Keyword Research in the AI Search Era: A Hands-On Workflow for Shifting from "Keywords" to "Questions and Intent"

Why the "keyword" is no longer the unit of research

In AI search, the smallest unit of research has already shifted from the "keyword" to "a single question with intent." Traditional SEO treats "water leak repair Taichung" as a term to fight over, but before ChatGPT, Perplexity, or Google's AI Overviews, users describe their whole situation in complete sentences and ask for an answer—and the AI breaks that question down into needs, then goes to various sources to grab "the passage that answers this need."

This brings two changes. First, the same commercial intent shows up in a huge variety of phrasings, so locking onto a few phrases means you only catch a small slice of those people. Second, AI doesn't hand the whole page to the user; it extracts and cites the passage that best answers the question. So the output of your research ultimately has to map to content where "a single passage can answer a specific question," not a list of keywords.

Step 1: Collect how customers actually ask—don't invent terms yourself

The best raw material for keyword research is the questions customers ask in their own words, not the polished terms a marketer dreams up sitting at a desk. Your goal is to collect the original sentences they'd genuinely say, because AI learns from real human language too—the closer you get to how people actually ask, the more easily your content matches their queries.

Write these original sentences down one by one; don't rush to rewrite them into keywords. A line like "I can't make sense of the renovation quote—are these line items really necessary?" carries far more intent and context than the keyword cluster "renovation quote line items."

• Comb through support tickets, LINE official account chats, and email inquiries: copy down exactly what customers typed, word for word.
• Ask frontline sales and store staff: what are the top 20 questions customers open with—write them in the staff's own words.
• Look at reviews, comments, and group posts about your product or competitors: watch for questions where "someone asks, and a string of people below say 'I want to know too.'"
• Type your core topic straight into ChatGPT or Perplexity and see what follow-up or branching sub-questions it raises—use them as inspiration to expand your phrasings.

Step 2: Sort questions by "intent," not by wording

Once you've collected a big pile of phrasings, the next step is to group them by intent—because what decides what to write and how to answer is intent, not the surface wording. The same intent often has a dozen ways of being expressed, and only by grouping them together do you realize that "this need" actually deserves one complete, well-answered piece of content.

In practice, sorting questions into four intent types works best. After grouping, you'll immediately see which type has the most questions and the least content—that's the topic to write first.

• Informational (wanting to understand): e.g., "How do you tell damp-related mold from a leak?"—needs the principles, ways to judge, and clear definitions.
• Comparative (making a choice): e.g., "Is hiring a general contractor or separately subcontracting better?"—needs side-by-side conditions and who each suits.
• Local/situational (looking for someone or making a decision): e.g., "How do I pick a recommended leak-detection service in Taichung's Xitun District?"—needs local info, selection criteria, and contact details.
• Transactional (ready to act): e.g., "Roughly how much is a leak-detection quote, and how do I book?"—needs price ranges, the process, and a clear next step.

Step 3: Expand into long-tail and conversational queries

Once you've gathered the main questions, expand each intent group in the direction of "longer, more colloquial, more specific," because conversational queries are exactly the mainstream way people ask in AI search. Long-tail and conversational queries have low competition and clear intent—they're often the easiest entry point for small and medium businesses to get a passage extracted and cited by AI.

When expanding, think from the angle of "how a person would add to the question": add the situation, add conditions, add judgment sentence patterns like 'why / should I or not / is it,' and a single main question can usually grow into five to ten sub-questions, each worth answering on its own.

• Add the situation: before or after the main question, tack on concrete circumstances like "old house," "rental," or "only leaks after it rains."
• Add conditions: add budget, area, timing, and scale—e.g., "under NT$30,000 budget," "needs to be done within two days."
• Swap the sentence pattern: rewrite the same intent into colloquial questions like "Should I…," "Is … worth it," "What happens with …," "Can I do … myself."
• Ask what comes next: list the next question users usually ask after getting their answer—these are often the best next passage.

Step 4: Map each question to a "self-contained passage"

The final output of your research is a mapping table of "question → one answer that stands on its own," and this step is the biggest watershed between AI-era keyword research and the traditional approach. When AI cites content, it pulls a "passage," not a "page," so every important question should have a subheading that states the question directly, a first sentence right below it that delivers the conclusion, and a passage you can understand without relying on what comes before or after.

Check each passage with one simple principle: if you copy this passage out on its own and show it to someone, can they understand it and get the answer without seeing the surrounding context? If they can, then AI can confidently—and conveniently—pull it out as a citation.

• Write subheadings directly as questions or clear topics, e.g., "How do you tell damp-related mold from a leak?" rather than "FAQ."
• Lead the first sentence of the paragraph with the conclusion or a direct answer; put details and reasons afterward, making the opening sentence easy to extract.
• Have each paragraph answer only one question—don't cram three questions into one big block.
• Present information like costs, processes, steps, and conditions as bullet lists or tables, so AI can extract them completely more easily.
• Use an FAQ block at the bottom of the article to hold the remaining long-tail phrasings, using the customer's real question as each title and keeping answers specific and actionable.

Step 5: Use traditional keyword tools alongside it—don't throw them out

Shifting to questions and intent doesn't mean keyword tools are useless; the right posture is to treat them as instruments for 'validating and filling gaps,' not as the starting point. First grow your topics out of real phrasings and intent, then use the tools to confirm which phrasings people actually search, fill in long-tail you didn't think of, and see what other questions are hiding in related searches and 'People Also Ask.'

Put simply: let real conversations decide 'what questions to answer,' and let keyword tools calibrate 'which questions are worth writing first, and what wording is most widely used.' Use both together and you'll neither miss demand that has volume nor get stuck on empty terms with no intent.

• Use keyword tools and search-box autocomplete to validate whether the phrasings you collected are actually searched, and to fill in synonymous wordings.
• Look at the search results page's 'People Also Ask' and related searches, and fold the branching questions into your FAQ.
• For core terms with clear search volume, still work them naturally into your title and subheadings to take care of traditional SEO and clicks.
• Feed the terms the tools find back into your 'question → passage' mapping table, rather than starting a separate, disconnected keyword list.
• If you want to know how clearly your current content answers questions and how easy it is for AI to extract a passage, start by running a free GEO health check to get a baseline, then add passages per the recommendations.

Turn the workflow into a repeatable worksheet

Keyword research in the AI era is essentially about engineering 'listening to customers' questions' into a worksheet you can run repeatedly, not a one-off pile of words. We suggest fixing it into a single table: on the left, the customer's real phrasings; in the middle, the intent classification and the long-tail questions you've expanded out; on the right, the corresponding passage subheadings and a one-sentence conclusion.

Update it once a quarter: add the new phrasings that surfaced this quarter, retire topics no one asks about, and flag the passages that got cited by AI or brought in inquiries so you can make more of the same kind. When your content is a series of passages that precisely answer real questions, you become much more likely to be the answer that gets extracted and cited—whether the user searches on Google or asks an AI.