Keyword research for GEO still starts with real search demand, but it no longer ends with a list of high-volume terms. When your goal is to get cited by AI chatbots, you research the questions people ask, the entities those questions revolve around and the subtopics an AI engine is likely to search for while it builds an answer. Search volume becomes one input, not the deciding filter.
That is the short version. Below you get the longer one: what changes, what stays, and how to order the research so your content plan works for Google rankings and AI citations at the same time.
Does search volume still matter when you plan for AI citations?
Yes, but it answers a narrower question than it used to. Volume is an estimate of how often a phrase is searched in a classic search engine. It does not tell you how an AI engine phrases its own searches when it composes an answer.
Google states that both AI Overviews and AI Mode may use a “query fan-out” technique, issuing multiple related searches across subtopics and data sources to develop a response. One question from a user can turn into several searches behind the scenes. Some of those sub-searches may have little or no recorded volume in your keyword tool.
An arXiv study of ChatGPT citations points in the same direction. Across two 90-day collections of answers to commercial questions (158,099 answers with 468,037 citations), the researchers found that every page ChatGPT cites is one its own search retrieved, for sub-queries the model rewrites on every answer and that rarely coincide with Google’s or Bing’s results. Which pages that search retrieves explained far more of the variation than page properties: 32% against 2.8%.
Keep the scope of that study in mind. It looked at commercial questions in two markets, with a replication in a market of small B2B vendors. It does not prove how every engine behaves for every query type. Still, the practical lesson for your research is clear enough:
- Use volume to estimate demand. Volume figures are estimates for a phrase. Weigh them alongside audience fit and how useful a page would be for your readers; low recorded volume does not prove that nobody asks the question.
- Do not use volume alone to drop low-volume questions. A long, specific question with little or no recorded volume can still match the kind of sub-query an engine runs.
- Weight coverage over single terms. The goal is to match the meaning of many related queries, not to rank for one phrase.
Where intent segmentation fits: before the keyword list
In classic research you often pull a big keyword export first and sort it later. For GEO, flip that order. Decide which intents you want to serve before you collect terms, so every keyword you find has a job.
Intent segmentation groups queries by what the searcher wants: to understand a concept, compare options, solve a problem or buy. Google describes AI Mode as particularly helpful for queries that need further exploration, reasoning, or complex comparisons, from exploring a new concept to comparing options. Our recommendation: give those intents their own place in your plan, since Google names them as the kind of queries where AI Mode is particularly helpful.
If you already have Search Console data, you can build these segments from your own queries. Our guide on how to segment queries by intent in Search Console walks through the setup. Once the segments exist, research becomes a question of filling gaps per intent instead of chasing whatever term has the biggest number.
A simple starting set of segments:
- Define: what is X, how does X work
- Compare: X vs Y, which option fits a situation
- Solve: why does X happen, how do I fix X
- Decide: cost, pricing, choosing a provider
Entity and question research vs traditional keyword clustering
The table below compares a simplified, volume-led keyword workflow with the entity and question workflow this article recommends. In that volume-led workflow, keywords are grouped by shared words or overlapping ranking pages, and each cluster is assigned to one URL. Entity and question research groups by meaning: which things a topic involves, and which questions people ask about those things. The comparison is a planning simplification, not a description of every clustering method.
| Simplified volume-led keyword workflow | Recommended entity and question workflow for GEO | |
|---|---|---|
| Uitgangspunt | Seed keyword and volume export | Intent segments and the entities in your niche |
| Unit of research | Zoekwoordcombinatie | Question plus the entities it mentions |
| Grouping logic | Shared words or overlapping ranking URLs | Shared meaning and shared subtopics |
| Main filter | Search volume and difficulty | Demand, fit with your expertise and coverage gaps |
| Uitvoer | One target keyword per page | A page that answers a question and its likely sub-questions |
Researching entities
An entity is a distinct thing a model can recognize: your brand, a product, a concept, a method, a competitor. Mapping entities helps you see which concepts must be explained clearly and consistently across your site.
The same study found that ChatGPT writes brands it knows into its searches, prefers websites it knows, and names brands without any supporting page. Your own brand is one of the entities in your research. List how it connects to the concepts you want to be known for, and keep naming consistent. For the on-site side of that work, see duidelijkheid over entiteiten voor AI-zoekopdrachten op uw hele website.
Researching questions
Within one website, the researchers found that pages whose title and paragraphs match the questions people ask are cited more, while FAQ blocks, structured data, summary boxes and page speed appeared effective only across websites. In plain terms: the question you choose to answer, and how directly your title and paragraphs answer it, matter for your own pages.
So your research output should be phrased as questions, not fragments. “GEO keyword research” becomes “How does keyword research change for GEO?” plus the follow-ups a reader or engine would naturally ask next.
A practical research workflow
- Set your intent segments. Pick the intents you want to own before collecting anything.
- List your core entities. Brand, products, concepts, methods and the competitors you are compared with.
- Collect questions per entity and intent. Use Search Console, sales and support conversations and classic keyword tools. Our guide on ontdekken welke vragen klanten aan AI-chatbots stellen covers sources.
- Expand each question into likely sub-questions. Think like a fan-out: which comparisons, definitions or conditions does a full answer need?
- Check demand, but keep low-volume questions with clear fit.
- Group by meaning into topic clusters. This is how coverage turns into erkend autoriteit op dit gebied.
- Assign each core question a page and link supporting pages to it.
What this research step is not
Keyword research for GEO is a planning step. It decides what you publish. Prompt testing is a separate step that checks how chatbots respond once content exists. Mixing them leads to plans built on a few chatbot outputs that change between runs.
When your plan is ready and you want to measure inclusion, move on to het opbouwen van een prompt-universum voor AI-zichtbaarheid. New to the wider discipline? Start with wat generatieve zoekmachineoptimalisatie is.
A good next move this week: take your ten most important keywords and rewrite each one as the question a buyer would ask, then list three sub-questions under each. That list is the start of a GEO content plan. If you would rather have that plan planned, written and published for you without building an SEO team, bekijk hoe Authora werkt.