A customer wants a sofa that fits a narrow room, survives life with a dog and can arrive before a move. Your filters know dimensions, materials and delivery dates. They do not necessarily know how to turn the whole situation into a useful shortlist.
This is the problem an on-site AI shopping assistant is meant to solve. It lets a customer describe a need in ordinary language, refine it through follow-up questions and move from an uncertain brief to products they can actually buy.
That can be useful. It is not automatically a conversion strategy.
An assistant built on incomplete product data, unreliable stock information or weak measurement may produce confident recommendations without improving the buying journey. Before adding another interface to the site, a retailer needs to decide where conversation genuinely helps, what the system is allowed to say and how success will be proven.
Answer in 60 seconds
An AI shopping assistant is worth testing when customers face a real product-finding problem that search, filters and category pages do not handle well. The strongest candidates usually involve several constraints, uncertain terminology, comparison or follow-up questions.
Do not start with a site-wide assistant. Start with one task, one part of the catalogue and a defined group of eligible visitors. Keep conventional search and navigation available.
Before the pilot, confirm that the assistant can retrieve authoritative product, price, stock, delivery and policy data. Decide what it must not infer, how personal data will be handled and who owns incorrect answers.
Measure the whole path from exposure to fulfilled order, including activation, useful recommendations, product views, basket additions, cancellations, returns and complaints. A shorter session or a high chat count is not, by itself, evidence of a better shopping outcome.
The decision should end in one of three places: run a bounded pilot, repair the foundations first, or do not build.
In this article
- What is an AI shopping assistant?
- When does conversational product discovery help?
- Does customer interest justify the investment?
- Five signs your ecommerce site is not ready
- What the assistant needs behind the interface
- How to run a bounded pilot
- How to measure the result
1. What is an AI shopping assistant?
An on-site AI shopping assistant is a conversational product-discovery interface operated by, or embedded into, a retailer’s website. It interprets a customer’s request, asks for missing information and recommends products using the retailer’s catalogue and commercial rules.
It is useful to separate four experiences that are often grouped under the same label:
| Experience | Main job | Typical risk |
|---|---|---|
| Customer-service chatbot | Answers questions about orders, delivery or policies | Gives generic or outdated support answers |
| On-site shopping assistant | Helps a visitor discover and compare products | Recommends unsuitable or unavailable products |
| External AI referral | Sends a shopper from an AI search or answer engine to the retailer | Attribution is incomplete or lost |
| Purchasing agent | Selects, orders or pays on the customer’s behalf | Permission, payment and accountability become more sensitive |
This article is about the second category. If the system can change a basket, place an order or trigger another action, it moves into a higher-risk product and security problem. Our guide to AI agent security and product controls covers that wider boundary.
The distinction matters because a pleasant chat interface does not prove that the underlying product-finding system works. The commercial value comes from the quality of the shortlist and the next step, not from the conversation alone.
2. When does conversational product discovery help?
Conversation is most useful when a customer’s need is difficult to express as a short keyword or a sequence of independent filters.
Good candidate tasks include:
- choosing among products with several interacting constraints;
- translating an intended use into attributes when the customer does not know the category language;
- refining a broad brief and explaining the final trade-offs.
Research on conversational recommender systems supports the basic mechanism: dialogue can help elicit preferences, resolve uncertainty and collect feedback. It does not establish a universal commercial uplift. Evaluation methods vary, and a recommendation that sounds helpful can still be wrong.
Controlled evidence points to a trade-off. A Microsoft experiment found faster, more satisfying consumer-choice tasks when the LLM result was correct, but also a risk of over-reliance when it was wrong. A smaller hybrid-store experiment found that conversation complemented normal browsing and did not improve every measured outcome.
Do not make a customer chat when they already know the product name, want to apply a simple size filter or need to see the whole range. Search, filters, comparison tools, category pages and merchandising still have jobs to do. Baymard’s product-finding research continues to document avoidable failures in these conventional interfaces. Adding AI does not repair them.
3. Does customer interest justify the investment?
Recent French surveys suggest that AI-assisted shopping is no longer a fringe behaviour. They do not show that every retailer needs an assistant.
