DEV Community

卜代文
卜代文

Posted on Originally published at we0.ai AI-assisted

Keyword Research for AI Search: From Search Volume to Question Chains

Originally published on We0 on November 11, 2026. This English edition was adapted with AI assistance.

Search volume still helps estimate demand. To plan content for conversational search, also map the follow-up questions, comparison criteria, and answer formats behind each customer decision.

Conclusion First: Search Volume Has Not Become Invalid, But It Is No Longer the Only Basis for Topic Selection

Keyword search volume answers 'how many people search with this set of words.' It does not answer 'what else will this person ask before making a decision.' AI search has turned the latter into a decisive variable: when a user asks a question with constraints in ChatGPT, Perplexity, Gemini, or Google AI Mode, the engine generates multiple sub-queries to retrieve information, then assembles the retrieved passages into an answer with citations. What determines whether you get cited is no longer just whether you rank for a certain word, but whether your page contains a passage that can be extracted on its own and precisely answers a sub-query.

So a more accurate statement is: search volume is still a useful demand-scale signal, but it is a starting metric, not an ending metric. Keyword research needs to add three things—the complete question chain behind this demand, the retrieval format corresponding to each question, and whether the page has extractable answer blocks.

Why Looking Only at Search Volume Misses Real Purchase Intent

The traditional process is 'primary keyword → related terms → content,' which assumes every search is an independent need with a fixed word form. User behavior has already changed: a real question often simultaneously contains budget, scenario, compatibility, and usage, for example 'a laptop under 15,000 yuan that can edit video, has good battery life, and is convenient for business trips.' In AI search, these conditions are split into different branches and retrieved separately.

Taking B2B procurement as an example, behind 'best CRM software' there is usually a series of follow-up questions: Which one is suitable for small businesses? Does it support email automation? Can it integrate with Shopify? Which has the lowest monthly fee? The retrieval value of these four questions may be higher than the original term itself, but monthly search volume is often close to zero—few people type the complete question into a search box, but they will ask AI this way.

This creates two gaps. First, low search volume does not equal low commercial value. Second, search volume gives no hint about answer format: it does not tell you that comparison needs require a table, process needs require numbered steps, and definition needs require a 30–50 word explanation that can stand alone. Yet it is precisely these formats that determine whether content can be retrieved.

What Is Query Fan-Out?

What Is Query Fan-Out?

Query Fan-Out refers to an AI search engine splitting a question into several model-generated sub-queries, retrieving them in parallel, scoring the returned passages, and then synthesizing an answer with citations. Google states in its AI features documentation that experiences such as AI Mode and AI Overviews may use this technique to present 'a broader and more diverse set of useful links' than a conventional results page (Google Search Central).

From an engineering perspective, it usually involves five steps: identify intent and complexity, generate sub-queries, filter highly duplicate sub-queries, retrieve in parallel, and synthesize the answer with citations. The third step is most easily overlooked: the system prefers coverage over repetition, and two pages with similar content generally will not both enter retrieval.

In terms of scale, public observations roughly are: a single AI Mode question produces about 8–12 parallel sub-queries (commonly 3–8 in practice); ChatGPT averages 4.7 sub-queries per online search; 'hundreds of searches' applies only to multi-round pipelines such as Deep Search, with an upper limit of about 20 rounds. More critical is the form of sub-queries—reverse analysis shows they average 70–80 words in length, about 17–26 times more complex than traditional three-to-four-word keywords (MaximusLabs). The engine does not search with your keyword; it searches with a complete description that includes context and usage scenario.

Fan-out also does not happen unconditionally. Relevant patents tie whether sub-queries are generated to input length, query frequency, result quality, token count, and server load. Short navigational or simple factual questions often trigger only a single retrieval close to the original question; long comparative and evaluative questions expand the largest branch tree. Budget should therefore be prioritized for complex commercial questions.

Search Intent Has Changed from 'One Word' to 'One Question Chain'

The four intent types—navigational, informational, commercial investigation, and transactional—still hold, but in AI search, one primary intent extends into multiple layers of sub-intent. Take 'How do I choose the best SEO tool for my business?' as an example. The primary intent is choosing a tool, but sub-intents include: What type of tool do I need? What is the budget? Which features are key? Do I need technical SEO capabilities? What suits beginners versus agencies?

