Keyword Research for AI Search: How to Find Prompts, Fan-Out Queries and Topic Trends

Quick quiz

How should you get mentioned by ChatGPT and other AI assistants?

Question 1 of 425%

What do you want to know or change?

Just see if AI assistants talk about my businessFollow it over time against my competitorsBe mentioned more often, starting now

Keyword research for AI search starts from a simple observation: people do not type keywords into an assistant, they ask questions with context. OpenAI's study of how people use ChatGPT (NBER working paper, September 2025) found that about 49% of messages are "asking" requests, and that practical guidance, seeking information and writing together account for nearly 80% of conversations. A list of two-word keywords misses most of that. This guide shows how to build the three layers, which signals to use for volume and trends, and how to turn them into pages that get cited.

What changes in keyword research when buyers ask AI?

Classic keyword research looks for short queries with a monthly volume. AI research adds two objects that keyword tools do not show well: the prompt, written in natural language, and the searches the assistant derives from it.

The three layers do not compete. The prompt tells you the intent and the context of the buyer. The fan-out queries tell you which pages the assistant will read. The keyword tells you whether the same topic also brings Google traffic. Our guide on how AI assistants choose their sources explains why the fan-out layer decides which pages get retrieved.

Does prompt volume exist?

Not as an official figure. OpenAI, Google, Anthropic and Perplexity do not publish how many times a given prompt is asked. The numbers sold as prompt volume are estimates built from panels, browser data or models extrapolating from search volume. They help rank topics against each other; they are not counts.

Three official signals come closer to the truth:

• Google Search Console: Google's documentation on AI features states that traffic from AI Overviews and AI Mode is reported in the Performance report, within the Web search type. Long, question-shaped queries in that report are the closest thing to prompts you can see.
• Bing Webmaster Tools: the AI Performance report lists the grounding queries that led Copilot to cite your pages. Since June 2026 Microsoft also groups them by intent and topic, with a citation share per query, which is a rare first-party view of AI demand.
• Classic keyword volume: fan-out queries are short and often match real keywords, so their search volume is a sound proxy for the demand behind a prompt.

How to do keyword research for AI search, step by step

1. Collect the questions your buyers really ask

Start from sources that contain full sentences: sales and support conversations, customer reviews, forum threads in your sector, and the longest queries in Search Console. Rewrite each one as a prompt a buyer would type, with its context (company size, budget, location, use case).

2. Sort the prompts by intent

Group them by what the buyer wants to know. Four groups work for most markets: understanding (how does it work), benefits (what will it bring me), comparison (which option should I choose) and quality (how do I recognize a good provider). Comparison and quality prompts are where buying decisions happen, so cover them first. The classic notion of search intent still applies, with more nuance per sentence.

3. Build a panel of 25 to 150 prompts

A panel that is too small misses topics; one that is too large cannot be tracked. Between 25 and 150 prompts covers most businesses. Our prompt library for SEO and GEO gives starting formulations you can adapt.

4. Capture the fan-out queries

Send each prompt to the assistants your buyers use and record the searches they run. Google describes AI Mode as breaking a question into subtopics and issuing a multitude of queries simultaneously, and its Deep Search can issue hundreds of searches for one report. These sub-queries are often longer and more specific than anything in a keyword tool, and they change from one model to another.

5. Check the demand behind each query

Run the fan-out queries through a classic keyword research tool to get volume, difficulty and intent. A sub-query that the assistants use often and that also has Google volume is a double win: one page can rank on Google and be retrieved by the AIs.

6. Group into topics and map one page per cluster

Cluster the prompts and sub-queries that share an answer, then assign one page per cluster to avoid cannibalization. The topical cluster generator does this grouping from a seed topic.

How do you track topic volume and trends in AI search?

Trends matter more than absolute numbers, because both demand and the assistants' behavior move. Our ChatGPT citation study shows that ChatGPT moved from 5.8 to 4.1 sources per answer between August 2025 and August 2026: the same prompt can open or close citation slots within a few months.

Two habits make these signals useful. First, compare periods rather than reading a single month, as our guide to Google Trends for SEO explains for seasonality. Second, track the same prompts every week: the share of answers that mention you, measured on a stable panel, is the most reliable trend line you can build on AI visibility.

Which mistakes should you avoid?

• Keeping only high-volume keywords: the prompts that lead to a purchase are long and specific, and their sub-queries often show little or no volume in keyword tools.
• Testing a prompt once: answers vary from one run to the next, so a single screenshot proves nothing. Repeat the same panel over several weeks.
• Ignoring the model your buyers use: sub-queries and cited sources differ between ChatGPT, Gemini and Claude. Prioritize the assistants your customers say they open.
• Writing one page per prompt: prompts that share an answer belong on the same page, otherwise your pages compete with each other on Google and dilute the passage the AIs could quote.

How does keyword research for AI search connect SEO and GEO?

The same research feeds both channels. Fan-out queries with Google volume become articles that rank; prompts with high buying intent become comparison and pricing pages that the AIs can quote; weekly tracking tells you which of them works. Keyword research for AI search starts from the buyer's question, follows the searches the assistant runs, and checks the demand with real data.

Sorank automates the loop. AI prompt research writes the prompts of your market from your site and sorts them by intent, AI query fan-out records the searches each assistant runs behind them, and AI mention tracking measures every week whether your brand is cited on ChatGPT, Claude and Gemini, plus Perplexity and Grok from the Pro plan. The fan-out queries then become keywords in your editorial calendar.