Most stores asking this should buy the app
We went looking for the firms that answer this question, by asking it the way a store owner would: who can add an AI chatbot to my ecommerce store that knows my products.
Every result was a product. Apps, tools and comparison posts about apps. Not one development firm appeared on the first page. We ran that search on 22 October 2026 and again on 30 October, and both times it came back the same way. That is our own measurement, on two dates, using the words a buyer would use, and you can repeat it in a minute.
The honest reading of that is not that the market has missed something. It is that the market is mostly right. For a store selling a stable catalogue, where customers ask about delivery, returns, sizing and whether something is in stock, an app is the correct answer. Entry plans sit around $10 to $30 a month, several have free tiers, and the work is an afternoon. We would rather tell you that than sell you a build you do not need.
So this page is about the other case: how to tell when you are in it.
What an app is genuinely good at
Be fair to the thing before you replace it. An off-the-shelf assistant reads your published pages and answers from them, and it does that well:
- Policy questions. Shipping times, returns windows, warranty terms, what happens if something arrives damaged.
- Pre-sales questions with published answers. Materials, dimensions, care instructions, what is in the box.
- Deflection. A real share of messages are questions your help pages already answer, and every one handled is one your team does not open.
- Out of hours. Most of these conversations happen when nobody is at a desk.
If that describes the bulk of your inbox, stop reading and go and install one.
The four places an app stops working
Each of these has a symptom you will recognise, and none of them is fixed by paying for a higher tier.
1. The answer depends on live data, not on a page. Many tools read your catalogue on a schedule, commonly overnight. The assistant then answers from a copy. For a stable catalogue nobody notices. If your stock moves during the day, or your prices change, or items are made to order, the assistant will confidently describe yesterday. The question to ask any vendor is whether it reads stock and price at the moment of answering or from a periodic copy.
2. The question is about the customer, not about the catalogue. "Where is my order", "can I change the delivery address", "has my refund gone through". These need the assistant to look something up against an identified customer. An app that only reads public pages cannot do it, and the conversation dies at exactly the point where it was about to be useful.
3. The assistant needs to do something, not say something. Starting a return, applying a credit, changing an address, booking a fitting. Answering is a different job from acting, and acting means a connection to your systems and rules about what is allowed without a person.
4. You cannot see why it was wrong. This is the one that decides it for most stores. When an app gives a wrong answer, you get a transcript and no mechanism. You cannot see what it read, you cannot correct the source, and you cannot test that your correction worked. If a wrong answer costs you a return, a chargeback or a complaint, "we have emailed the vendor" is not a fix.
A useful way to sort yourself: if a wrong answer is embarrassing, an app is fine. If a wrong answer is expensive, it is not.
What a build costs, and what it costs every month after
Our prices are published, and the second number is the one almost nobody publishes:
| Work | What it covers | Fixed price |
|---|---|---|
| AI fit checkThe first step | Send the site and the task. A one-page plan with a price, in two working days | Free |
| Assistant on your own content | Answers from your pages, catalogue, policies and help articles, with the source shown under each answer | from $3,500 |
| With a store connection | Adds live lookups against your store: order status, stock, delivery dates | from $4,500 |
| Extra source or connection | A helpdesk, an ERP, a second content source, per connection | $600 to $2,000 |
| Running costIn your own account, not billed through us | Model usage. Hosting is usually under $20 a month on top | $40 to $200 a month |
| Care planOptional, cancel any month | Tuning, monitoring, changes, a named developer who knows your setup | $250 to $1,000 a month |
Two things worth saying plainly. The model fees are billed to your own provider account, so you can see them, cap them and take them with you. And we do not take jobs under $2,500, because below that the time spent scoping it costs more than the work returns for both sides.
Set against an app at $10 to $30 a month, a build has to earn its difference. It earns it when one of the four conditions above is costing you real money, and not before.
A test you can run this week, before you spend anything
This takes about an hour and it answers the question better than we can from the outside.
Write down twenty real questions
Not questions you imagine. Take them from your inbox, your chat transcripts or your returns reasons, in the customer's own words, including the badly worded ones.
