The blind spot keyword tools cannot see
Every keyword research workflow you learned in the last decade starts with a volume number. You type a topic into a tool, it returns a monthly search volume, and you decide whether the prize justifies the effort. The logic is sound for traditional search. It is almost useless for AI-driven search.
When someone types "best CRM for a professional services firm with 12 people and a US and Singapore presence" into ChatGPT, Perplexity, or Google AI Mode, that query never appears in any keyword tool. It has no volume. It exists once, in that session, and it is gone. But it represents a buyer at exactly the moment of active vendor evaluation. Missing it is not a minor gap.
This is the structural problem: keyword tools measure what people typed into a search box with a ten-blue-links interface. They do not measure what people prompt into conversational AI interfaces, because those sessions produce no click stream to harvest. The data simply does not flow into the tools you are accustomed to trusting. HubSpot's State of Marketing 2026 found 41% of marketers had already updated their SEO strategy in response to AI search shifts. The other 59% are running a strategy calibrated for a channel that is shifting underneath them.
The fix is not a new tool. It is a different data source. Keyword research has become prompt research.
Building a prompt set from what you already have
A prompt set is a curated list of the conversational queries your buyers use when they are evaluating, comparing, or trying to understand a topic your product or service addresses. Building one from your own data is not guesswork.
The method has three inputs, each from a source you already control.
Input 1: GSC question-format queries. Export your last 90 days of query data from the GSC Performance report. Filter for queries longer than four words (the conventional keyword length ceiling) or for question starters. Group them by intent cluster: evaluative ("how do I choose"), comparative ("X vs Y"), process-level ("how to"), and outcome-level ("best way to achieve"). These clusters are the topic spine of your prompt set.
Input 2: Bing grounding queries. Pull the grounding query report from Bing Webmaster Tools. These queries reveal how Copilot internally translates user prompts into search phrases to retrieve your content. A grounding query like "AI marketing consultant Singapore B2B" tells you Copilot is fielding user prompts about AI marketing consultants in Singapore and finding your page relevant. A grounding query on a topic your page addresses poorly is a gap to close.
Input 3: Prompt-pattern inference. Work backwards from grounding queries and GSC long-tails to infer the original user prompts that generated them. A grounding query of "first-party data strategy B2B SaaS" most likely came from a user prompt like "how should a B2B SaaS company handle first-party data in 2026." That inferred prompt belongs in your prompt set; it is what your content must answer.
The output of this process is a structured list of 30 to 60 prompt-level questions your buyers are asking AI systems. It is not a keyword list with volumes. It has no monthly search estimates. What it has is intent fidelity: every item on the list represents a real moment of active evaluation, not a historical aggregate of clicks.
Worked prompt-mining template
The steps below are the operational sequence leapbuzz uses when building a prompt set from a client's existing data. Each step has a defined input, a defined output, and a copyable query filter or export instruction.
Turning long-tail prompts into posts
The prompt set tells you what to write. The conversion step is mapping each prompt to a content format that answers it well enough to be retrieved and cited.
Not all prompts produce the same content format. Evaluative prompts ("how do I choose a digital analytics partner") call for decision frameworks: criteria, trade-offs, common mistakes, a structured comparison. Process prompts ("how to set up a first-party data pipeline") call for step-by-step structure with named outputs at each step. Outcome prompts ("what results should I expect from influencer marketing") call for honest benchmarks, context for variation, and the variables that matter most.
The structural principle that applies across all of them: answer the prompt in the first 150 words. AI retrieval systems quote from the opening of a section more often than from its middle. The Princeton GEO study (arXiv:2311.09735) found that citation-adding content patterns, including direct quotable statements and structured fact blocks near the top of a section, lifted AI visibility by up to ~40% on their benchmark. Front-loading the answer is not a stylistic choice. It is a retrieval mechanic.
Each prompt that clears the intent threshold becomes a standalone blog post. Not a section in a longer page. Not a paragraph buried in a pillar. A post with its own URL, its own structured data, and its own BLUF block that answers the prompt in two sentences. This is how leapbuzz approaches content architecture for AI visibility: one prompt, one URL, one authoritative answer. The post you are reading was seeded by exactly this process.
For clients across SG, AU, US, CA, and MY, the prompt-set approach surfaces market-specific demand patterns that aggregate keyword tools miss entirely. A query like "what are the GST implications of influencer marketing payments in Singapore" has essentially no measurable keyword volume, but it represents a live compliance question from a marketing procurement team. A post that answers it precisely gets cited. A pillar page that mentions it in passing does not.
The measurement loop: from grounding query to pipeline
Publishing prompt-set posts without a measurement loop is planting seeds and never checking the soil. The loop is straightforward once the two data sources are configured.
Set a monthly review cadence. In Bing Webmaster Tools, track grounding-query-to-citation conversion rate by page: which pages are retrieved but not cited? That gap is where you improve answer structure, add a BLUF, or sharpen the first paragraph. In the GSC AI performance report, watch which pages gain impressions in AI features month over month. Impression growth without traffic growth is normal right now. AI answers resolve queries without sending clicks. That is brand presence in the answer layer, not a failure state.
For the post-to-pipeline connection, use UTM parameters on CTAs within prompt-set posts. When an AI answer cites your post and a reader clicks through, that session is trackable. The post is the citation vehicle; the CTA within it is the lead channel. Each prompt-set post is a tiny landing page with a specific action and a specific audience.
The Semrush/Indig study published July 1, 2026, found that ChatGPT's reasoning mode generates roughly 1,130 searches per 100 user prompts, against roughly 245 for fast mode. That ratio means reasoning-mode queries drive substantially more retrieval activity per user session. For B2B buyers doing deep research in reasoning mode, the content that gets retrieved and cited is the content that addresses the full prompt with precision, not the content optimised for a two-word keyword. Prompt research is the only method calibrated for that retrieval pattern. A keyword tool cannot help you there.
For deeper coverage of how AI answer engines select and cite sources, the GEO playbook covers the full retrieval-and-citation architecture. The AEO step-by-step guide walks through the entity and schema layer that makes content machine-readable. And ChatGPT reasoning mode citation patterns unpacks the specific mechanics of how reasoning models decide what to quote. If you want to understand the gap between how Bing Webmaster Tools surfaces AI citation data today and what a fuller measurement stack looks like, leapbuzz's visibility optimization practice is where to start.