What chatbots for retail should answer before a sale

Retail shoppers ask a short, repetitive set of questions before they commit: is it in stock, will it fit, when will it arrive, can I send it back, and are you open. Chatbots for retail earn their place by answering those five instantly from your own written policies, and by saying plainly when they cannot.

Retail is a narrow question space and a wide product range. A shopper does not want a conversation. They want one fact, in under ten seconds, so they can either add to basket or leave. That is a much easier problem than open-ended customer service, and it is why a document-grounded assistant fits retail better than it fits almost anything else. The broader picture sits in the ecommerce chatbot guide; this article is about the retail specifics.

What are chatbots for retail?

A chatbot for retail is an assistant on a shop's website that answers pre-purchase questions from the shop's own material: delivery policies, size charts, returns terms, product descriptions and opening hours. It replaces the search that shoppers do not do and the email they cannot be bothered to send.

That is a deliberately small definition. It excludes the things vendors like to promise and retail buyers should be sceptical about: live stock lookups, order tracking against a warehouse system, personalised recommendations from browsing history. Those require integrations, not documents. Start with the questions documents can answer, because that is most of them.

The five questions shoppers ask before they buy

These five cover the majority of pre-purchase contact in most shops. What matters is not that you can answer them, it is where the answer lives so the assistant can find it.

The pattern is worth noticing. Four of the five are policy, and policy is exactly what a document-grounded assistant does well. Only stock is genuinely live, and that is the one to be careful about.

Worth knowing: Baymard Institute puts the average documented cart abandonment rate at 70.22% across 50 studies, and the top reasons given, excluding shoppers who were only browsing, are extra costs being too high at 40%, delivery being too slow at 20%, an unsatisfactory returns policy at 13%, and not being able to see or calculate the total cost up front at 12% (Source: Baymard Institute, 2025). Three of those four are delivery and returns questions. An assistant does not fix a slow courier, but it does stop a shopper guessing.

Why do shoppers ask instead of reading the page?

Because the page is not read. Nielsen Norman Group's research on reading behaviour found that "users have time to read at most 28% of the words during an average visit; 20% is more likely" (Source: Nielsen Norman Group, 2008). Your delivery page can be complete, accurate and three clicks from the basket, and the shopper will still not find the line about Northern Ireland.

There is also a second, less obvious reason. Retail questions are conditional. "How much is delivery" has no single answer once you have regional pricing, a free-delivery threshold and an express option. A page has to present all the branches at once, which is why delivery pages read like tax guidance. A question can collapse the branches: the shopper says where they are and what they want, and gets one line back.

That is the useful mental model for chatbots for retail. Not a replacement for your content. A way of answering the specific version of a general question.

Stock and sizing: where a document-based assistant needs care

Be honest about these two, because getting them wrong is how a shop ends up apologising.

Stock. An assistant that answers from documents does not know your stock levels. If you upload a spreadsheet of stock counts on Monday, it will confidently quote Monday's counts on Friday. There is no timestamp on a retrieved passage and no way for the model to know it is stale. So do not upload stock counts at all. Instead, write a curated answer that tells the shopper where the live number is: "Availability is shown on each product page, under the price. If a size shows as unavailable, use the notify option and we will email you when it is back." That is true on Monday and true on Friday.

Sizing. This works well, but only if the chart is a table rather than a paragraph, and only if the awkward exceptions are written down. Shoppers do not ask "what is your size chart", they ask "I am usually a 10 in jeans, what should I get in the wide leg". That question can only be answered if somebody wrote down that the wide leg runs large. Most shops know this and have never put it in writing anywhere. Writing it down is the whole job, and the structural side of it is covered in the how to write source material an assistant can retrieve.

Products themselves are handled separately from documents. Each product gets an entry with a name, price and description, and when a question is genuinely about a product, up to four matching products are shown as cards beside the answer rather than described in the text. Questions about returns, delivery and opening hours do not trigger cards, which is the behaviour you want: nobody wants a jacket recommended at them when they asked about parking.

What should a retail chatbot do when it does not know?

Say so. This is the one design decision that separates an assistant you can put on a shop's website from one you cannot.

chatfor.site answers only from the material the owner supplied, and is instructed to say plainly when the documents do not cover a question rather than fill the gap. Ask it who the managing director is, with nothing about that in the uploads, and it says it does not have that to hand. In retail that matters more than in most sectors, because the invented answers are all expensive ones:

• A returns window that is longer than your real one, which you then have to honour or argue about.
• A delivery date you cannot hit, which becomes a complaint rather than a query.
• A material or allergen claim that is wrong, which is a safety problem, not a service problem.
• A price from a competitor's range that the model has seen somewhere and reproduced.