The results need careful reading because the studies ask different questions:
| Study | What it suggests | What it does not prove |
|---|---|---|
| FEVAD/Odoxa, January 2026 | In a survey of 1,500 French online shoppers, 31% reported using generative AI during shopping. Trust was higher before purchase than at the point of purchase. | That an on-site assistant improves conversion or that shoppers will delegate a transaction |
| Adyen/Censuswide, May 2026 | In a survey of 2,000 French consumers, 42% were open to an AI-supported journey extending to payment. | Actual use, or unconditional trust in autonomous purchasing |
| Checkout.com/Censuswide, March 2026 | In a survey of 2,002 French respondents, substantial groups expressed reluctance to delegate or uncertainty about who should manage an agent. | That customers reject all forms of AI-assisted discovery |
These figures should not be averaged: using AI during research, accepting AI-supported payment and delegating a purchase are different behaviours. Together, the studies suggest both interest and hesitation. That favours assistance first - useful, transparent and easy to leave - while the business case comes from a specific journey problem, not adoption headlines.
4. Five signs your ecommerce site is not ready
1. Search and product data are already unreliable
If customers cannot trust availability, variants, delivery dates or product attributes, the assistant will inherit the same weaknesses. Repair the source data and ordinary discovery path first.
2. There is no defined high-friction task
“We need AI on the site” is not a use case. “Help customers choose a compatible replacement part without knowing the model number” might be. Start with a customer decision that is valuable, repeated and currently difficult.
3. The system cannot reach an authoritative answer
Recommendations may depend on catalogue attributes, live stock, regional delivery, promotion rules, returns policies or compatibility data. If the assistant cannot retrieve the right source at the right time, it must not improvise.
4. Nobody owns multi-turn quality
The first answer is only the start. Recommendations can forget constraints or deteriorate later in the comparison. A recent preprint shopping benchmark also found weaker performance on optional criteria and later turns. Test complete conversations against expected outcomes and assign an evaluation owner.
5. There is no measurement or stop decision
Without a defined eligible audience, control group, event model and exit criteria, the team may celebrate usage without knowing whether the assistant improved the journey.
Decide in advance what would justify expansion, correction or shutdown.
5. What the assistant needs behind the interface
The visible conversation is only one layer. A dependable implementation needs five foundations:
- Structured, current product data. Use stable product and variant identifiers, accurate attributes, price, availability and any delivery or returns information needed for the decision. The assistant should retrieve product facts, not reconstruct them from marketing copy.
- Approved sources. Define which systems are authoritative for catalogue, stock, price and policy answers. The assistant must be able to say that it cannot confirm something.
- A hybrid experience. Let customers open product pages, edit filters, compare alternatives and return to browsing. Preserve their constraints and explain recommendations with checkable facts.
- Privacy, transparency and control. Make the AI interaction clear, explain relevant data use and do not request personal information that the task does not need.
- Operational ownership. Assign owners for catalogue quality, model behaviour, analytics, incidents and commercial results.
For France and the EU, the exact obligations depend on the implementation and data flow. European Commission guidance states that relevant Article 50 AI transparency obligations apply from 2 August 2026. The CNIL and CIANum have also highlighted the additional privacy, cybersecurity and responsibility risks created by agentic systems with memory, external connections or complex action chains.
This is not a substitute for a legal assessment. Map the provider, deployer, processor and data-controller roles for the actual product before launch.
If the pilot may later become a permanent capability, use the same governance discipline described in our AI pilot-to-production checklist.
6. How to run a bounded pilot
A useful pilot is small enough to diagnose and realistic enough to affect a genuine decision.
- Choose one shopping job. Use service transcripts, search logs, zero-result queries, exits and customer research to find a repeated difficulty. Write it as an outcome: “Help a first-time buyer shortlist three compatible products within budget and explain the trade-offs.”
- Limit the catalogue and audience. Select a category with usable data and meaningful choice. Define who is eligible and who remains in the comparison group.
- Define the answer contract. Record approved sources, facts requiring live verification, prohibited inferences, uncertainty language, hand-off rules and the route out of the conversation.
- Build an evaluation set. Include ambiguous briefs, conflicting constraints, unavailable products and follow-ups that change the request. Score constraint retention, accuracy, shortlist quality, explanation and final hand-off.