Here we need to distinguish two concepts that are often confused: keyword clustering groups similar words by intent, semantics, and SERP overlap, addressing how content is organized; Query Fan-Out describes how one question splits into multiple retrieval paths, including topics the user never explicitly mentions, addressing how retrieval happens (SEO for Today). The former is an editorial method, the latter an engineering mechanism; they cannot replace each other.

Table: Eight Types of Sub-Queries and the Content Formats They Require

After categorizing fan-out branches, a single question usually produces eight types of sub-queries, each requiring a different passage format. This table can be used directly as a content self-check list:

Sub-query Type Example Required Content Format
Exact match best CRM software An answer block that directly gives options
Semantic variant top CRM platforms Restate the same answer using another buyer phrasing
Navigational a vendor's pricing page A crawlable, clearly structured product or pricing page
Comparative A vs B for small business Table rows with explicit attributes
Definitional what is pipeline attribution A 30–50 word short definition that can stand alone
Process how to migrate CRM data Numbered steps, one action per step
List CRM tools for agencies A scannable list with filtering criteria
Scenario CRM for a 3-person team A qualifying sentence naming team size and usage scenario

Pure prose pages usually cover only two or three of these: without a table, the comparison branch has nothing to extract; without numbered steps, the process branch fails; if team size is never mentioned, the scenario branch cannot match. Format here is not a layout preference but retrievability itself.

Expanding from One Keyword to a Complete Question Chain: A Five-Step Method

Expanding from One Keyword to a Complete Question Chain: A Five-Step Method

Step one: list seed prompts rather than keywords. Collect 10–20 questions that ideal customers would actually type before booking a demo; the more colloquial, the better.

Step two: run each seed through Perplexity. Its interface shows the actual retrievals performed, and it is currently one of the few places where fan-out behavior can be observed.

Step three: record the 5–8 related questions it surfaces one by one, keeping the original wording without rewriting them into keywords.

Step four: cross-check the same set of seeds with a question tool such as AlsoAsked to fill in missed branches.

Step five: map the surviving sub-questions to H2s and H3s on existing pages. The principle is 'one question cluster, one page,' not 'one sub-query, one page': building a thin page for each sub-query dilutes authority and may touch Google's policy boundary on scaled content abuse (MaximusLabs).

It should be clear that the real branch tree is conversational and personalized, and cannot be fully replicated externally. The question chain is a reconstructed approximation, suitable for coverage planning but not for reporting externally as precise data.

Content Clusters: Turning the Question Chain into a Three-Layer Structure

In implementation, it can be divided into three layers: pillar pages answer 'the complete framework for how to make this decision'; sub-pages address sub-questions with independent search demand, such as comparisons of certain solution types or the price structure of a certain industry; FAQ blocks address peripheral questions such as compatibility, refunds, and minimum configuration.

To decide whether a sub-question deserves its own page, use three checks: Would it be asked separately in a real conversation? Is its answering task something the pillar page cannot handle without going off-topic? Can you state in one sentence how it does not overlap with sibling pages? Only when all three pass should you build a new page.

At the same time, be aware of priority differences in the bottom of the funnel. Fan-out for purchase-type questions usually concentrates on pricing, plans, free tiers, refund policies, and competitor comparisons, so comparison pages and pricing pages are more likely to enter the purchase path than definition pages; definitional questions are often answered by the engine itself, and winning them mostly earns a mention.

How Should FAQ and Product Pages Divide Responsibilities?

The two undertake different retrieval tasks, and mixing them makes both worse. The goal of a product page is to clearly state 'whether it suits me': who it applies to, what capabilities it includes, where the boundaries are, what the pricing structure is, and how to get started. It should explicitly state team size, usage scenarios, integrations, and limitations, because these are exactly the qualifying information used by scenario-type sub-queries for matching.

The goal of an FAQ is to remove decision friction and answer peripheral questions that do not belong in the main body but that users will certainly ask. Each answer must be able to stand independently of context: replace pronouns with entity names, give the conclusion first and then the conditions, and keep the length to a few dozen words.

The internal order of a page is also a retrieval decision. In citation pattern observations, about 44% of citation share is concentrated in the first 30% of the body, then gradually declines (MaximusLabs). The most commercial answers should therefore appear in the first third of the page, not at the end.

Can Search Engines and Conversational Search Be Covered Simultaneously?

