How AI Is Changing Shopping, and What That Changes for Retailers

What company disclosures through September 2026 show about discovery, checkout, customer ownership and the services marketplaces still perform.

AI shopping raises a straightforward possibility: customers could increasingly ask an assistant to find and compare products, choose between sellers, and complete a purchase. That could change where shopping begins, how customers make decisions and which businesses they interact with along the way. For retailers and marketplaces, the question is how much of their existing role an assistant takes on—and which services customers still need them to provide.

Company disclosures offer an early view of how this is developing. Some businesses are improving recommendations inside their existing stores. Others are building their own assistants or accepting purchases from outside AI platforms. These arrangements can look similar to the shopper while having different consequences for the business. A retailer’s own assistant selling through WhatsApp, for example, has a different role from an independent assistant choosing among competing retailers.

The findings below address two questions in turn: how AI is changing the shopping experience, and how those changes affect retailers and marketplaces. The most useful evidence comes from differences in actual results and commercial choices, alongside management’s expectations about what comes next.

What is changing in how people shop

Better matching can help before conversational features do

Expedia’s August earnings call separated two sets of AI products. Ranking, recommendations and personalization were already improving conversion. Natural-language search, Property Expert and AI Compare were not yet doing so, although they gave Expedia more than 60% more information about traveler intent.¹

The two sets ask different things of the traveler. Ranking, recommendations and personalization can improve the results a traveler already receives without requiring them to adopt a conversational interface; the conversational features require that adoption first, and then still have to resolve a decision rather than only inform one. That difference in what each product requires is a possible explanation for the gap in measured conversion, not a demonstrated cause — Expedia did not attribute the difference. What the company reported is better matching on one side and richer intent information on the other, with only the first showing up in conversion so far.

Rakuten supplies a measured version of the first effect. In what management described as an A/B test, LLM query understanding and ranking raised orders by 0.52% and merchandise sales by 0.87%.² These are small increments applied to an existing shopping process, and the company disclosed no sample, duration or randomization detail, so the figures support this intervention rather than an industry estimate of what AI is worth.

Larger headline numbers need a different reading again. Walmart said average order value rises 40% when customers engage with Sparky.³ Management did not specify the comparison group or disclose a controlled test. The figure does not isolate what Sparky caused, because shoppers who choose to use an assistant may already buy differently. The Walmart and Rakuten figures are not comparable: one is an observed association with assistant use, the other tests a single ranking change.

The useful product question is narrower than whether an assistant attracts users. Does it improve the match, shorten the decision, or complete the purchase? Expedia’s own results show those can develop at different speeds inside one company.

A retailer’s assistant on someone else’s surface

Magazine Luiza shows shopping moving to another interface while the retailer continues to supply the shopping service. Its Lu assistant sells through WhatsApp. By the August call, Lu had reached BRL 100m of cumulative GMV since its December 2025 launch; management said the milestone had been disclosed in July. The company separately reported more than 18m conversations; the two totals were not disclosed as a linked conversion measure.⁴

The GMV figure is a transaction measure rather than an engagement measure, which makes it unusual among owned-assistant disclosures. It is cumulative, and Magazine Luiza did not isolate incremental sales, WhatsApp channel costs or profit. The structural point is the arrangement: the retailer supplies the agent while another company supplies the interface. An independent assistant comparing offers across retailers takes on a different product-selection role, choosing among sellers rather than within one seller’s catalog. Both arrangements move activity off the retailer’s website; the owned agent still leaves the retailer supplying the service the customer uses.

Reach for owned assistants is already large at some incumbents. Amazon disclosed 350m Alexa for Shopping users over twelve months; Alibaba disclosed 250m first AI-shopping experiences; Allegro disclosed 500,000 monthly assistant users in September.⁵, ⁶ These count different activities over different periods and cannot be set against referral shares. What they show is that these particular incumbents can put AI shopping tools in front of substantial existing audiences. Usage at that scale is not evidence of net benefit to the retailer.

Outside agents remain a small channel, and prominence is not demand

Disclosed shares for externally operated agents are low, with different trajectories:

Company	Disclosure	Period
Booking	Significantly below 1% of room nights; little recent change	August 2026
Etsy	Below 1% of traffic	August 2026
Shopify	AI traffic and orders tripled year over year, from a small base	Q2 2026
Hemnet	Significantly below 0.1% of sessions	July 2026

Booking’s figure covers both paid and unpaid LLM channels and has not moved materially over recent months or quarters. Shopify’s grew far faster but management still described AI as small relative to the overall business. Hemnet already had a live ChatGPT integration when it reported.⁷, ⁸, ⁹, ¹⁰ These measures describe different activities and periods; they cannot be pooled into a market-share estimate. Taken together they establish that integrations and partnerships can precede material purchasing activity. They also leave room for influence that no referral metric captures, since a shopper can consult an assistant and then arrive directly.