This is a reduction in the surface for invention, not a repeal of how language models work. The mechanism is narrow and worth understanding: the retrieved passages are the only permitted source of fact, and the instruction not to guess is explicit. It makes the failure mode "I do not have that" instead of a plausible fabrication. You can also open any past conversation in the dashboard, see the exact passages an answer was drawn from, and overrule it.

Worth knowing: a shop assistant who says "let me check" is trusted more than one who guesses. The same is true here. Treat "I do not have that to hand" as a report on your knowledge base rather than a failure of the assistant, and each one tells you which document to write next.

Shops with a door as well as a website

The word retail gets used as if it meant online retail. It does not. In the second quarter of 2026, US retail e-commerce was $340.2 billion of $1,986.5 billion in total retail sales, which is 17.1% (Source: US Census Bureau, 2026). More than eight in every ten retail dollars are still spent somewhere other than a website.

That changes what the assistant on your site is for. If you have a shop with a door, a large share of the questions it receives are not about buying online at all:

1. What time do you close today, and are you open on bank holidays.
2. Do you have this in the Leeds branch.
3. Can I collect an online order in store, and how long until it is ready.
4. Is there parking, and is the entrance step-free.
5. Can I return an online order to the shop.
6. Do you do repairs, alterations or fittings, and do I need an appointment.

Every one of those is answerable from a single short document about the shop itself, and every one of them is a question that currently rings your phone during trading hours. Opening hours are the highest value thing on that list and the most commonly out of date, so keep them in a curated answer you can edit in seconds rather than in a PDF you have to re-export.

What one costs to run

The measured cost of serving an answer is small. Across a real support corpus, an answer averages 1,794 input tokens and 87 output tokens, which at Anthropic's Claude Sonnet rates works out at roughly $0.0045, under half a cent. Questions that match a curated answer closely enough never reach the model and cost nothing at all.

That is the underlying cost rather than the price. What you pay is a plan:

For a single shop, 100 answers a month on the Free plan is enough to find out whether shoppers use it before spending anything. Starter removes the chatfor.site badge and adds email support. Growth adds conversation history and search, which is what you want once you are reading transcripts to decide which document to write next, plus the option to place the chat inside a page rather than only in a corner bubble. A multi-branch retailer running an assistant per brand or per region wants Growth or Business for the assistant count.

Visitors are limited to 30 messages a minute each, enforced centrally rather than per browser, so one bored person cannot burn a month of answers.

If you want to see the shape of these conversations before writing anything, the worked examples of real question and answer exchanges are more use than a feature list. And if you already have someone answering chat manually and are trying to work out whether to automate it, the comparison of automated answers against staffed live chat is the decision to make first.

Frequently asked questions

Can a retail chatbot check stock levels?

Not from documents. An assistant answering from uploaded material has no view of your stock system, and uploading a stock spreadsheet means it will quote stale counts with confidence. Write a curated answer that points shoppers to the availability shown on the product page instead, and keep stock counts out of your documents entirely.

What should I upload first for a retail shop?

Five short documents: delivery, returns, sizing, payment, and the shop itself with its opening hours and address. Then add a product entry for each line you sell. That covers the majority of pre-purchase questions. Add more documents only when a real conversation shows a gap.

Do chatbots for retail work for a physical shop rather than an online one?

Yes, and often better. US retail e-commerce was 17.1% of total retail sales in the second quarter of 2026, so most retail questions are about opening hours, stock in a named branch, click and collect, parking and access. All of those are answerable from one short document about the shop.

Will it recommend products to shoppers?

It shows matching products as cards when the question is genuinely about a product, up to four at a time, drawn from the product entries you have added. Questions about delivery, returns or opening hours show no cards. It does not build a profile of the shopper or recommend from browsing history.

How long does it take to set up?

An afternoon for the documents and about five minutes for the install, which is one script tag on your site. The documents are the real work, and they are worth doing properly: the quality of the answers is almost entirely a function of the quality of what you upload.

Sources

• Baymard Institute, "49 Cart Abandonment Rate Statistics", last updated 22 September 2025. https://baymard.com/lists/cart-abandonment-rate
• Nielsen Norman Group, Jakob Nielsen, "How Little Do Users Read?", 5 May 2008. https://www.nngroup.com/articles/how-little-do-users-read/
• US Census Bureau, "Quarterly Retail E-Commerce Sales, 2nd Quarter 2026", released 18 August 2026. https://www.census.gov/retail/mrts/www/data/pdf/ec_current.pdf