What Your Customers Ask Before They Buy
Two prompts that rank the questions in your customer chats and draft the answer your homepage is missing.
This guide is for people who run an online store or service and have a chatbot on their website. You export 10 to 20 recent chats as text files, upload them to Claude, and run two prompts. The first ranks what customers ask before buying, counts how many different chats each point appears in, and quotes the customer word for word. The second finds the most common question your homepage does not answer, then drafts a short answer and three headlines in your customers' own words. There is nothing to install: you only need a Claude account that lets you upload files. Claude can still misquote or miscount, so you check the quotes against your files. Your chat text goes to Anthropic. On Free, Pro and Max plans, check the Model Improvement setting first, and remove personal details before you upload.
The question every visitor has to ask
Say you run a tea business. Customers chat with the chatbot on your website before ordering, but nobody counts what they keep asking. Each question is a doubt that can stop an order, and you never see which one your homepage misses.
Customer chats are market research that is already written. Claude can read a batch of them and count. This guide gives you two prompts. The first ranks the questions customers ask before buying, with each customer's exact words beside it. The second finds the most common question your homepage does not answer and drafts the answer. In the tea example in the reel, the top question is caffeine, and the homepage never mentions it.
Completion test: take the top row of your table. You can open the file it names and find the customer's exact message, and your homepage does not answer it. If you cannot do both, the result is not ready to use yet.
What you need
There is nothing to set up. You use a normal Claude chat.
| Need | What it means |
|---|---|
| A Claude account | One that lets you upload files |
| Chat exports | 10 to 20 recent conversations from your website chatbot, as text files |
| Your homepage text | Copied into a text file, or ready to paste |
| One decision-maker | The person who approves the wording that goes on the page |
What Claude accepts. Anthropic's upload page lists up to 20 files per chat and up to 500 MB per file. It accepts TXT, CSV and JSON files, as well as PDF, DOCX and others. It also says extra token limits can apply when the text in a file is very long, so keep each chat short and your batch modest.
What it costs. Nothing beyond your Claude plan. There is no connector, no install and no code.
Get your chats ready
A few minutes of preparation decides how much you can trust the result.
- Export recent chats as text files. Most chatbot tools let you export conversations. The menu differs by tool, so look in yours.
- Choose 10 to 20 chats from a normal stretch, such as the last two to four weeks. Include chats that did not end in an order.
- Remove personal details. Replace names, email addresses, phone numbers, street addresses and order numbers with a placeholder such as [name]. Keep the customer's wording and the date.
- Name each file with a number and the date, for example chat-07_2026-10-12.txt. The prompt asks Claude to cite the file name and date for every quote, so these names are what you check later.
- Keep one chat per file if you can. If you have more than 20 chats, put several in one file and start each with a line such as Chat 07, 2026-10-12. The prompt counts different chats, so each one must be clearly separated.
- Put your homepage text in a separate file, for example homepage.txt, or keep it ready to paste for the second prompt.
Checkpoint. You have no more than 20 files, every file name carries a date, and no file contains a customer's name or contact details.
Prompt 1: rank what customers ask
Open a new Claude chat, upload your chat files, and paste this prompt as one message. Use the wording exactly as written the first time you run it.
You are a customer research analyst. I'm giving you chat files from the chatbot on my website. The people chatting are my customers. Use only what is in the chats. Read every file in full. Then give me four sections: 1. Questions customers ask before buying. 2. The exact words customers use to describe their problem. 3. What stops them buying (doubts, price, trust, delivery). 4. Requests the chatbot couldn't meet. For every finding: give it a short label; count how many different chats it appears in (a customer repeating it in one chat counts once); quote the customer's message word for word, with the file name and date; rank findings from most chats to fewest; mark any finding that appears in only one chat as "one-off". Rules: quote only customer messages, never the chatbot's; never paraphrase or invent a quote; if you can't quote it, leave it out; ignore thank-you messages unless they show a problem or a request; show the result as a table: Finding | Chats | Quote | File and date.The rest of this playbook is below.
Enter your email to read it. You also get one new playbook every week.
One email a week. Unsubscribe any time.
What you get. Four sections: the questions customers ask before buying, the exact words they use to describe their problem, what stops them buying, and the requests the chatbot could not meet. Each section is a table with one row per finding, in this shape:
| Finding | Chats | Quote | File and date |
|---|---|---|---|
| [short label] | [number of different chats] | "[the customer's exact message]" | [file name, date] |
Rows run from the finding that appears in the most chats to the fewest. A finding that appears in only one chat is marked one-off.
Why it is written this way. It gives Claude one role and limits it to your chats. It defines the count, so one chatty customer cannot inflate a row. It demands a word-for-word customer quote with a file name and date, so you can check every row. And the rule "if you can't quote it, leave it out" is what keeps invented findings off the list.
Check the result before you trust it
Claude can misquote or miscount, and the prompt is built to make that easy to catch. Spend ten minutes here before you use any number.
| Check | How |
|---|---|
| The quotes are real | Pick five rows. Open the file each one names and search for the first few words of the quote. The wording must match exactly |
| The quotes are the customer's | Each quote must come from the customer, not from the chatbot |
| The top counts hold | For the top three findings, search your files for the key word and count the different chats yourself. Expect small differences where customers used other words |
| One-offs are marked | A finding that appears in a single chat says one-off |
| File and date match | The file name and date beside each quote are the ones in your files |
Pass: every check holds for the rows you looked at.