Angi is the clearest disclosed case against reading visibility as demand. Management said its share of voice in LLM answers was roughly twice its nearest competitor’s, yet that prominence was not converting the way SEO had.¹¹ The two metrics measure different things: share of voice counts presence in an answer, while an organic click came from someone who had already declared intent and taken an action. One possible explanation is that an assistant answer resolves part of the comparison the click used to represent, leaving fewer clicks behind it. Angi did not identify the driver, and the disclosure does not establish one.

There is also direct counterevidence to automatic search displacement. Shopify said traditional search sessions to merchant storefronts stood at 1.3 times their level two years earlier and still made up roughly a third of sessions, while AI traffic and orders tripled.⁹ Both channels grew at once. That does not demonstrate incremental demand, and it does not rule out substitution elsewhere, but it means AI referral growth cannot by itself be read as search traffic lost.

Nor is conversation preferred for every task. Baltic Classifieds said filters worked well and that customers had shown no preference for fully conversational search.¹⁰ That is evidence that the value of a conversational interface depends on the task rather than holding uniformly across shopping.

What that changes for retailers and marketplaces

Different businesses have different reasons to open the catalog

Shopify and MercadoLibre have taken different positions on external access. Shopify wants merchant catalogs, carts, checkout and business rules to function across AI interfaces. MercadoLibre welcomes visibility in external search but has questioned giving ChatGPT its full catalog for end-to-end buying, and is building its own LLM search.⁹, ¹²

The positions are consistent with what each company sells. Shopify supplies infrastructure to merchants and is paid when transactions run through it, so an order that begins in an AI interface can still be one Shopify serves. MercadoLibre also operates the destination itself — the place where buyers discover products and sellers pay to be found. An external agent performing more of the discovery could shift that activity, and the advertising sold against it, outside the marketplace. That interpretation helps explain the stated positions without establishing which strategy will produce better returns.

The contrast is not adoption versus resistance, and not binary. MercadoLibre is deploying AI inside its own service while raising questions about full external catalog access — a stated reservation rather than a disclosed restriction on access it currently grants.

Retaining checkout does not settle who earns the profit

Etsy kept its payment role in purchases originating through OpenAI, but said in October 2025 that it paid a success-based fee on completed purchases and was not passing that fee to sellers. Neither the rate nor the calculation base was disclosed.¹³

That separates two questions usually treated as one: who processes the order, and what the order cost to obtain. Etsy can retain the transaction while absorbing a new distribution cost, and retained GMV says nothing about the second question.

Whether the arrangement is worse depends in part on what the purchase displaced. A fee on a customer who would have arrived directly is an added cost holding other things constant; a fee that replaces paid search spending may be cheaper than the channel it replaces; a fee on a genuinely new customer is ordinary acquisition cost. The disclosed terms cannot distinguish these cases.

Familiar acquisition mechanics are reappearing inside the new channel. Booking is testing cost-per-click advertising with OpenAI. Angi described ChatGPT advertising as producing roughly $3m in annualized revenue at a positive margin, below its Google and Meta channels, with management expecting comparable economics eventually but not yet achieving them.⁷, ¹¹ The Angi disclosure is the more informative of the two: the channel is small, it is already profitable, and by management’s own account it has not yet reached the economics of the channels it would need to replace.

Shopify said it charges its normal transaction fees with no separate agentic charge.⁹ That describes what Shopify charges merchants. It does not describe the merchant’s total cost of selling through an external AI interface, which may include additional fees charged by the interface.

None of the disclosures cited here reports contribution per agent-referred order after fees and acquisition spending, measured against the channel it displaced — the comparison that would show whether retaining checkout is worth what the referral costs.

The second purchase separates acquisition from ongoing access

eBay’s September remarks gave an early clue about what happens after an agent delivers a buyer. Within what management described as a small channel, agent-acquired buyers came back directly for their next purchase about half the time. No cohort size, measurement period or precise denominator was given.¹⁴ This is not a 50% repeat rate across acquired customers; it is a statement about the route taken for a subsequent purchase, with limited supporting detail. Walmart described high organic repeat among new and returning customers arriving through frontier-model partnerships, without a disclosed cohort analysis or precise return-route measure.³

The distinction bears on what a fee of the Etsy kind amounts to. Repeat direct visits may amortize the acquisition cost across later purchases that carry no fee. If instead the agent remains the route for subsequent purchases, a similar fee recurs on customers the retailer already has. Neither pattern is established. eBay’s figure indicates that direct return occurs within one small channel; it does not show how durably it persists, or how it varies by retailer or category.