Can Search Engines and Conversational Search Be Covered Simultaneously?

In most cases, yes, and the overlap is larger than imagined. The preferences of several major entry points do not conflict: AI Mode values sub-topic coverage and passage clarity; ChatGPT values self-contained Q&A blocks and expertise signals; Perplexity values timeliness and source transparency; Claude's research mode values depth and methodology. Their intersection is answers up front, visible sources, plain language, and genuine depth in the body.

One directional change is worth noting: in public observations, AI Overview citations' dependence on top 10 organic results fell from about 76% to about 38%, with a considerable share of citations coming from positions outside the top 100 (MaximusLabs). This does not mean 'rankings are useless,' but that 'rankings are no longer the only threshold for citation'—whether a passage can be extracted and directly answer the question has clearly risen in weight.

Fundamentals still need to be done: technical crawlability, content structure, internal links, and structured data remain prerequisites for entering AI search experiences (Google Search Central).

What Businesses Should Do Now: A Six-Item Checklist

Priority Action Completion Standard
1 Replace the keyword list with 10–20 real customer questions Each is a complete colloquial question
2 Run them through Perplexity and record follow-up questions Record at least 5 branches per seed
3 Self-check core commercial pages for coverage of the eight sub-query types Cover at least six of the eight
4 Add an independent answer block under each H2 Replace pronouns with entity names; conclusion before conditions
5 Move the most commercial blocks to the first third Appear within one screen after the first screen
6 Assign a unique owning page to each question cluster No two pages answer the same question

At the same time, retain some search volume judgment: it is still a low-cost signal for estimating demand scale, setting priorities, and judging seasonality. What should truly be abandoned is the single-line rule 'write if search volume is high, don't write if it is low.'

How we0 Carries This Chain

For the question chain to become a content asset, it must ultimately land on the website: page structure, FAQ blocks, comparison tables, structured data, multilingual versions, and subsequent update cadence. The path offered by we0.ai is to put website building and content growth in the same workspace: describe requirements in natural language, have multiple agents collaboratively generate a runnable page, adjust it on a visual canvas, deploy it to your own domain, then continuously maintain content through the CMS backend and handle SEO and GEO-related settings on the site side (we0.ai).

For teams that need to quickly validate question clusters, the value of this integration is not 'how fast it generates,' but that from a sentence of requirements to page publishing and then to checking coverage by question chain, there is no need to shuttle content and copy back and forth among three or four tools.

FAQ

Q: Do we still need to look at keyword search volume?
Yes, but use it to judge demand scale and priority, not as the sole basis for topic selection. Search volume cannot reflect sub-query branches, answer formats, or intent layers; these three must be supplemented by question chain research.

Q: Are Query Fan-Out and keyword clustering the same thing?
No. Keyword clustering groups similar words by intent and semantics for content organization; query fan-out is the mechanism by which the engine splits one question into multiple sub-queries for parallel retrieval, used to explain retrieval behavior.

Q: How can I tell whether my page can be cited?
Start with extractability: whether each section has an answer block that can stand alone, whether it names entities and applicable scenarios, whether it has tables or steps, and whether sources are visible. These can all be checked manually without relying on third-party scores.

Q: What if a small team has no budget for complete research?
Start with the minimal closed loop: choose 10 real customer questions, run conversational search and record branches, fill core commercial pages with tables, steps, and answer blocks, and rerun quarterly with dates noted to distinguish whether result changes come from content adjustments or engine updates.

Related Tools

  • Google Search Central's AI features documentation: confirm official wording and basic requirements.
  • Perplexity: the interface shows the retrieval process, suitable for observing fan-out behavior.
  • AlsoAsked: retrieve question clusters by keyword and fill in missed branches.
  • we0.ai: an integrated workspace from website building to content operations and SEO/GEO settings.

Related Links

Summary

Search volume has not become invalid; it has simply receded from a 'topic selection basis' to a 'demand scale reference.' The new work of keyword research is to expand a word into a complete question chain, prepare extractable answer formats for each question, and then allocate pages by question clusters rather than individual words. This serves both traditional search results and conversational search that rewrites queries on its own. The judgment standard also changes accordingly: it is not how much search volume a word has, but whether your site can still be the one answer cited when a user asks a fourth or fifth follow-up question.

Top comments (0)