# Using AI to analyze customer psychology: Reading fears and purchase barriers on your landing page

URL: https://opcaios.org/en/blog/ai-phan-tich-tam-ly-khach-hang-landing-page
Date: 2026-11-11
Author: Lê Quốc Nghĩa

# Using AI to Analyze Customer Psychology and Rewrite Landing Pages

Using AI to analyze customer psychology is using a language model like [Claude](/tien-ich/claude) to read a large volume of real text left behind by customers, then group them into categories of concerns and purchase barriers. The text includes reviews, comments, support emails, chat logs and survey responses through Google Forms, for example 200 customers answering in 2 weeks. The goal is to rewrite the landing page right where customers are hesitating. I need to be direct: the title uses the phrase "read minds", but AI cannot read anyone's thoughts. The model only summarizes and categorizes what customers have already said, so the result is a hypothesis that needs testing, not truth about each person.

This post makes no promise about how much conversion will increase and contains no actual figures. As of the date this source was read, 25/10/2026, the post uses no external research data. The post has 4 parts: (1) types of barriers; (2) the "four forces" framework at a conceptual level; (3) a six-step process from data collection to A/B testing; (4) a hypothetical example and ethical boundaries.

![Diagram of the process for analyzing purchase barriers using AI: collect data with consent, anonymize, group topics using AI, manually verify, map to landing page blocks, test A/B](https://opcaios.org/images/blog/ai-phan-tich-tam-ly-khach-hang-landing-page/so-do.webp)
*From what customers say to each block of the landing page. Diagram: OPCAIOS.*

## What can AI do and what can't it do?

AI is good at summarizing, grouping and naming topics from hundreds of comments, but doesn't know which group matters most to revenue. For example, with 200 customer responses, Claude can group them into 5 topics, pull representative quotes and suggest how to rewrite the headline or FAQ in just a few minutes. However, AI cannot guarantee it won't miss anything or misunderstand irony, doesn't know which option wins without testing, and only guesses when asked to infer the "hidden psychology" of one person.

Easy rule to remember: AI gives you a list of hypotheses with evidence, only real customers through experiments give you the answer. So ask AI to pull exact quotes from the data, read part of it yourself, then only bring hypotheses into the landing page. Finally, record the analysis date and data version so you can compare next time.

### Comparison table of four tasks

| Task | AI can do it | AI cannot do it |
|---|---|---|
| Summarize, group hundreds of comments | Yes, quickly | Know which group matters most to revenue |
| Name topics, pull representative quotes | Yes | Guarantee nothing is missed or nuance misunderstood |
| Suggest how to rewrite headline, FAQ | Yes | Know which option wins without testing |
| Infer the "hidden psychology" of one person | Only guess | Read actual thoughts |

## What are the five common purchase barrier groups?

Purchase barriers are reasons why customers are interested but haven't clicked buy. Five common groups are price, trust, risk, time and comparison. The price group sounds like "too expensive", so the landing page needs a pricing section, value comparison and clear statement of what's included. The trust group sounds like "who's behind this, is it real?", so it needs introduction of the person, real proof and clear contact. The risk group sounds like "what if I buy and it doesn't work?", so state the refund policy or free trial if you actually have one.

The time group asks "how long until I see results?", so you need a timeline and actual effort level per week, for example 2 hours per day for 30 days, calculated for 2026. The comparison group asks "how is this different from that option?", so you need an honest comparison table, including cases where you shouldn't buy. The list is just a starting point: your data might reveal other groups, such as concern about security or difficulty using Excel, Zalo or Facebook.

### Table of five barrier groups

| Barrier | Customers usually say | Landing page section that handles it |
|---|---|---|
| Price | "Too expensive", "don't know if it's worth it" | Pricing and value comparison section, break down costs, state what's included |
| Trust | "Who's behind this?", "is it real?" | Introduction of the person, real proof, clear contact |
| Risk | "What if I buy and it doesn't work?" | Refund policy/free trial if you actually have it, easy-to-read terms |
| Time | "How long until I see results?" | Timeline, actual effort level, hours per week |
| Comparison | "How is this different from that option?" | Honest comparison table, include cases where you shouldn't buy |

## What are the "Jobs to be done" Framework and the "Four Forces"?

Jobs to be done is a way of viewing purchasing as "hiring" a product to make progress in a specific situation, rather than just looking at demographics. In this approach to analyzing purchase decisions, people often talk about 4 forces. Push is the current discomfort. Pull is what's attractive about the new solution. Anxiety is the fear of switching, and habit is familiarity with the old way.

A good landing page usually increases pull and reduces anxiety. For example, someone who has used Excel and Google Sheets for 3 years might feel pushed by losing 10 hours a week on reconciliation, but held back by habit and anxiety. OPCAIOS only borrows this framework as a way to label when grouping. The details of each variation of the framework haven't been verified from original sources (checked on 25/10/2026), so you can use the framework as a suggestion rather than a standard.

## Six-Step Process

### Step 1: Collect data with consent
Useful sources: product reviews, comments under posts, pre-purchase customer emails, support chat logs, short surveys, frequently asked questions. Only use data you have permission to use for this purpose; if you collect personal information, you need to disclose the purpose and have a valid basis under personal data protection regulations (reference only, not legal advice; consult an expert when needed). Don't collect data by violating platform terms.

