Keyword research in the AI era: from counting volume to understanding questions
The keyword research spreadsheet most small businesses keep has three columns: keyword, volume, difficulty. That sheet is not wrong, but it is going blind. Users are switching to full-sentence questions typed into AI assistants and conversational search boxes, while the volume column only sees short phrases. This article is about the replacement process: researching your customers' real questions, a resource a small business already owns in greater quantity than any paid tool can sell.
Quick summary: Volume-counting keyword research misses the fastest-growing query layer: long, naturally phrased questions put to AI assistants. Around 93% of the keywords in Ahrefs' database get fewer than 10 searches a month, which means most real demand sits below what the tools can measure. Instead of guessing from a tool, SMEs should mine real customer questions from Zalo and Messenger chats, the on-site search box and consultation notes, group them by intent, score them for priority, then write pages that answer each one directly. Volume still helps pick big topics, but real questions decide what each page says.
Why is counting volume no longer enough?
The old method rests on one assumption: that market demand is adequately reflected in the search counts of short phrases. That assumption was never entirely true, and it is drifting further off. According to Ahrefs' long-tail research, around 93% of the keywords in their US database get fewer than 10 searches per month, meaning the vast majority of queries live in a tail that volume reports barely display. Google itself confirmed back in 2017 that 15% of the queries it sees each day are completely new, never searched before.
The AI layer stretches that tail further. Semrush's 2025 clickstream research across tens of millions of records found that purely conversational questions put to an AI chatbot average about 23 words, while traditional-style search queries average only about 4.2 words. Semrush's early-2026 update adds another data point: search-enabled prompts on chatbots nearly doubled in length within a year, from 4.7 to 8.7 words. Users no longer type "spa district 7 price"; they ask "my skin breaks out easily, should I book a scalp treatment or a facial first, and roughly what does it cost". The second phrasing never shows up in a volume tool, but it is the one that becomes an order.
A note on scope: these figures were measured mostly on US market data. There is no equivalent study of the same scale for Vietnam yet, but the direction of travel matches what we read in our clients' query data: questions are getting longer, carry a subject, and come with context.
Where do your customers' real questions live?
Good news for small businesses: the most valuable data source for this work is not inside a paid tool but inside your own inbox. A coffee shop with a Zalo OA, a spa with a Messenger page, an online store with livestream comments each own an archive of verbatim customer questions that no tool can buy. Four sources worth collecting in parallel:
- Zalo OA and Messenger chats: reread the last 30 days of messages and copy every customer question verbatim, including the ones with typos or the wrong terminology. The way customers name their problem is the real keyword.
- The internal search box on your website: the internal search log shows people who were already on your site and still could not find what they needed. Every internal query with zero results is a missing page.
- Notes from consultants and the call line: the questions customers ask again and again over the phone rarely get typed into the web, but they are still questions your site should answer up front.
- Comments and messages during livestreams: this is where customers ask in the most everyday language, for example "will this work on oily skin?".
Three external sources complete the picture: Google Search Console filtered for queries longer than 4 words, search box autocomplete when you type the topic plus question words, and the "people also ask" block on the results page. The Chạm AI question-based keyword suggestion tool gathers these clusters quickly for any topic.
A 5-step question research process for SMEs
The process below is designed so one person can run it in about a week, with no paid tools. The destination is not a keyword list but a prioritised question map, where each cluster of questions corresponds to one page that needs writing.
Five steps, from chat inbox to page outline:
- Collect. Gather 50 to 100 verbatim questions from the four in-house sources above, using the last 30 days as the window. Copy them word for word; do not "translate" them into marketing language.
- Normalise and deduplicate. Merge questions that mean the same thing into one line, keep the most common phrasing as the "representative question" and record how many times it appeared.
- Group by intent. Assign each cluster to one of the four search intent types: informational, navigational, commercial investigation, transactional. The framework is explained in detail in our article on search intent in the AI era.
- Score for priority. Give each cluster two scores from 1 to 5: frequency of appearance and closeness to a purchase decision. Multiply the two and rank; work on the highest products first.
- Write pages that answer directly. Each cluster becomes a subheading, the answer sits in the opening paragraph, and the FAQ at the bottom uses the customers' phrasing verbatim. The formatting principles follow our guide on how to write FAQs AI will cite.
A worked example from a spa in District 7, Ho Chi Minh City: after collecting 80 questions from Zalo and Messenger over one month, the owner found the cluster "can sensitive skin have this treatment" appeared 14 times, "how much is the full package" 11 times, and "do I need to avoid anything afterwards" 9 times. None of the three had ever made it into the old keyword sheet, because the tools reported zero volume for them, yet they accounted for nearly half of the chats that ended in a booking. Three pages answering those three clusters are worth writing before any volume-driven "top 10 spa services" article.
How do keywords and questions work together?
Question research does not replace keyword research; it reassigns the roles. Volume numbers remain best at one job: showing which topics are big enough to invest in and how dense the competition around them is. Real questions are best at the other job: deciding, inside that topic, what a page must answer and in whose voice.
The clean division of labour: use a volume tool to choose 3 to 5 pillar topics for the quarter, then use the question map to build the outline of every page within those topics. Pages written this way are also exactly the kind that generative engines quote back, which is the subject of our article on what GEO is. The two extremes to avoid: chasing volume alone produces pages on the right topic that answer the wrong questions, while chasing scattered one-off questions produces fragments that never form a strong topic cluster.
How often should you rerun the research?
Customer questions shift with the seasons, with promotions and even with the news. A sensible rhythm for an SME is one full 5-step cycle per quarter, plus a quick 30-minute monthly pass to pick up newly emerging questions from the chat inbox. Businesses with a clear peak season, such as weddings or travel, should run the full cycle about 6 to 8 weeks before the season so the new pages have time to be written and indexed.
Frequently asked questions about keyword research in the AI era
How is question research different from traditional keyword research?
Traditional keyword research starts from a volume tool: pick phrases with high search counts and write articles around them. Question research starts from real people: collect customers' questions verbatim, group them by intent, then write pages that answer each one directly. The two approaches complement each other, but the question layer is exactly the part that volume numbers cannot see.
Where can a small business find its customers' real questions?
Four sources are free and already in-house: the business's own Zalo OA and Messenger chat inboxes, the internal search box on the website, notes taken by sales staff or the call line, and comments plus messages during livestreams. Collecting questions from the last 30 days is usually enough to gather 50 to 100 of them and start grouping and writing.
Should you stop using keyword volume tools?
No. Volume numbers are still useful for choosing big topics and estimating competition. The problem is that most long questions sit below the tools' measurement threshold, so relying on volume alone means missing exactly the query layer that is growing fastest. Use volume to pick topics, and use real questions to decide what each page should say.