OTTO's shopping assistant: guided discovery before agentic checkout

External-source research synthesis, not an OfferSignal client result. Based on the retained October 2026 research and source-specific dates.

A shopper can know the need—an allergy-suitable vacuum or a washing machine for a one-person household with a dog—without knowing which catalogue filters express it. OTTO launched a Gemini-based conversational advisor in its German consumer app on 5 March 2026, with text and Voice Chat, narrowing questions and product recommendations. The feature was rolled out in stages as a beta; independent German trade coverage confirmed progressive enablement rather than a completed universal release. 1 4 5

What the beta actually does

The advisor is a product-discovery and decision-support layer. OTTO describes using product descriptions and structured attributes across more than 18 million articles, with product reviews planned next. An OTTO annual-report story illustrates a refrigerator conversation that gathers household size, appliance type, capacity and desired features before recommending products. The shopper remains the chooser and buyer. 1 3

This is separate from OTTO's customer-service assistant, which handles orders, delivery, returns and invoices through internal service connections and human escalation. Those integrations do not show that Smart Shopping can access an individual's order, change it, assemble a cart or pay. OTTO places such capabilities in a prospective “could” horizon. Google's case describes routine reordering and agent-to-agent commerce as exploratory and says the system should not operate on autopilot. 1 2 5

Data and orchestration matter more than the chat surface

Google Cloud's supplier case says OTTO began building the advisory experience in November 2024 and rebuilt the voice path on Gemini Live native audio in November 2025 after an initial speech-to-text/text-to-speech approach produced pauses and terse commands in remote tests. The text assistant retains guardrails, recommendations and search; the audio model handles dialogue, with Chirp 3 for German speech recognition. Candidate products are retrieved and re-evaluated for fit by Gemini 2.5 Flash, while orchestration controls include a constraint against medical advice. These are provider-reported implementation details, not an independent audit. 2

The catalogue figures need careful reading. Google's case refers to about 19 million products, while OTTO's release says more than 18 million articles; Google also mentions 6,100 marketplace partners. These are catalogue-scale figures, not users or successful matches. Its result panel says the advisory scaled “from 5 to 50 categories,” while the body says five expert-crafted categories were extended to “fifty more.” The endpoint is ambiguous, and no category-quality or attribute-completeness measure is published. OTTO's 12 million active customers and the app listing's 10M+ downloads are retailer/app scale, not assistant adoption. 1 2 3 7

The reported result is engineering evidence

Google reports end-to-end voice latency falling from 8–9 seconds in early prototypes to under two seconds after the native-audio rebuild. The case gives no sample, task mix, percentile, device or network conditions, or independent replication, so this is not a benchmark of recommendation quality or commercial value. 2

Google also says voice entered open beta in early May 2026 and that production data was only days old. It reports richer contextual detail in voice and illustrates one lamp-replacement session in which a customer mentioned a broken wire. That is a qualitative observation, not a controlled comparison. No reviewed source reports assistant users, query volume, recommendation acceptance, conversion, average order value, return-rate change, satisfaction, revenue, cost-to-serve or return reduction. OTTO's suggestion that better advice could reduce wrong purchases and returns remains a proposed mechanism. 1 2 3 4 5

A predecessor test in July 2023 covered 180,000 products on desktop and app and used existing descriptions and reviews to learn about generative AI. It is not performance evidence for the 2026 beta. 6

An illustrative merchant test

For a distributor or manufacturer with a large consumer-facing catalogue, this is an illustrative application, not a reported OTTO result. Start with one category where buyers routinely describe requirements in ordinary language. Version the canonical product identity, structured attributes, approved alternatives, exclusions, price and stock authorities, and review provenance. The assistant may clarify and recommend, but the commercial owner remains authoritative for fit, availability, price and purchase.

Replay real queries against conventional search and the conversational flow. Measure clarification, correct-family and wrong-variant rates, attribute completeness, stale-price or stock exposure, recommendation corrections, latency and returns over an agreed window. Include requests where the system should abstain. Keep approval before cart or payment. OTTO's case supports governed discovery and structured product data; it does not support RFQ pricing, negotiation, equipment-fit guarantees, cart placement or payment authority.

Sources and next reading

• OTTO Smart Shopping announcement — official launch and phased-beta description.
• OTTO consumer app on Google Play — public app route; listing does not establish Smart Shopping eligibility.
• Google Cloud OTTO voice-shopping case — supplier case on implementation and voice beta; not an end-user access link.

Explore OfferSignal · Focused Business & AI Workshop · All research

References