- Instrument the journey. Connect assistant events to product views, baskets, orders, fulfilment, returns and support outcomes. Keep the unassisted route measurable.
- Launch gradually. Use controlled exposure, review real failures and maintain a kill switch.
7. How to measure the result
The measurement chain should follow the customer, not the interface:
Eligible visit → assistant exposure → activation → first useful result → constraint refinement → product view → basket → order → fulfilled order → return, cancellation or complaint
Use that chain to avoid five mistakes:
- Exposure is not activation. Report who saw the assistant and who chose to use it.
- Conversation volume is not success. More messages may mean engagement, confusion or recovery.
- A placed order is not a retained order. Watch cancellations, returns and support contacts.
- Speed is not always efficiency. A short session can be a fast decision or an early exit.
- Assisted and unassisted customers differ. A controlled rollout or randomised experiment is stronger than comparing self-selected users after the fact.
Before launch, set:
- a primary customer or commercial outcome;
- guardrail metrics for accuracy, returns, complaints and latency;
- a minimum evidence window appropriate to your traffic and buying cycle;
- explicit expand, repair and stop criteria.
If the team cannot agree on those four items, the pilot is not ready.
Conclusion
An AI shopping assistant can improve a difficult product-finding journey. It can also add an impressive interface to unresolved catalogue, UX and measurement problems.
The right starting question is not “Which model should we add?” It is “Which customer decision is currently hard, and can conversation improve it without reducing accuracy, control or trust?”
If the answer is specific, the product data is dependable and the result can be measured through to the retained order, run a bounded pilot. If those foundations are missing, repair them first. If conversation adds no clear advantage over search, filters or comparison, do not build it.
Turn one product-finding problem into a testable decision
Bring us one journey where customers struggle to choose. In a 30-minute scoping call, we will examine the task, catalogue readiness, customer path, measurement requirements and risk boundaries.
You will get an honest first view: pilot, repair the foundations, or keep the existing experience.
Frequently asked questions
It is a conversational interface that helps customers discover, refine and compare products using the retailer’s catalogue and rules. Unlike a support chatbot, its main job is product selection rather than order service.
Usually not. Conversation is better suited to exploratory or constraint-heavy tasks. Search, filters, navigation and product pages remain more efficient for exact or simple requests. A hybrid experience lets the customer use the right tool at each point.
It may improve a specific discovery journey, but there is no universal conversion guarantee. The result depends on the use case, product data, recommendation quality, UX and measurement design. Test it against an appropriate comparison path and track downstream outcomes.
The exact fields depend on the category. Common requirements include product and variant identifiers, attributes, compatibility, price, stock, delivery and returns information. Each fact should have an identified authoritative source.
GDPR can apply when personal data is processed. The obligations depend on the data collected, purpose, legal basis, vendors, retention, profiling and international transfers. The assistant may also fall within AI transparency requirements. Complete a data-flow and legal review for the actual implementation.
Start with one repeated product-finding problem in a limited category. Test factual accuracy, constraint retention across follow-ups, shortlist usefulness, hand-off quality and the full journey to fulfilment and returns.
Sources and evidence note
This article was developed from independent research rather than from a single market article. Material sources included:
- FEVAD/Odoxa research on French consumers and AI-assisted shopping
- Adyen France consumer research on AI-supported shopping
- Checkout.com France research on delegation and trust
- Conversational recommender systems: survey and research directions
- Shopping Reasoning Bench: multi-turn product-recommendation evaluation
- Microsoft Research randomised comparison of traditional and LLM-based consumer search
- NIM hybrid-store experiment on conversational shopping
- Baymard product-finding research
- Google product structured-data guidance
- Shopify catalogue search documentation for agents
- European Commission guidance on AI Act transparency obligations
- European Commission Q&A on Article 50 transparency obligations
- CNIL and CIANum note on agentic AI
The cited surveys use different samples and question wording, so their percentages are not directly comparable. The experimental and benchmark literature tests particular tasks and settings; it does not establish a universal ecommerce revenue effect. Legal and regulatory status should be rechecked immediately before publication.