What a Website Chatbot Actually Is, and How It Went From Menus to Real Conversation
You have almost certainly used one already. The small chat window in the bottom-right corner of a store’s website, the one that slides up with “Hi, can I help you find anything?” while you browse. That is a chatbot.
Behind that little window there can be a few different things. Sometimes a real person is on the other end, a support agent typing back to you. More often, especially after hours, it is automated. For years that automation was simple: a short menu of preset buttons (“Track my order,” “Returns,” “Sizing”) with canned replies, and a way to reach a human when none of them fit. Useful, but closer to a phone menu than a conversation.
What changed recently is what sits behind the window. Today it can be a genuine AI that reads your question the way you actually wrote it and answers the way a real assistant would: finding the dress, checking the size, telling you when it ships. And you can still ask to be handed to a person whenever you want one. Same little window, a very different thing behind it.
This article is about that window on your own store: what it is, how it went from button menus to real conversation, and why the tool matters far less than how you train it.
Stage 1: the menu with a chat skin
The ancestor is ELIZA, built at MIT in 1966. It worked by pattern matching: spot a keyword in the user’s sentence, return a prepared reply built around it. The mechanism is not conversation. It is a lookup table with good manners.
The first generation of store chatbots followed the same logic, in two common forms:
- Preset questions and buttons. The shopper picks from a list (“Track my order,” “Returns,” “Sizing”), and the bot returns a scripted answer. Under the hood this is a decision tree: each button leads to the next fixed node.
- Intent bots. The next step up. Instead of exact keywords, the bot classifies what the shopper seems to want (“where is my order”) and fires the matching answer. To work, it needs a labeled training set: someone has to collect example phrasings and tag each one with its intent. The answers are still fixed.
Both versions share one weakness that matters for sales. They only work while the shopper stays on the script. Ask about a dress “for a wedding in October, not too formal, I’m between sizes” and the bot has no node for it. The result is a canned apology or a handoff to a human, and the shopper, who came with a buying question, leaves. The mechanism of the loss is simple: a sales conversation is made of off-script questions, and a script cannot hold them.
Stage 2: AI inside
Today’s on-site chatbot has real AI in the middle. That changes three things.
- It understands ordinary language. The shopper types the wedding-dress question as written, with the hedging and the missing details, and the bot parses it instead of looking for a keyword.
- It can work across the whole catalog and inventory. Connected to your product data, it filters by what exists and what is in stock, not by what a script author predicted someone would ask.
- It answers the actual question. Product details, fit, shipping, returns, in a few sentences rather than a link to a policy page. It runs 24 hours a day, which matters for any store with customers in more than 1 time zone.
The practical shift: Stage 1 was a menu. Stage 2 is a conversation, and conversation is where hesitant shoppers get unstuck.
One piece of background, kept short on purpose: shoppers increasingly discover products inside AI tools such as ChatGPT, but the purchase itself still lands on the brand’s own site. That makes the chat on your site the last meter before the sale, which is the reason it deserves attention now.
It is already running at large fashion and beauty brands
Several large brands run customer-facing chat today. Treat the numbers below as reported by the companies or their vendors, not as audited results.
- Zalando built a ChatGPT-powered shopping assistant in a 5-week sprint. Per vendor-sourced data, it has been live across all 25 markets since October 2024 and reached about 10 million users in Q1 2026.
- Frasers runs “Ask Frasers” on its site and has reported a 25% conversion lift versus standard site search.
- Kering has a ChatGPT-based assistant called “Madeline” (reported by masterofcode.com).
- Fenty Beauty offers a beauty advisor, “Rose Amber,” inside WhatsApp, which shows the same idea works in a messenger channel as well as on a website.
- DRESSX launched the DRESSX Agent for AI try-on and shopping in September 2025.
- Ulta runs “Ulta AI,” built on Gemini, for beauty shoppers.
A FashionNetwork survey reported that roughly 1 in 3 brands (33%) have adopted AI agents for shopping, with fashion leading. It is a reported figure, and the survey’s definition of “AI agent” is broad, so read it as direction, not a benchmark.
The market map: classes of tools, not a ranking
Search for “best AI chatbot for ecommerce” and you will find lists written by vendors, which contradict each other. There is no honest top 1. What exists is a set of classes, each built for a different job. Prices below differ across sources, so check them with the vendor before you budget.
| Class | Tools named in the sources | What it is for |
|---|---|---|
| Platform-native | Shopify Inbox, Sidekick | Free with the platform; sources describe it as less capable than specialized tools such as Fin or Lyro |
| Support and helpdesk | Gorgias, Intercom Fin, Zendesk AI | Support volume and tickets. Gorgias is most often recommended for Shopify (reported ~$0.90 to $1.00 per automated reply on top of the plan). Fin is mature, priced per resolution (~$0.99), and not ecommerce-native, so it is weaker on Shopify order actions. Zendesk AI is enterprise (from ~$55 per agent per month) and fits teams already on Zendesk |
| Small store, on-site widget | Tidio (Lyro engine) | Budget entry; free live-chat plan exists. Lyro is metered separately, so the real bill runs above the base price |
| Conversion and on-site selling | Rep AI, Zipchat | Built to sell on the site, not to deflect tickets |
| Quiz and product matching (DTC) | Octane AI | Guided product finder |
| Build your own | Voiceflow | You design the bot and choose the model; for stores that outgrew templates |
| Enterprise | Ada, Yellow.ai, Salesforce Agentforce | Large-scale deployments; pricing is often not published |
The decision follows the job. If your pain is support load, start in the support class. If the pain is shoppers who browse and leave, look at the conversion and matching classes. If you cannot say which one it is, that is the first thing to settle, before any demo.
The tool is commodity. The training is the work.
Every tool in that table can hold a conversation. That is exactly why the tool alone is not an advantage: your competitor can license the same one this afternoon. What differs is what the bot knows when a shopper asks.
An AI chatbot is only as good as the data it is connected to and the rules around it:
- Catalog and inventory. If stock data lags, the bot recommends items that are gone. Sync frequency becomes a sales issue.
- Policies. Returns, shipping times, exchange conditions. A bot that guesses a policy is making a promise on your behalf.
- Brand tone. Your boutique voice is not a default setting. If it is not specified, the bot will sound like every other bot.
- Connections. Where the bot can look things up (orders, sizing data) decides whether it answers or merely apologizes.
A poorly trained bot does not fail quietly. It tells a shopper that a jacket runs large when it does not, or states a return window that does not exist, and it does so on your storefront, in front of the customer, with your name on it. In a support inbox, a wrong answer is caught and fixed before anyone sees it. A chatbot has no such buffer, so its errors are the costliest ones in the stack. (Testing a bot before launch with 50 awkward real customer questions sounds tedious. It is also the cheapest insurance on this list.)
A working rule for choosing and running one
Name the job (support, conversion, or product matching), pick the class that serves it, and then spend most of your effort on the data: catalog, stock, policies, tone. Test it against real questions before it goes live, and re-check answers after every policy or catalog change. The bot is the easy part; the training is where results come from.