What customers actually ask an e-shop chatbot
100 real conversations from the chatbot on antoniadis.com.cy: 59% are product searches, 53% start outside opening hours, and Sunday is the busiest day.
Anyone thinking about adding a chatbot to their e-shop usually pictures a specific scenario: customers asking “where’s my order?” and “how much is shipping?”, and a bot that saves the team from answering the same ten emails every day. It’s a reasonable picture. It just isn’t what we see when we open the logs.
This article is based on 100 real conversations from the chatbot on antoniadis.com.cy, a PrestaShop store in Cyprus selling electronics, STEM and robotics kits, craft supplies and educational products. It’s our own shop and our own bot, which is why we can show the numbers, including the ones that don’t flatter us.
The short answer
- 59% of conversations are product searches: the customer couldn’t find something with the site search and asks in plain language.
- 53% start outside opening hours. Sunday, the only day the shop is closed, is the busiest day of the week for chat.
- 83% are resolved in three questions or fewer.
- 14 in 100 needed a human, and in one, the bot didn’t realise in time that it should bring one in.
Where the data comes from
We looked at the 100 most recent conversations on the chat widget, from 23 August to 19 September 2026: 28 days and 339 visitor messages. The bot handled 96 of them; 4 were passed to a person.
Two caveats, so the numbers are read correctly. First, these are the most recent conversations, not a random sample. Second, this is one shop with a very specific catalogue: a clothing store would most likely look different. All data here is aggregated and anonymised: no names, emails, order numbers or verbatim conversations. The examples below are paraphrased.
What they ask: the categories
Many conversations cover more than one topic: 1.44 on average, with 62 of 100 sticking to a single one. That’s why the percentages don’t add up to 100: they show how many conversations each topic appeared in.
| Topic | Conversations |
|---|---|
| Product search / finding a product | 59% |
| Technical specs, compatibility, equivalents | 20% |
| Location, opening hours, contact, pickup | 19% |
| Custom work: cutting, dimensions, quantities | 11% |
| Shipping costs, times, destinations | 10% |
| Availability / when it’s back in stock | 7% |
| How to order / payment methods | 7% |
| Where’s my order / tracking | 6% |
| Order or payment problem, complaint | 3% |
| Off-topic / B2B pitch / spam | 2% |
“Where’s my order?”, which dominates most chatbot sales pitches, is just 6% here. Something else dominates.
Why chat is a second search box
Nearly 6 in 10 conversations start because the visitor has already failed to find something with the site search. Chat isn’t working as a customer-service channel here; it’s a second search, in natural language. Three patterns explain why the first one fails:
- A bare part number. Someone types “TPS54331” or “BC550”, or a code with “smd” tacked on. They know exactly what they want. The site search doesn’t catch variants, suffixes or different packages of the same component.
- A description instead of a name. “Wooden wheels that spin, like a wheel of fortune.” “A round light for crafts.” “Something to hold up two posters.” There’s no keyword to type.
- A problem, not a product. “I want to repaint an old fridge with chalkboard paint — how?” “What thickness of plexiglass for a tabletop?” “What do I need for decoupage?” The right answer isn’t a product; it’s a basket.
Add the language mix on top: many questions arrive in Greeklish (Greek typed in Latin characters), or with Greek and English in the same sentence, typos included. Classic site search can’t handle that. A language model can.
When they ask: Sunday
The shop is open Monday, Tuesday, Thursday and Friday 09:00–18:30, Wednesday 09:00–14:30, Saturday 09:30–13:30. It’s closed on Sunday. Against those hours, 53 of 100 conversations started when nobody was there to answer.
| Out-of-hours window | Conversations |
|---|---|
| Sunday (closed all day) | 17 |
| Weekday after 18:30 | 12 |
| Wednesday afternoon (after 14:30) | 9 |
| Weekday before 09:00 | 8 |
| Saturday outside 09:30–13:30 | 7 |
| Total | 53 |
By day: Monday 13, Tuesday 12, Wednesday 14, Thursday 17, Friday 16, Saturday 11, Sunday 17. The day the shop is completely closed is the busiest day of the week for chat, tied with Thursday. And Wednesday, which closes at 14:30, is the only working day with more conversations outside hours (9) than inside (5).
Overall, 28 conversations happened at the weekend and 14 between 22:00 and 08:00. Without chat, more than half of these questions would have waited for the shop to open, and the peak falls exactly when there’s nobody there.
Who asks, and how long a conversation lasts
68% of conversations were in Greek and 32% in English. 82 came from Cyprus, 7 from Greece and 11 from ten other countries, from the UK to Australia. One arrived entirely in Hungarian, and was answered in Hungarian. The 18% from outside Cyprus is a second argument: multilingual support without a multilingual team.
Most conversations are short:
| Visitor messages | Conversations |
|---|---|
| 1 | 27 |
| 2 | 28 |
| 3 | 28 |
| 4–8 | 14 |
| 14, 29 and 53 | 3 |
83 of 100 are done in three questions or fewer. The three very long ones are the same story with different subjects: someone is building a project (an Arduino build, a school assignment, party decorations) and uses the bot as an adviser. There, the bot isn’t answering a question; it’s helping put together a whole order.
What it shouldn’t answer on its own
14 of 100 conversations needed a human decision. Only 4 were actually passed to a person through the chat. The cases fall into three groups:
- Custom work and dimensions (11%). Plexiglass cut to size, five pieces at a specific size, sixty favours for an event, polystyrene cylinders in a size that isn’t in the catalogue. The bot can collect the specs. Whether it can be done, and at what price, is a human decision.
- Order or payment problems (3%). A card charged while the page showed an error; an order running late and a customer saying they’ll buy elsewhere. Here, every minute of delay costs a customer.
- A complaint that escalates. This is the part that didn’t go well. In one conversation, a customer objected to the shipping charge to Greece, and the tone rose with each message until they called us scammers. The bot kept politely explaining how shipping is calculated. At no point did it offer to put the customer in touch with a person.
The problem wasn’t what the bot said; it was that it kept saying it. An angry customer doesn’t want a more accurate explanation: they want to be heard by someone. The lesson for anyone setting up a chatbot: handoff rules can’t rely only on the topic (“payment”, “refund”). They have to catch the tone too.
What to do with this if you run an e-shop
If you already have a chatbot, open last month’s logs and do three things:
- Collect the searches that found nothing. Some are products you stock but describe differently from how people search: fix those titles and descriptions. Others are products you don’t stock: that’s your chat log telling you what to add to the catalogue.
- Turn recurring questions into content. Opening hours, pickup, shipping, “what do I need for…”: put them on your pages, with proper structured data. It’s the same content AI engines read when deciding whether to recommend your shop, and so do their crawlers. If you’d rather hand that off, it’s what our SEO & GEO for e-commerce service does.
- Write down explicitly when the bot brings in a human, based on topic and tone.
If you don’t have a chatbot, ask one simple question: who answers in your shop on Sunday? If the answer is “nobody”, and your catalogue has things people describe rather than name, you fit the profile these numbers describe.
The Pythia Works AI chatbot is the same bot that produced this data, with natural-language product search, live availability checks and handoff to a human. For what it costs per plan and per model, see the AI chatbot pricing. If your e-shop is still on the drawing board, start with how much a website costs in Cyprus. And if you’d like to see what your own logs would show, drop us a line.