### Step 2: Anonymize before giving to AI
Delete names, emails, phone numbers, addresses, order codes. If using an external AI tool, read their data terms and see [5 data security rules for business when using AI](/blog/bao-mat-du-lieu-kinh-doanh-khi-dung-ai-5-quy-tac).

### Step 3: How to group topics using AI?

Grouping topics is the step of giving anonymized data to AI in batches and asking for 4 things. First, list repeated themes. Next, count how many times each theme appears. Then, include 2 to 3 verbatim quotes for each theme. Finally, note sentences that don't belong to any group. Asking for verbatim quotes helps you check whether the AI is making things up.

Suggested prompt: "Only use content from the data; if there isn't enough evidence, write 'insufficient data'." For example, you paste 50 comments per batch into Claude, and after 4 batches you have a themes table for 200 customers. Counting is only directional since AI can miscount, so cross-check with Excel or Google Sheets before using the count to make decisions.

### Step 4: Manual check
Read a few dozen random original lines and compare with the groups AI assigned. Check the counts, sarcasm, and rare but serious opinions (like fraud suspicions). With little data, the counts are only directional.

### Step 5: Map barriers to each block of the landing page

Mapping barriers is assigning each concern group to a landing page block that has the job of answering it. There are 5 blocks to consider. The headline answers what benefit customers get, for whom, so state specific results and don't oversell. Proof answers why to trust, so only use real reviews with consent and cite the source. FAQ answers the most common worries, so directly answer the top questions from the data.

The guarantee block answers what happens if it doesn't work, so only state what you will actually do. The CTA block answers what the next lightest step is, so use a clear button and describe what happens after clicking. Example assumption: if 60 out of 200 comments worry "not for beginners," the headline and FAQ need to say who the course is for. You can draft in Notion or Google Docs in one session, then review by hand.

### Step 6: A/B test with sufficient traffic

A/B testing is randomly splitting visitors into two landing page versions to compare a pre-chosen metric. Change only one main element at a time, set what to measure beforehand, and randomize. Warning: with small traffic, a few conversions apart is just noise, so don't conclude "version B wins" too early. For instance, version A getting 3 signups and version B getting 5 in 2 weeks isn't enough to conclude.

If traffic is low, it's better to make evidence-based changes then carefully compare before and after, or interview 5 to 10 more customers, than trust a test with insufficient sample size. You also need to account for seasonality and different traffic sources between periods, like Facebook visitors versus Google visitors. OPCAIOS doesn't give a specific sample threshold because we haven't verified a trustworthy source.

## Hypothetical example (not a real customer): someone selling design courses

The hypothetical design course seller is an illustrative case with 200 comments and emails asking before purchase. After anonymizing and grouping with Claude, AI produced 3 topics: "don't know if it suits beginners", "worried I won't be able to use it after learning", "price compared to free courses". The seller read back 40 original lines, found groups one and two accurate, while group three appeared only a few times.

Next, the seller added a section "For whom, not for whom", made an FAQ part directly addressing the first two concerns, and ran a headline test when traffic was sufficient. Finally, the seller recorded results in Google Sheets week by week, starting from November 2026. This is an illustration of the process, with no actual result numbers, and makes no promise that every case will produce similar results.

## Ethical boundaries

- Don't manipulate through fear: don't exaggerate risks to force a purchase. Handle concerns with accurate information and real guarantees.
- Don't fake scarcity: fake countdown timers, "only 3 slots left" when untrue.
- Don't fake reviews, don't write testimonials on behalf of customers, don't use AI to generate comments.
- Comply with advertising rules (result claims must be backed up) and personal data rules where you do business. This content is for reference only, not legal advice.

## OPCAIOS's experience

OPCAIOS has a blog, document store and training courses, email sequences and internal management tools, so we use AI to read and summarize reader feedback and questions, rewrite FAQ, then review by hand. We don't publish conversion rates or gains from this approach, and have no verified customer case studies. Our traffic isn't enough to run every statistically meaningful A/B test, so many changes get careful assessment rather than confirmation. Read more [how one person runs with AI](/blog/he-thong-ai-van-hanh-mot-minh) and [AI RAG chatbot with Dify and Supabase](/blog/chatbot-ai-rag-dify-supabase) if you want a chat channel to gather more customer questions (with consent).

## Frequently asked questions

### Can AI really read customer psychology?
No. AI analyzes text the customer already wrote and puts forward a hypothesis. To know for sure, check by hand and run experiments.

### How much data do I need to start?
No standard number. A few dozen to a few hundred genuine responses already point a direction, but a small sample is only indicative.

### Can I feed customer comments into an AI tool?
It depends on data source, purpose and terms. Anonymize, read the tool's policy and current personal data rules.

### What if I have no customers yet?
Use short interviews, surveys and questions from the community; avoid letting AI "imagine" customers then treat that as data.

### Where can I learn how to do this properly?
See [OPCAIOS's AI application training course](/services/dao-tao-ung-dung-ai).

*Source: Jobs to be done framework and four forces mentioned at concept level only, sources not cross-checked yet (read 25/10/2026); legal rules not verified in detail; example is hypothetical, not a real person. Experience is OPCAIOS's.*
