From keywords to questions: rebuilding your SEO research for AI search
AI-assisted research. This note was prepared with the assistance of AI tools. It may contain errors and is not professional advice. See our Legal Disclosures.
Keyword research has shaped content strategy for two decades: find terms with search volume, map them to pages, and optimize. That method still has value, but it describes only part of how people now look for information.
An AI assistant invites a different kind of input than a search box. Instead of a fragment like "marketing automation pricing", a buyer can ask "Which marketing automation tools work well for a 20-person B2B company with a small budget?"
Content research needs to follow that shift, from keywords to questions.
Why questions matter more now
A keyword is a compressed question with the context stripped out. Assistants accept the context back, and that changes what research should capture.
Questions carry context. A question includes the asker's situation, constraints, and goal. That context decides which answer is useful and which brands get mentioned. "Pricing" means something different to a founder with a small budget than to a procurement lead comparing contracts.
Answers are assembled, not ranked. Assistants combine information from several sources into one response. Content that directly answers a specific question is easier to use in that process than a page written to cover a keyword broadly. We go deeper on this in Citations, Not Rankings.
Long questions rarely show up in volume data. A question with a specific team size, budget, and use case may be asked in many slightly different forms. Keyword tools are built to count repeated strings, so they tend to miss these. Teams need research methods that do not depend on volume.
Where to find real questions
The best sources are places where buyers already speak in their own words. Each one has a bias, so we recommend drawing on several.
- Sales calls and notes: the questions buyers ask when they are close to a decision, often about fit and risk.
- Support tickets and chat logs: the questions customers ask after buying, which often reveal what pre-sale content failed to explain.
- Community discussions: forums, professional groups, and review sites, where people describe problems before they know your category's vocabulary.
- Site search: internal search queries that show what visitors came for and could not find.
- Assistants themselves: ask an assistant what questions people typically have about your category, then confirm those against real conversations.
Treat the last source as a prompt for ideas, not as evidence. An assistant's list of likely questions reflects its training, not your buyers.
Call notes and tickets often contain personal information. Strip names, contact details, and account data before any of it goes into a research tool or an agent. Check with your privacy team, and counsel where needed, on what you may use and how long you may keep it.
Organizing questions into a map
A pile of questions is not yet a plan. The next step is to group them by the stage of the buying process and the intent behind them. We use four groups.
- Problem questions: "Why does our pipeline reporting take so long?"
- Approach questions: "Should we automate reporting or hire an analyst?"
- Option questions: "What tools automate pipeline reporting?"
- Decision questions: "How does Tool A compare with Tool B for small teams?"
Within each group, merge near-duplicates and keep the clearest phrasing as the label. Keep a few original wordings under each label too, because the variations show the constraints buyers care about.
Then check coverage. For each question, mark whether your existing content answers it clearly, answers it poorly, or ignores it. The pattern is usually more useful than any single gap. A team that answers decision questions well but ignores problem questions is invisible to buyers who have not yet named their category.
A worked example
Consider a hypothetical B2B software team that pulls one month of material: 60 sales call notes and 200 support tickets. A research agent extracts every question it can find and returns several hundred raw questions.
After merging duplicates, the team is left with about 90 distinct questions. Sorting them into the four groups shows a lopsided map. Most of the team's content addresses option and decision questions. Almost nothing addresses the 25 problem questions, which describe the frustrations that lead people to look for a tool at all.
The team does not write 25 new pages. It picks the eight problem questions that came up most often in sales notes and writes one clear page for each, each answering the question in its first paragraph. It then adds those eight questions to its monthly sampling set to see whether assistants begin to draw on the new pages.
Where agents help
This work suits agent support because the volume is high and the judgment calls are narrow. A pipeline of two or three agents can handle most of the mechanical steps. Our post on multi-agent content pipelines covers how to divide that work.
- Extract: pull every question from notes, tickets, and transcripts, keeping the original wording.
- Group: merge near-duplicates and assign each to a stage and intent.
- Count: tally how often each theme appears and in which source.
- Audit coverage: check whether your site answers each question and summarize the gaps.
People still decide which questions matter most and how to answer them honestly. An agent can find that buyers keep asking about migration risk. It cannot decide how candid your answer should be.
Failure modes to watch
Question research has its own ways of going wrong. We see three that are worth checking for.
Over-weighting loud sources. Support tickets skew toward problems, and community threads skew toward strong opinions. Balance them against sales notes, which skew toward evaluation.
Writing to the wording, not the need. A page that repeats a question word for word but answers it vaguely will not help a reader or an assistant. Answer the question first, then explain.
Letting the map go stale. Buyer questions change with products, pricing, and competitors. Refresh the map each quarter from new material.
Pair it with measurement
A question map is also a measurement plan. Once you know which questions matter, you can track how assistants answer them over time. See Measuring Share of Answer for a method, and Structuring Content for Answer Engines for how to write the pages themselves.
The question we are studying next is how stable buyer questions are across a year. We expect problem questions to change slowly and decision questions to change quickly, and we want to know how often each group needs to be refreshed.
How to apply this: pull one month of sales call notes or support tickets, remove personal details, and extract every question a buyer asked. Group them into the four stages, then check how many your website answers directly. The gaps are your next content plan.
Related research
Citations, not rankings: how brands get referenced in AI answers
AI answers often have no position to rank for. What matters is whether you are cited or mentioned, and what earns that overlaps with good marketing.
Structuring content for answer engines: entities, claims, and evidence
AI answer systems favor content that is easy to interpret and check. Pages built on clear entities, plain claims, and visible evidence are easier to use.
Answer engines and the shrinking click: planning for AI search in Q2
When an AI answer resolves a question on the spot, the visit may never happen. Plans should value being the source of the answer, not only the destination.
Put this research to work
Marketix turns these patterns into software: agents, automated workflows, and a hub that keeps people in control. Sign up for early access.
