AI Retail Customer Questions — What We've Learned After Seven Months in Stores

Remi has now been running in real independent retail stores for about seven months. That's enough time to start seeing patterns in what customers actually ask an AI concierge — patterns that I would not have predicted from the outside, and that I think are useful for any retailer thinking about deploying conversational AI on the floor. This post is observational. I'm not going to invent specific numbers; I'll describe the categories of questions in their relative frequency and let you draw your own conclusions.

The framing: across the conversations we've logged across our pilot stores, what kinds of questions show up, how often, and what does that tell us about indie retail customers.

The question categories, ordered roughly by frequency

These are the buckets we've ended up grouping conversations into. The order is rough — varies a bit by store and category — but the rank is fairly stable.

1. "What should I bring to / pair with X" (highest frequency)

This is the dominant question category in liquor stores, wine shops, and specialty foods. The customer has a specific occasion or food in mind and is asking for a product fit.

Sub-patterns within this:

• "I'm bringing wine to dinner, they're cooking [protein]." Very common. Almost always under a $30 budget. The customer wants a recommendation that won't be embarrassing.
• "What goes with [dish] for a dinner I'm hosting." Smaller volume but higher tickets — the host is buying for multiple courses.
• "What's a good gift for someone who likes [spirit category]." Surges in December but present year-round for birthdays.

The takeaway: a meaningful share of indie retail customers are not shopping for themselves. They're shopping for a social moment. The recommendation needs to account for the occasion, not just the customer's own taste.

2. "Do you have / where is X" (high frequency)

Customers asking whether the store carries a specific product, and where it's located if so. This is a "save the customer a walk around the store" question.

Two interesting subpatterns:

• In-stock check. Customers often ask whether the store has a specific brand or SKU. The kiosk's value here is binary — yes/no with a shelf location. Saves 90 seconds and saves the customer asking a clerk who's at the register.
• Out-of-stock landing. When the store doesn't carry the item, the kiosk's job is to suggest a near-alternative. This is where the conversation often leads to a sale that wouldn't have happened — the customer was about to walk out.

The data we capture here is also valuable to the store: a running list of "things customers asked for that I don't carry" is purchasing intelligence the store didn't previously have.

3. "Recommend me something like X but [different attribute]" (moderate frequency)

The variant question. The customer knows a product, and is asking for something similar but cheaper, smoother, lighter, more local, or different in some specific way.

This is the question where AI concierges most clearly outperform static signage or printed materials. The variant is usually personal ("I usually drink X but I want to try something") and the answer needs to be specific to the store's actual shelf. Generic "if you like X try Y" lists tend to be wrong because they don't know what the store carries.

This is also the question where staff knowledge matters most — an experienced clerk handles this question well; a new hire often can't. The kiosk extends the experienced clerk's knowledge to every shift.

4. "What's new / what's good right now" (moderate frequency)

The exploratory customer. They don't have a specific need; they're browsing. They want to know what the store thinks is interesting.

This question category is interesting because it's where the store's voice — the curation, the editorial point of view — most directly shows up. The kiosk persona is configured per store, so a wine shop with a strong sommelier perspective gets a different "what's new" answer than a convenience store does.

It's also a high-engagement category. Customers who ask "what's new" tend to ask follow-ups, tend to buy, and tend to come back asking the same question on later visits. They're treating the kiosk as a discovery tool, which is exactly what it's designed for.

5. "How much is X" / pricing questions (moderate frequency)

Direct pricing checks. Lower volume than I'd expected, actually — customers often check the shelf tag rather than asking the kiosk. But the question shows up enough to matter.

The interesting pattern: when the price is asked, it's often followed up with a comparison ("is there a cheaper version of this") which routes back into category 3. Pricing questions are often the start of a recommendation conversation, not just an isolated lookup.

6. "Is X good / what's it taste like" (moderate frequency)

The opinion question. The customer is considering a specific product and wants a tasting note, a description, or a credibility signal.

This is the category where we are most careful. We don't generate fake tasting notes. If the store has provided notes for a SKU, the kiosk surfaces them. If not, it gives a category-level description ("this is a typical mid-shelf bourbon, sweeter than rye, good for sipping or in cocktails") and offers to compare it to something the customer might already know.

The data here points to an important truth: customers want a credible-sounding opinion before buying. The store that has internal tasting notes loaded has a meaningful edge. The store that doesn't is leaving money on the table.

7. "Hours / location / event" questions (low to moderate frequency)

Logistical. When are you open, where are you, do you do tastings, can I bring my dog. These are fine and the kiosk handles them well, but they aren't the high-leverage conversations. They mostly free up the phone line.

