Ecommerce Chatbot: How to Choose One for a Small Shop

An ecommerce chatbot is a chat window on a shop's website that answers customer questions automatically, usually about delivery, returns, sizing and stock. The useful ones answer from the shop's own written material rather than from general knowledge, which is what keeps them from inventing a returns policy you never had.

This guide covers what the three kinds actually are, what one costs to run per answered question, the failure mode nobody selling you one mentions, and how to install one on the platform you already use.

What is an ecommerce chatbot?

An ecommerce chatbot is software that holds a text conversation with a shop's visitors in order to answer their questions or guide them to a product. Modern ones use a language model to understand the question and compose a reply; older ones follow a decision tree of buttons the shop owner drew in advance.

That distinction matters more than any feature list, so it is worth being blunt about the three things sold under the same name.

The third kind is what most people now mean by an conversational AI in a retail setting, and it is the only one of the three whose failure mode is honest.

Worth knowing: the technical name for the third approach is retrieval-augmented generation. The assistant searches your documents for passages relevant to the question, then the model composes an answer using only those passages. Nothing about the shop is baked into the model itself.

Why put one on a shop at all?

Because the answer to most pre-purchase questions is already written down somewhere on the site, and the customer will not go looking for it.

The evidence on this is unusually good. The average documented cart abandonment rate across 50 studies is 70.22% (Source: Baymard Institute). When shoppers are asked why, excluding those who were simply browsing, the reasons at the top are:

1. Extra costs too high, including shipping, tax and fees, at 40%
2. Delivery was too slow, at 20%
3. Did not trust the site with card information, at 19%
4. The site wanted them to create an account, at 18%
5. Checkout too long or complicated, at 17%
6. Website errors or crashes, at 17%
7. Returns policy unsatisfactory, at 13%
8. Unable to view or calculate the total order cost upfront, at 12%

(Source: Baymard Institute, page last updated 22 September 2025)

Now be honest about which of those a chatbot touches. It does not make your shipping cheaper or your delivery faster. If a customer leaves because next-day delivery costs $12, an assistant that tells them so faster has not saved the sale.

The two that an assistant genuinely speaks to are the last two on that list, and they are worth separating. A shopper who finds your returns policy unsatisfactory has read it and decided against it; no assistant fixes that. A shopper who cannot find or cannot work out the returns terms or the total cost is a different person entirely, and that person is answerable. At 13% and 12% respectively, each is a meaningful slice on its own, and part of each is a findability problem rather than a policy problem.

Do not add those percentages together. Baymard's respondents could select more than one reason, so the figures overlap and a total is meaningless. Any vendor who sums them to claim a bigger number is doing arithmetic the survey does not support.

That is a real problem and a smaller claim than the one most vendors make. It is also the claim you can actually keep.

For scale, US retail e-commerce was $340.2 billion in the second quarter of 2026, 17.1% of all retail sales (Source: US Census Bureau). The questions are not going to stop arriving.

What does an ecommerce chatbot cost?

Per answered question, far less than most pricing pages imply. Here is the real arithmetic, using chatfor.site's own measured figures rather than an estimate.

An average answer on chatfor.site consumes 1,794 input tokens and 87 output tokens. At the model's published rates of $2.00 per million input tokens and $10.00 per million output tokens, that works out as:

Under half a cent. A thousand answered questions is $4.46 of model spend.

Two things follow. First, any vendor charging you per conversation at a dollar a head is selling margin, not compute. Second, the interesting cost in a chatbot is never the model, it is the fixed monthly hosting underneath and the labour you spend maintaining the documents.

Work out your own number the same way, in the full breakdown of what a chatbot costs to run.

Worth knowing: ask any vendor what a single answer costs them to serve. The ones who cannot tell you are usually reselling someone else's API with a markup they would rather not itemise.

The failure mode nobody sells you

A language model with no restrictions will answer a question about your returns policy whether or not it has ever seen your returns policy. It will produce something plausible, in your brand's voice, that you never agreed to. The customer has no way to tell.

This is the single most important thing to test before you put a chatbot on a shop, and it is easy to test. Ask it something it cannot possibly know.

• "Who is your chief executive?"
• "Do you ship to Norway?" when you have never written down where you ship
• "Is this cotton organic?" when the product page does not say

A well-built assistant says it does not have that to hand. A badly built one invents a name, a country list, or a fabric claim. The second kind is a liability rather than a support tool, because a fabricated delivery promise is still a promise your customer heard from your website.

The design that avoids this is restriction: give the assistant only your material, and instruct it to refuse rather than reach. That is what chatfor.site does, and it is deliberately the least impressive-looking behaviour in the product. See it happening in worked examples including one where the assistant refuses.

