Ecommerce chatbot examples, with the source of every answer

Below are six ecommerce chatbot examples taken apart: the exact question a visitor typed, the document the answer was drawn from, and the reply. One of them is a question the shop's documents do not cover, where the assistant says so instead of inventing an answer, because that case matters more than the five that work.

Most published examples are screenshots of a happy path. They show a tidy answer and hide where it came from, which makes them useless for judging whether the tool would tell your customers the truth. Every example here names its source. If you want the wider picture first, start with the ecommerce chatbot guide and come back.

What counts as a usable example

An ecommerce chatbot example is a single exchange you can inspect end to end: the question a shopper typed, the document the answer came out of, and the words the assistant replied with. Anything short of all three is a screenshot, and a screenshot proves nothing about accuracy.

The shop in these examples is a composite of the document set almost every small retailer already has: a delivery page, a returns policy, a sizing guide, a set of product entries, and a handful of answers the owner wrote out themselves. No named brands, no invented case studies, no revenue figures. Just the documents and what comes back out of them.

Six worked ecommerce chatbot examples

1. A delivery question with a surcharge zone

Question typed: "do you deliver to Hawaii and how long does it take"

Source: the shop's delivery page, pasted in as a text source rather than uploaded as a file.

Standard delivery: 2 to 4 working days, $4.95, free over $50. Express: next working day, $9.95, order before 1pm. Alaska and Hawaii: add two working days and a $9.95 surcharge. We do not deliver outside the United States.

Answer given: "Yes, we deliver to Hawaii. Standard delivery takes 2 to 4 working days plus two extra working days for Hawaii, and there is a $9.95 surcharge on top of the usual $4.95. Express is available on the same terms at $9.95."

The useful part is the surcharge. It sits in the fourth line of a delivery page that a shopper on a phone will not scroll to, and it is the kind of detail that produces an angry email when it is discovered at checkout instead.

2. An exchange question with a deadline in it

Question typed: "I ordered the wrong size, can I swap it? it's been about three weeks"

Source: the returns and refunds policy, uploaded as a PDF.

Returns are accepted within 30 days of delivery. Items must be unworn with original tags attached. One free exchange per order; start it from the link in your dispatch email. Return postage is paid by the customer unless the item is faulty. Sale items are final and cannot be returned or exchanged.

Answer given: "Three weeks is inside the 30 day window, so yes. You get one free exchange per order as long as the item is unworn with its tags on, and you can start it from the link in your dispatch email. Return postage is yours unless the item is faulty."

Notice what it did with "about three weeks". It compared a vague number to a stated 30 day rule and answered the actual question rather than reciting the policy at the shopper.

3. A sizing question the chart cannot answer directly

Question typed: "I'm 5ft 10 and 165lb, which size mid layer do I need"

Source: the sizing guide, a chest measurement table with a fit note underneath.

Chest: S 35 to 37 in, M 38 to 40 in, L 41 to 43 in, XL 44 to 46 in. The mid layer is a regular fit designed to be worn over a base layer. If you are between two sizes, take the larger.

Answer given: "Our chart goes by chest measurement rather than height and weight, so measure around the fullest part of your chest. Medium covers 38 to 40 inches and large covers 41 to 43. The mid layer is a regular fit meant to go over a base layer, so if you land between two sizes, take the larger one."

This is the example most worth studying. The honest answer to the question as asked is "I cannot tell from that", and a guessing assistant would have said "medium" and generated a return. It gave the shopper the one thing that would get them to an answer instead.

Worth knowing: an assistant that converts height and weight into a size your chart does not mention is not being helpful. It is manufacturing a fact, and the cost lands on you as a return.

4. A product compatibility question that shows cards

Question typed: "does the 18v drill take the same batteries as the sander"

Source: two product entries, plus a compatibility note pasted from the tool range page.

All tools in the 18V range share the same battery and charger. Bare tool versions are sold without a battery or charger. Batteries from the older 14.4V range are not compatible.

Answer given: "Yes, both are in the 18V range and share the same battery and charger. The sander is sold as a bare tool, so it does not include one. Older 14.4V batteries will not fit."

Two product cards appeared underneath, because the question was genuinely about products. That distinction is measured rather than guessed: against a real catalogue, questions actually about a product scored 0.61 to 0.79 on the best match, while questions about returns, delivery, accounts and opening hours topped out at 0.53. The gate sits between them, so a returns question never gets a drill card pushed at it.

5. A question answered in the owner's exact words

Question typed: "can I collect in store today"

Source: a curated question and answer pair the owner wrote themselves, not a document.

Q: Can I collect from the shop? A: Yes. Order online, choose Collect in store at checkout, and we will email you when it is ready, usually within two hours during opening hours. Bring your order number and photo ID.

Answer given: the owner's words, verbatim.

This one never reached the language model. A curated pair that matches the question closely enough is served directly, which is measured against a real pair: the exact question scored 1.000, a trivial rewording 0.993, and a genuine paraphrase 0.853. The cutoff sits at 0.92, above the paraphrase, so anything that actually needs rewording still goes to the model. Everything above it is free to serve and identical every time.