8. Off-topic questions (low frequency, but interesting)

A small share of conversations are off-topic — customers asking the kiosk about the weather, sports, current events, or just chatting. We mostly redirect politely back to "I'm here to help you find what you need at this store, but happy to chat for a moment." We do not pretend to be a general-purpose AI; the persona is store-specific.

This redirect matters for trust. If customers think the kiosk is ChatGPT-with-a-store-skin, the conversation drifts and the store value evaporates. The persona has to be focused.

9. Complaints / problems (low frequency, very high importance)

Customers occasionally use the kiosk to register a complaint — wrong price, expired product, bad experience with a staff member. These are rare in absolute terms but the most important signal in the system.

We surface them immediately to the store dashboard so the owner sees them within hours, not at end-of-week. A store that handles a complaint within a day keeps the customer; one that learns about it a week later usually doesn't.

Patterns that surprised me

A few observations from seven months that I would not have predicted:

Customers introduce themselves more than I expected. Not formally, but they volunteer context — "I'm new to the area," "I usually drink wine but I want to try a cocktail tonight" — at high rates. People talk to the kiosk in a way that gives it useful context, even without being prompted to.

Voice and text usage are roughly even. I had assumed voice would dominate; it depends on the store environment. Quieter stores see voice; loud or shy stores see text. The kiosk supports both, which has turned out to matter more than I'd realized.

Repeat customers come back to the kiosk. I had assumed novelty would be most of the engagement. It isn't. Customers who used the kiosk on a first visit often come back to it on later visits, treating it as a tool. This is a stronger signal than first-visit engagement.

Bilingual conversations are routine. A non-trivial share of conversations switch between English and Spanish within the same session, especially in our San Diego pilot stores. The kiosk handles this; the staff would have struggled to.

Cross-category questions surface gaps. Customers ask questions that span categories the store organizes separately ("what wine goes with the cheese over there"). The kiosk handles this naturally because it sees the whole catalog as one. The store often doesn't.

What I'd do with this data if I were running a store

If you're an indie retailer thinking about how to use this kind of data when you have it, here's what I'd prioritize:

Mine the out-of-stock asks. The list of "products customers asked for that I don't carry" is your purchasing input for the next quarter. This is the single highest-leverage report.

Read complaint conversations the day they come in. Don't let them sit. The dashboard alerts on them; the owner should respond same-day.

Look at the variant questions. "I usually drink X but I want Y." This tells you what's an acceptable substitute for what, in your customers' minds. That's information you can use in shelf adjacency, in display, and in staff training.

Configure the persona to match your shop. A good wine shop's "what's new" answer should sound like the owner. A good convenience store's should be fast and direct. Don't run the default persona — set the voice.

Tag the high-value conversations for review. A small set of conversations are unusually long, unusually rich, or unusually tied to high-ticket purchases. Reviewing them weekly is a good staff training input — these are the conversations your floor staff would want to be having.

For more on the framework that informs all of this, see the great-grandfather test and the AI upsell vs POS prompts piece. If you want to see the dashboard side of this data in a live demo, book a session.

Caveats

This is observational data from a small set of pilot stores over seven months. The patterns are stable enough that I'm comfortable describing them, but I'm not making category-wide claims. I'm also not citing specific percentages — frequency varies by store and the absolute numbers are small enough that single-store percentages would mislead. When we have a year of data across more stores, I'll publish a quantitative version. For now this is the qualitative read.

Frequently asked

How long until you'll publish actual numbers?

When we have a representative sample across enough stores and enough time. Probably another six to nine months for a defensible quantitative version. We'd rather wait than publish numbers that overstate confidence.

Can store owners see their own conversation data?

Yes. Every store has a dashboard view of their own conversations, broken down by category, with filters and a "asked for but not in stock" report. The data the store generates belongs to the store.

Do customers know their conversations are logged?

Yes — disclosure is on the kiosk. Conversations are stored to improve the service and inform purchasing; they're not sold to third parties. We treat the conversation as a store asset, not ours.

What's the single highest-value question category for a store to optimize for?

The "out-of-stock" asks — the product requests the store can't fulfill. That list, read weekly, becomes a buying input that pays back the kiosk subscription many times over in most stores.

How does this differ from in-store traffic analytics?

Traffic analytics tell you that customers were in the store. Conversation data tells you what they wanted. The first is an input to layout and staffing; the second is an input to merchandising and inventory. They're complementary.