Two further protections worth insisting on:

• Show the source. For any answer, you should be able to open the conversation later and see which passage of which document the answer came from. Without that you cannot audit a complaint.
• Let yourself overrule. When an answer is wrong, you need a way to correct it that does not involve re-uploading a 40-page PDF.

What do you feed it?

Whatever you already have in writing, and the quality of that material sets the ceiling on the assistant's answers. This is the part shop owners underestimate.

Good sources, roughly in order of value:

1. Returns and refunds policy, in full
2. Delivery information, ideally as a table of destinations, costs and times
3. Sizing and fit guidance, per product category
4. Care and materials information
5. The questions your inbox actually receives, written as short answers
6. Product descriptions, including the details buyers ask about

A single 40-page PDF answers noticeably worse than eight short focused pages, because retrieval works on passages: the tighter each document's subject, the more precisely the right passage can be found. Tables of shipping rates beat paragraphs about shipping for the same reason.

The full method is in how to write documents an assistant can answer from.

Chatbot or a person?

Both, and the split is not subtle. The question is whether an answer already exists in writing.

An assistant handles the repeated, documented and boring. A human handles the specific, the emotional and the exceptional. A shop that routes the second category into a bot is not saving money, it is losing customers more cheaply.

The trade-offs in detail, including when live chat is the better buy, are in the comparison between an assistant and staffed live chat. Sector-specific patterns for shops with physical stores are in chatbots in a retail setting.

Installing one on what you already run

Modern assistants install as a single script tag, the same shape as adding Google Analytics. You paste one line into your theme and the chat bubble appears. No plugin, no rebuild, no developer.

The exact menu path differs per platform:

• installing on Shopify
• installing on WordPress and WooCommerce
• installing on Squarespace
• installing on BigCommerce

If you run something else, or hand-built HTML, the platform-agnostic version is adding a chatbot to any website.

How to choose one: the checklist

Run every candidate through this before you pay anything. Most of it can be tested on a free plan in twenty minutes.

• Ask it something it cannot know. Does it refuse, or invent?
• Can you see which document each answer came from?
• Can you correct a wrong answer without re-uploading everything?
• What is the per-answer allowance, and what happens at the limit? A bot that silently stops answering mid-month is worse than no bot.
• Can you restrict it to your own domains, so nobody embeds it elsewhere and spends your allowance?
• Where is the data stored, and what is retained?
• How long is the first token? A four-second wait reads as broken.
• What does it cost per answered question, worked out rather than guessed?

For how the current tools measure against that list, see the ranked comparison for ecommerce shops. If you are already paying for a helpdesk and wondering whether to switch, the specific cases are covered in alternatives to Intercom, alternatives to Tidio and how Chatbase compares.

Frequently asked questions

What are ecommerce chatbots?

Ecommerce chatbots are automated chat assistants placed on an online shop to answer customer questions about delivery, returns, sizing and stock. The current generation uses a language model restricted to the shop's own documents, so answers reflect the shop's actual policies rather than general knowledge.

How does a chatbot work in ecommerce?

A visitor types a question. The system searches the shop's uploaded documents for relevant passages, then a language model writes an answer using only those passages. Because the source material is restricted, the assistant can be instructed to say it does not know when the documents do not cover the question.

Should I add a chatbot to my ecommerce site?

If you answer the same handful of questions every day and the answers already exist in writing, yes. If your questions are mostly order-specific or emotional complaints, a chatbot will not help and may harm the experience. Test it on a free plan before paying.

How do I integrate an AI chatbot into my ecommerce store?

For most platforms you paste a single script tag into your theme, which takes a few minutes and needs no developer. Shopify, WordPress, Squarespace and BigCommerce each have a different menu path for custom code, covered in the platform guides linked above.

How much does an ecommerce chatbot cost?

Model costs are under half a cent per answered question, around $4.46 per thousand answers at current rates. What you pay a vendor is mostly hosting and margin: plans for small shops typically run from free through to a few hundred dollars a month depending on answer volume.

Will a chatbot make things up about my shop?

An unrestricted one can. One restricted to your uploaded documents and instructed to refuse rather than guess should not, but you should verify this yourself by asking it something your documents do not cover before you trust it with customers.

Sources

• Baymard Institute, cart abandonment rate review across 50 studies: https://baymard.com/lists/cart-abandonment-rate
• US Census Bureau, Quarterly Retail E-Commerce Sales, Q2 2026: https://www.census.gov/retail/ecommerce.html
• Per-answer token and cost figures: chatfor.site's own measured averages, recorded in lib/plans.ts as MEASURED_TOKENS_PER_ANSWER and MODEL_PRICE_CENTS_PER_MTOK