Delivery and returns bound cross-retailer selection

NEXT identified a physical constraint on shopping assembled across retailers. In its March results discussion, management described average order value after returns of roughly £70 against a delivery cost of about £5, and argued that an agent building a basket from several retailers would multiply delivery costs and complicate clothing returns.¹⁵ This is a management scenario, not a measured effect of agent shopping, and the comparison depends on the actual basket and the delivery terms available.

The mechanism is nonetheless specific. If an agent chooses the best individual items across several retailers without accounting for the combined delivery and returns process, it can leave the customer with a more expensive or cumbersome purchase. Consolidating items into one delivery and handling returns through one process are part of the product, not overhead attached to it — which is why an agent selecting better individual items can still produce a worse purchase.

DoorDash and Instacart emphasize the same class of capability from the other direction: live inventory, stored customer preferences, basket construction and fulfillment.¹⁶ Access to product descriptions does not supply what is needed to deliver a usable order.

These services can preserve a retailer’s or marketplace’s role in the transaction. They do not guarantee its current fee. An agent can leave the fulfillment intact while changing how customers compare providers and what execution is worth.

What sellers will pay for is the open question

MercadoLibre and OLX hold different views about what advertisers buy in an AI-mediated process. MercadoLibre argues that better knowledge of customer intent raises the value of conversion-oriented advertising even if clicks decline. OLX executives suggested boosted listings and paid visibility could become less relevant, with monetization shifting toward better leads and transaction services; OLX also said it had not yet observed a cost from traffic disintermediation.¹², ¹⁷ That last point makes the OLX position a forecast about its own model rather than a reading of something already happening to it.

Both positions concern the product sold to sellers, not whether the marketplace is needed. A platform can remain essential for inventory, trust and transaction execution while finding that prominent placement no longer commands a premium; equally, a more accurate recommendation could produce a lead worth more than the clicks it replaced. Fewer clicks alone does not determine revenue. It depends on whether sellers get better outcomes, whether those outcomes can be attributed, and what the platform can charge for them.

---

Notes

1. Expedia Group, Q2 2026 call, Aug. 5, 2026, on conversational tools: “Those are not driving conversion right now”. ↩
2. Rakuten Group, Q2 2026 call, Aug. 10, 2026: “attributed orders grew by 0.52% and attributed GMS grew by 0.87%”. ↩
3. Walmart, Goldman Sachs Global Retailing Conference, Sept. 15, 2026: “their average order value jumps 40%”; “new and returning customers have a very high organic repeat rate with Walmart”. ↩
4. Magazine Luiza, Q2 2026 call, Aug. 7, 2026: “we reached BRL 100 million of GMV”. ↩
5. Amazon, Q2 2026 call, July 30, 2026: “Over 350 million customers have used it in the last 12 months”. ↩
6. Alibaba, earnings call, Aug. 20, 2026: “their first AI-driven shopping experience”; Allegro, Q2 2026 investor presentation, Sept. 17, 2026, p. 9: “AI Assistant: already serving 500k monthly users”. ↩
7. Booking Holdings, Q2 2026 call, Aug. 4, 2026: “still significantly below 1% of our room nights”. ↩
8. Etsy, Q2 2026 call, Aug. 6, 2026: “traffic from Agentic experiences is still less than 1% of our overall traffic”. ↩
9. Shopify, Q2 2026 call, Aug. 5, 2026: “AI-driven traffic and also orders to Shopify stores tripled year-over-year”. ↩
10. Baltic Classifieds Group, FY2026 call, July 2, 2026: “fully conversational within chat experience is just not what works for customers”; Hemnet, H1 2026 call, July 17, 2026: “traffic from LLMs to Hemnet remains very low, still significantly below 0.1% of our total sessions”. ↩
11. Angi, Q2 2026 call, Aug. 5, 2026: “we’re roughly $3 million revenue run rate”; “spending profitably”. ↩
12. MercadoLibre, MoffettNathanson Media, Internet & Communications Conference, May 13, 2026: “retain that customer inside our platform”. ↩
13. Etsy, Q3 2025 call, Oct. 29, 2025: “a success-based transaction fee”; “we are not passing that through to our sellers”. ↩
14. eBay, investor remarks, Sept. 8, 2026: “about half the time, they come back to eBay directly to drive their next purchase”. ↩
15. NEXT plc, FY2026 results discussion, Mar. 26, 2026: “our average order value, net of returns, be around GBP 70”; “Cost of delivery to the consumer is around GBP 5”. ↩
16. DoorDash, Q2 2026 call, Aug. 5, 2026: “knowing where all the inventory sits. What’s in stock?”; Maplebear (Instacart), management remarks, Sept. 9, 2026: “know about your purchase history, we’re going to know about your preferences”. ↩
17. Prosus / OLX management discussion, Feb. 24, 2026: “lots of techniques will become not relevant in the future, like boosted listings”. ↩