Stop: a quote you cannot find, a quote from the chatbot, a count that is off by more than a few chats, or a row with no quote. Ask Claude to remove every row whose quote it cannot show word for word in the files, and to recount. Then check again.
An optional extra check. This prompt is ours, not part of the two prompts above:
Check your own table. For each row, find the quoted message in the files and confirm it is word for word and from the customer. List every row where the quote is not an exact match, and recount the chats for the top five findings.This is the same model checking its own work, so it helps but does not replace opening the files yourself.
Prompt 2: find the question your homepage does not answer
Stay in the same chat. The second prompt refers to sections 1 and 2 of the first answer, so it only works after the first prompt, in the same conversation. Upload your homepage text or paste it in, then send:
Now read my homepage text. Which of the questions in section 1 does it not answer? Take the most common one. Write a short answer for the page using my customers' own words, and give me 3 headline options built from the phrases in section 2. Keep the customer quote beside each line so I can check it.What you get. The most common question from section 1 that your homepage does not answer, a short answer for the page written in your customers' own words, and three headline options built from the phrases in section 2. The customer quote sits beside each line.
Before you use anything. Claude drafts and you approve. A draft answer states facts about your product, such as what is in it or how long delivery takes. Check each fact against your own product information. In the tea example, check the caffeine content of the actual tea before the answer goes on the page.
If your homepage text is inside an image, Claude cannot read it. Anthropic's upload page says that for files other than PDFs, Claude extracts the text only. Copy the text out first.
Put the answer on the page and measure it
- Edit the draft. Fix any fact and adjust the tone so it sounds like your brand.
- Put it where the visitor decides. Near the top of the homepage, or on the product page. Try one of the three headlines.
- Note the date you published it.
- A month later, export new chats and run Prompt 1 again. Compare how many chats still ask the same question. If the page is doing its job, fewer should. This is a signal, not proof, because other things change too.
Keep your first table so you have something to compare against.
If you will repeat this each month, you can keep the exports in a Claude Project. Anthropic's upload page says project files stay available across conversations, with a limit of 30 MB per file and text extraction only.
What stays private
- Claude is a hosted service. Your chat text and the files you upload go to Anthropic.
- Check the Model Improvement setting first. Anthropic's privacy page for Free, Pro and Max says it uses chats to improve its models if you choose to allow it. It may also use conversations flagged for safety review, or ones you explicitly opt in to training. Your setting is in your Privacy Settings.
- Incognito chats are different. The same page says Incognito chats are not used to improve Claude, even when Model Improvement is on.
- The thumbs-up and thumbs-down buttons share the whole conversation. The page says feedback stores the entire conversation for up to 5 years and may be used to train models. Do not click them on a chat that holds customer data.
- Business plans have their own rules. The page covers Free, Pro and Max. Claude for Work and the API are covered by a separate page.
- Strip personal details first, as in section 3, and check your chatbot's privacy notice before you send customers' messages to a third party. This is not legal advice.
Starter checklist
Today, no setup
Prepare
Run
Publish and measure
Troubleshooting
| What you see | What to do |
|---|---|
| Claude will not take all the files | Anthropic lists up to 20 files per chat. Combine several chats into one file, each starting with a line that gives the chat number and date |
| The result looks cut short or skips chats | Extra token limits can apply to long text. Run the first prompt on smaller batches, then compare the tables yourself |
| Claude summarises instead of quoting | Reply: Quote only customer messages, word for word. If you cannot quote it, leave it out. Then ask for the table again |
| A quote comes from the chatbot | Ask Claude to remove every row whose quote is a chatbot message, then re-rank the rest |
| A count looks too high | Ask Claude to list the chat files behind the count for the top five findings, then check them |
| Prompt 2 says the homepage answers everything | Check that Claude received the homepage as real text, not an image. Paste it into the chat |
| The tables are very long | Ask for only the top 15 findings in each section |
| Results differ between runs | Counts can shift a little between runs on the same files. Rely on the rows you have checked against the files |
Honest limits
- It reads a sample. 10 to 20 chats do not show everything customers ask. A question missing from the sample may still exist, and counts from a small sample move around a lot.
- It only sees people who chat. Customers who bought without chatting, or left without asking, are not in the data.
- Claude can misquote or miscount. That is why section 5 exists. Check the quotes against your files.
- A better page does not guarantee more sales. Answering one question removes one doubt. Whether orders rise depends on much else.
- Your customers' words are their data. Remove personal details and check your own privacy notice before you upload.
- The tea example is an illustration. The caffeine result comes from the reel's example, and is not a benchmark for your store.
How this guide was made. This guide is built from Anthropic's upload help page (updated July 23, 2026) and its privacy page on model training (updated March 16, 2026). Sources were checked on November 5, 2026. No chats were run for this guide, and the two prompts are used exactly as supplied. Settings, limits and screens may change, so follow the linked pages where they differ.
Links
- Upload files to Claude (Anthropic Help Center)
- Is my data used for model training? (Free, Pro and Max)
- Is my data used for model training? (Claude for Work and the API)
- Your Privacy Settings in Claude
What this changes
Every question a customer has to ask your chatbot is one your page could have answered first. Answering it once on the page saves every visitor from having to ask.
Customer chats are market research that is already written. Start with 10 chats and one question.
