How Shoppers Type in Chat: Fragments, Not Questions (51 Characters on Average)

TL;DR: People do not ask chatbots questions. They type fragments. Across 5,470 real website chat conversations analyzed by Chatonbo in 2026, the average opening message was 51 characters and only 35.7% contained a question mark. 58.7% of first messages were about a specific product or service, 12.7% were just "hi" or a test, and 5.8% were about an order. An assistant trained to expect polite full sentences will stumble on "blue one in medium". One trained on your product pages, given the current page as context and told to infer the missing words, will not.

What does a typical opening message look like?

We aggregated the first visitor message from every conversation held on 246 live assistants between 13 June and 11 September 2026. The numbers, method and the rest of the dataset are on the What shoppers ask an AI chat assistant study page. Two facts stand out before you even look at topics:

• The average opening message is 51 characters. That is shorter than this sentence.
• Only 35.7% of opening messages contain a question mark.

People type into a chat box the way they type into a search box. The study's own examples are "shipping to canada", "blue one in medium" and "track order 1042". No greeting, no verb, no question mark, and a very clear intent if you already know what the visitor is looking at.

What are the first messages about?

Keyword classification of the opening message, first match wins:

Nearly six in ten first messages are about a specific product or service. Add recommendations, price, size and stock and you are at almost three quarters of all openers being some form of "tell me about this thing". Only 2.3% start by asking for a human or contact details.

Why do people type fragments?

Three reasons, all visible in the data.

They are on a phone. 47.7% of conversations came from a phone and 44.8% from a desktop. Nobody writes a paragraph with their thumb.

They are already on the page. 20.4% of conversations started on a product page and a further 4.7% on a collection page. The visitor is looking at the blue jacket. From their point of view "in medium?" is a complete question, because the product is right there on the screen.

They have been trained by search engines. Twenty years of typing "hotel tel aviv cheap" into a search box carries over. The chat box is a search box that talks back.

What does this mean for training an assistant?

If the input is a fragment, the assistant needs two things a generic chatbot does not have: facts and context.

1. Ground it in your product and service pages, not a script. 58.7% of openers are about a specific item. A decision tree with "Sales / Support / Other" buttons cannot answer "does the 40L come in green". An assistant that has read your catalog can. On a store this is the entire game: 46.9% of all Shopify conversations in the study were product questions.

2. Give it the current page. "How much is this" is unanswerable in isolation and trivial with the page URL. Good assistants receive the page the visitor is on and resolve "this", "it" and "the blue one" against it. If your widget does not pass page context, a quarter of your conversations start with a guess.

3. Tell it to infer, then answer. The instruction that changes the most in practice is: "Visitors write short fragments. Work out the most likely full question from the fragment and the page they are on, and answer it directly. Only ask a clarifying question if two products or options genuinely match." Without it, models tend to reply "Could you tell me more about what you are looking for?", which is exactly the friction that made the visitor type a fragment in the first place.

4. Handle "hi" with a prompt, not a lecture. 12.7% of first messages are a greeting or a test. The right reply is one line: "Hi! Ask me about any product, shipping, or your order." Not three paragraphs about the company.

5. Map the small categories to real data. Size and fit (3.1%), stock (1.2%) and order status (5.8%) are fragments that need live or structured data: a size chart the assistant can quote, real inventory, and an order lookup. A Shopify AI agent connected to the store handles "track order 1042" by reading the order, not by guessing.

Do fragment openers turn into conversations?

Often. 53.7% of conversations were a single question and answer, but 46.3% went further: 27.7% ran to two or three exchanges, 13.0% to four to six, and 5.6% to seven or more. The average conversation had 4.8 messages. On general business sites 50.5% of conversations were multi-turn, against 35.8% on Shopify, where the shopper gets the answer and returns to the product page.

This matters for design. The first fragment is often a probe. If "shipping to canada" gets a precise answer with the price and the delivery time, the next message is frequently "and the blue one in medium?" or an email address. If it gets "please contact our support team", the conversation ends there, and so does the sale.

How to test your assistant with fragments

Before you trust an assistant with real traffic, throw the study's categories at it in the shape visitors actually use:

• "shipping to canada"
• "blue one in medium"
• "cheaper than the pro?"
• "track order 1042"
• "which one for a 2 year old"
• "discount code"
• "hi"

A good assistant answers each in one or two sentences with a real fact from your site or store, and asks a clarifying question only when it genuinely has to. You can run this test on your own pages without an account on the live demo: paste your URL, then type fragments, not questions.

The visitors will not change how they type. The assistant has to meet them there.