outreach, review acquisition | Long (months) |
| Brand.com | 30% | Structured data, machine-readable specs, comparable descriptions | Medium (crawl and re-index lag) |
| Retailer content (product listings) | 27% | Listing hygiene, attribute completeness | Short (fully in your control) |
| UGC (reviews, posts) | 5% | Review volume and freshness | Medium |
| Affiliate sites | 2% | Quality of partner content | Short |
The reading is straightforward. The retailer product listing accounts for only 27%, while earned media and brand.com together make up 66% of the citation surface. Building out a strong Walmart listing remains necessary, but it addresses a quarter of what the recommendation is built from.
Azoma founder and CEO Max Sinclair put it this way.
Whether Sparky recommends your product comes down to how you show up across the sources it reads, and most of those aren't your Walmart listing.
The company operates out of London and Toronto, and in March 2026 published the Agentic Merchant Protocol (AMP), a standard for distributing product catalogue data. Where OpenAI's ACP and Google's UCP connect product data to buyer-side discovery, AMP is framed as a merchant-side layer that lets a brand define its product information once and push it out across agents. L'Oréal, Unilever, Mars, Beiersdorf and Reckitt were named as launch adopters.
The numbers deserve care, though. This breakdown is first-party analysis from a vendor selling AI visibility tooling. The category composition, the measurement window, and how much of the sample comes from agents other than Sparky are not disclosed. It is useful for orientation, not for lifting directly into your own targets.
Reading the 40% as Causation Will Cost You
Walmart's 40% figure compares customers who use Sparky against customers who do not. Treating it as evidence that driving assistant usage will lift order value by 40% will badly distort your expected return.
The same issue has already been flagged around Amazon's Rufus. One retail media analysis reported that Rufus-assisted sessions made up roughly 40% of all sessions during Black Friday yet drove about 66% of purchases, converting at 3.5 times the rate of non-Rufus sessions. The same analysis attached explicit caveats. Shoppers who reach for Rufus are plausibly the higher-intent shoppers to begin with. Counting any brief interaction as a "Rufus session" can inflate the gap, and the work relies on modelled panel data. The conclusion was that this is a strong directional signal rather than causal proof (Retail Media Breakfast Club).
Walmart's 40% carries the same selection-bias risk. A world in which deliberate shoppers try the assistant and deliberate shoppers also buy in bigger baskets produces exactly this number.
There is a second premise worth questioning: that Sparky recommendations carry no paid placement. Walmart quietly tested advertising inside Sparky in late 2025 and has been expanding it during 2026 in the form of sponsored prompts, which appear when shoppers ask the assistant for product recommendations. Alongside it, an advertiser-facing assistant called Marty is rolling out to brands buying search ads (eMarketer).
There is no confirmation that the recommendation itself has become purchasable. But the surface is clearly being treated as a monetisation target. The idea that you can win on information quality because the placement cannot be bought is best read as a current window, not a permanent structure.
Sparky No Longer Lives Only Inside the App
One more shift is easy to miss.
Walmart announced its OpenAI partnership in October 2025 and piloted direct purchasing through Instant Checkout inside ChatGPT. In March 2026, after conversion rates came in well below its own channels, it moved away from that implementation and instead began embedding Sparky itself into ChatGPT and Gemini. Shoppers authenticate with a Walmart account, carts sync, and checkout completes on Walmart's side (Retail Dive).
The citation-mix question is therefore no longer confined to Walmart's app. When a conversation that starts in a general assistant flows into a Sparky recommendation, the surfaces you need in order stretch well beyond the retailer's own walls.
The Order in Which Operators Should Work
Turning all of this into something an e-commerce team can act on: Sparky is not available in Japan, but the structural point about citation sources applies directly to shopping answers in ChatGPT and Gemini, to Amazon Rufus, and to AI search inside marketplaces.
The governing principle is to start with whatever takes longest to take effect. Earned coverage has to be placed, then indexed, then picked up by the models, which runs to months. Structuring your own site comes next. Listing hygiene, being entirely within your control, can safely come last. Most teams work this in reverse and stall after polishing the 27% that finishes fastest.
On your own site, the work that pays is removing inconsistencies in product names and model numbers, holding specifications as attributes rather than prose, and putting comparison language into descriptions: who the product suits, and how it differs from the alternatives. Assistants decompose a question into several queries before searching, so a listing without use-case vocabulary never enters the candidate set in the first place.
Measurement also has to be designed differently from conventional SEO. Because answers are generated per query, the same product can appear for one phrasing and vanish for a near-identical one. Manual spot checks cannot capture that state. What you need is a fixed set of representative prompts per category, monitored on a schedule for how often you and your competitors surface. Deciding up front how many prompts and how often is the practical starting point.
Finally, resist using Azoma's percentages as your own target allocation. The weight of earned media and the influence of UGC vary considerably by category. Collect the URLs actually cited for your own representative prompts, and derive your own citation mix from there.
Conclusion
What Walmart's disclosure establishes is that assistant-led shopping has reached a scale measurable in order value. What it does not establish is causation, and the advertising surface is opening up in parallel. There is not yet enough evidence for a confident verdict.
Even so, the shape of the problem, with most of the citation surface sitting outside the product listing, looks durable for now. The things to watch next are how close Walmart brings sponsored prompts to the recommendation itself, and how much transaction volume Sparky ends up carrying inside ChatGPT and Gemini. The disclosure covering the holiday quarter will be the first real checkpoint.