Use curated pairs for the answers where wording is a commercial or legal decision: price match promises, guarantee terms, anything a lawyer has read.

6. The question the documents do not cover

Question typed: "who owns the company and what was your turnover last year"

Source: nothing. The uploads are delivery, returns, sizing and products. There is not a word about ownership or accounts.

Answer given: "I do not have anything about the company's ownership or accounts to hand. I can help with delivery, returns, sizing or any of the products, if that would be useful."

This is the example the other five exist to set up. An assistant that will guess here will also guess about your returns window, and you will not find out until a customer quotes it back at you.

The reason it behaves this way is a rule in the prompt, not a filter. There is a similarity floor that throws out obvious junk, but the measured scores overlap in a way that makes the floor useless as a judge of answerability: on a sample corpus a direct hit scored 0.72, a paraphrase 0.54, an unanswerable but on-topic question 0.51, a wholly unrelated question 0.43 and gibberish 0.41. A threshold cannot separate 0.54 from 0.51. So the guarantee comes from instructing the model to answer only from the supplied passages and to say plainly when they do not cover the question.

Write about this as refusal by design. It shrinks the surface for invention by restricting the source material. It does not repeal how language models work, which is why the owner can open any conversation, see the passages an answer was drawn from, and overrule it.

What the six examples have in common

Four patterns run through all six:

• Every answer is two or three sentences. Nobody reads a policy quotation in a chat bubble. The assistant is instructed to keep simple answers short and to use a list only for procedures.
• The source is always inspectable. For each of these, the owner can open the conversation and see which passages fed the answer.
• Nothing was authored for the chatbot. All five answerable examples came out of documents the shop already had.
• The refusal is a feature of the same mechanism. It is not a separate safety layer bolted on; it is what the retrieval rule produces when retrieval comes back empty.

Which questions should you teach it first?

Start with the questions that cost you money when they go unanswered. Baymard Institute puts the average documented cart abandonment rate at 70.22% across 50 studies, and among people who abandon during checkout rather than while browsing, 20% cite delivery being too slow, 13% an unsatisfactory returns policy, and 12% not being able to see or calculate the total cost upfront (Source: Baymard Institute, 2025).

In practice that means this order:

1. Delivery. Options, prices, cut-offs, surcharge zones, and what happens when a parcel is late.
2. Returns and exchanges. The window, the condition, who pays postage, and the exceptions you actually enforce.
3. Sizing, fit and specification. Charts as tables, fit notes as prose, compatibility rules stated explicitly.
4. The three or four questions you answer by email every single day. Write those as curated pairs so the wording is yours.
5. Product entries for the lines you want surfaced as cards.

Anything beyond that is optimisation. If you sell into a specific trade or run a physical shop as well, the retail specific version of this list covers stock, opening hours and collection questions in more detail.

How do you build these examples for your own shop?

You do not build the examples. You upload the documents and the examples are what comes out.

• Create an assistant. On chatfor.site the free plan is $0 forever and covers 100 answers a month, one assistant and 25 documents, which is enough to run a fortnight of real questions past it.
• Upload the four document types above. Uploaded files, pasted text, product entries and curated question and answer pairs are all accepted.
• Paste one script tag into your site, the same shape as adding an analytics tag.
• Then do the part everybody skips: ask it twenty real questions from your own inbox, including three you know are not covered, and read the answers with a critical eye.

Worth knowing: the twenty question test is the entire evaluation. Any tool can be made to look good on a demo question chosen by the vendor. Choose the questions yourself, from your own email, and include the awkward ones.

If you are still deciding which tool to run this test against, our comparison of the options for a small shop sets out the axes that matter. If you want to understand the machinery producing these answers rather than just the output, read how retrieval actually works in plain terms.

Frequently asked questions

What are ecommerce chatbots?

Ecommerce chatbots are chat windows on an online shop that answer a visitor's question in the moment. The useful kind answer from the shop's own delivery, returns, sizing and product information rather than from a scripted decision tree, and tell the visitor when they do not have the answer.

How does a chatbot work in ecommerce?

Your documents are split into passages and indexed. When a visitor asks something, the passages closest in meaning to their question are retrieved and handed to a language model with an instruction to answer from those passages only. The answer is generated from your text, not from general knowledge.

What should a chatbot do when it does not know the answer?

Say so in one sentence and offer something it can help with. Any other behaviour is worse for you: a guessed returns window or delivery date creates a complaint you have to honour or argue about, and you will not know it happened unless the conversation is stored and readable.

Can a chatbot recommend products?

Yes, when the question is genuinely about a product. In these examples the product cards appeared only for the compatibility question, because product matching is gated on a measured similarity score. A delivery or returns question does not trigger cards, which is the behaviour you want.

Do I need to write new content for it?

No. All five answerable examples above came from documents the shop already had: a delivery page, a returns PDF, a sizing chart and product descriptions. The only thing worth writing fresh is a handful of curated question and answer pairs for wording you want controlled exactly.

Sources

• Baymard Institute, Cart Abandonment Rate Statistics: https://baymard.com/lists/cart-abandonment-rate (updated 22 September 2025)