Mark each one before you test it
Against each question write where the answer lives: a published page, your live store data, or a person's judgement. Do this first, so you are not grading on a curve afterwards.
Ask all twenty of whatever you have
Your current app, a free trial, or the chat built into your platform. Paste each question as written.
Count only answers that are correct and specific
A fluent paragraph that does not answer the question scores zero. So does a correct answer to a question you did not ask.
Read the result
Sixteen or more correct, keep the app. Below that, look at what the failures have in common: if they are mostly "live store data" or "needs an action", that is the case for a build. If they are mostly "needs judgement", the answer is neither, and the work is a better handover to a person.
Keep the sheet. If you do go on to a build, with us or anyone, those twenty rows are the specification, and they are worth more than a brief.
A note on what we can and cannot tell you
The price band for apps is a band on purpose. We checked vendor pricing in October 2026 and the comparison sites contradict each other, sometimes by a factor of three on the same product, and several tools charge per conversation or by catalogue size rather than a flat fee. Check the app's own pricing page, and look for the usage charges under the headline number.
The part we are confident about is the shape of the decision, not any vendor's price. We have built assistants that should have been apps and we have watched apps fail on exactly the four conditions above. If the test puts you on the app side of the line, that is the better outcome, and it costs you nothing to find out.
Questions people ask
We already pay for a chatbot app. Is that spend wasted if we build?
No, and it is the cheapest research you will do. The app has been collecting the real questions your customers ask, in their own words, which is the single most useful input to a build. Export the transcripts before you cancel anything. We read them to decide what the assistant has to handle, what it should hand to a person, and what is not worth automating at all. Several of the jobs you thought you needed usually turn out to be rare enough to leave alone, which makes the build smaller and cheaper.
How much traffic makes a build worth paying for?
Volume matters less than what the questions need. A store with 200 conversations a month where most answers depend on live stock, delivery dates or order status will get more from a build than a store with 5,000 conversations that are all about the returns policy. Policy questions are what an app is good at. The test at the end of this page measures the thing that matters, which is how many of your real questions can be answered from your published pages alone. If that number is high, stay on the app whatever your traffic is.
Can we start with an app and move to a build later?
Yes, and that is the order we recommend. The app costs little, proves whether your customers will use a chat window at all, and produces the transcripts that make a build accurate. The one thing to protect is your data: check you can export conversations before you sign up, because some tools make that awkward. Nothing in a build depends on the app, so there is no technical migration to do. You turn one off and the other on.
Our catalogue changes every day. Does that rule out an app?
It rules out some of them. Many tools read your catalogue on a schedule, often nightly, so the assistant is answering from yesterday's prices and stock. For a stable catalogue that is fine. If you run flash stock, frequent price changes or made-to-order items, a scheduled copy will tell customers something that was true last night, and you will find out when someone complains. Ask any app you are considering one question: does it read stock and price live at the moment of the answer, or from a copy?
What happens when the assistant does not know?
That is the behaviour to test before you buy anything, because the failure modes are very different. A good assistant says it is not sure and offers a person. A bad one invents a plausible answer, and on an ecommerce site a plausible wrong answer about compatibility or delivery becomes a return. Ask the tool three questions you know it cannot answer and watch what it does. If it guesses confidently on all three, the price is irrelevant.
Is a build harder to keep running than an app?
It is more yours, which cuts both ways. There is no vendor who can change your pricing, your limits or your features without asking. There is also nobody else to call, which is why every build we do ends with an optional care plan at $250 to $1,000 a month and why we hand over the accounts, the prompts and the code either way. The running cost is model usage, usually $40 to $200 a month, and it is billed to your own account rather than through us.
Will an assistant on our site help us show up in AI search?
No. They are unrelated jobs and it is worth separating them, because the two get sold together. An assistant answers people who are already on your site. Being named when someone asks ChatGPT or Gemini for a recommendation depends on what other sites say about you, not on what your own site can do. We wrote separately about how AI search chooses who to cite, and the short version is that a widget on your page has no effect on it.