Skip to main content
Back to list
GEO PracticePublished By Zheming Digital Communication Research Institute

How to Collect and Group Prompts for GEO

Collect buyer questions from actual conversations, group them by intent, and decide which pages should answer each group.

"How do users actually ask about our brand in AI engines?" Sina Finance reported January 27, 2026, citing QuestMobile: AI search engines now reach 680 million monthly active users in China, and QuestMobile's 2026 AI Application Market Half-Year Report shows traditional search usage per user fell 19.1% year over year in May 2026. Users hand decisions to AI; brands get cited only if question-matching content exists.

Format matters. SparkToro data compiled by Cintra shows prompts to AI run 2-3 times longer than traditional queries: "CRM recommendation" becomes "what CRM should a 10-person SaaS startup already using HubSpot for email choose?" These context-rich prompts are GEO's ammunition. Seven channels and a four-step method follow.

Why "how users ask" is GEO's first-order data

Traditional SEO optimizes keywords; GEO optimizes the questions users ask AI. Every prompt is a content-demand sample carrying intent, context and budget constraints.

Bain & Company research shows 68% of AI users use AI to research and summarize information, and 42% ask for shopping recommendations. Being cited means presence at the decision point. QuestMobile's Q1 2026 AI Application Insights (April 21, 2026) counts Doubao at 345 million MAU; CNNIC's 57th report, 602 million generative AI users.

This shift makes question data the input for content production: teams measure from real prompts what users ask most. Questions are demand; frequency is priority — that's question clustering (see answer-centric content optimization).

Seven channels to collect 100-500 real prompts

Treat "questions users have asked" as an asset: 100 prompts as the starting point, 500 as the saturation line.

ChannelQuestion typeVolumeNotes
Customer service & pre-sales chatFirst-party, high intent1-3 per chatAnonymize, strip identities
On-site search & website inquiriesFirst-party, purchase intentDozens monthlyConnect with analytics
Zhihu / Xiaohongshu / communitiesSecond-party, real scenariosDozens per topicNote date and heat
Search suggestions & related searchesSecond-party, wide coverage5-10 per keywordPrefer question stems
E-commerce "Ask Everyone" & reviewsSecond-party, transactional5-20 per productPain points from reviews
AI engines' "people also ask"First-party, frontier3-8 per topicDoubao, Yuanbao, Qwen, DeepSeek
Competitor FAQs & help centersBenchmarking20-50 per siteFind gaps competitors miss

Follow four rules: deduplicate and normalize (merge different phrasings), tag source and date, anonymize (no names), and tag intent (informational, comparative, decision, transactional). Compliance: chat logs are personal information; authorize and anonymize before use per China's PIPL.

Four-step clustering: from scattered prompts to a content map

With 100-500 raw prompts, cluster in four steps to produce a question-to-content mapping.

Step 1 — Normalize. Rewrite colloquial prompts as standard questions ("is this thing expensive" → "what is the price of product X?"), merge identical ones, count frequency.

Step 2 — Layer by intent. Tag each question by decision stage: informational, comparative, decision or transactional.

Step 3 — Cluster and weight by frequency. Group by topic (brand, product, price, service, competitor comparison), batch-cluster with LLMs, review manually. Rank by frequency x intent; high-frequency first.

Step 4 — Map outputs and test live. High-frequency informational questions become FAQ and knowledge-center entries (see enterprise AI knowledge base guide); decision questions become answer-centric comparison pages; long-tail items enter the content matrix (GEO content depth guide). After publishing, re-feed the questions to Doubao and Qwen to check citation.

In one anonymized case (industry observation), a B2B equipment company collected 260 prompts from sales, service and community logs. Clustering showed "lead time," "repeat-purchase discount" and "old equipment disposal" accounted for 41% of questions, none covered on the website. After publishing FAQ and comparison pages, AI-answer brand mentions improved. This loop ties clustering to the zero-click era: brands must appear inside AI answers, not links (see zero-click search GEO strategy).

FAQ

Q1: How many real prompts do we need?

Start at 100, 500 as the saturation line. 100 covers a niche's core question skeleton; 300-500 supports a mid-sized company's content matrix. Beyond 500, new prompts fall into existing clusters.

Q2: Where can we get real user questions fastest?

First-party channels are fastest: sales and service chat logs, on-site search logs and inquiries; then Zhihu, Xiaohongshu, communities and e-commerce "Ask Everyone."

Q3: Do we need AI tools for question clustering?

No. Tag topic-intent-frequency manually, use LLMs for batch clustering, finish with human review. Manual works for small samples; keep human validation for large ones.

Q4: How is a collected question different from a traditional SEO keyword?

Keywords are words; prompts are sentences. AI questions are longer, carry context and constraints, and directly determine answer-centric hit rates.

Q5: How do clustering results become GEO content?

Prioritize clusters by frequency x intent: informational questions become FAQ and knowledge-center entries, decision questions become answer-centric comparison pages. After publishing, re-feed the questions to Doubao and Qwen to test citation.


Turning "how users ask" into content input is GEO's basic skill in 2026. Shanghai Zheming offers user-question research, question clustering and answer-centric content in one service. Call +86 18917757529 or email jaysun@widesight.cn.

Written by Zheming Digital Communication Research Institute.