Show me your keyword list and I can usually tell you the generation it came from. If everything is "seo services" and "marketing agency Vancouver" with search volumes in the thousands, you're working from a 2018 playbook. It's not that those keywords are dead — it's that they're the ones AI engines already answer so capably that nobody needs to scroll to see whose site is selling.
The demand that's still up for grabs is fragged into a thousand small questions. And the tools built for ten blue links barely see them.
The shift is structural, not cosmetic. For two decades, keywords were how we described a searcher's intent to a search engine. Now the search engine — and increasingly the answer engine — describes intent through natural-language questions, and the winners aren't the sites that optimized a phrase, but the ones that comprehensively answer a question. LLM keyword research is how you find those questions while they're still underpriced.
Why classic tools miss this
Search-volume tools measure typed queries against a results page. AI answers don't work that way. A buyer doesn't type "best seo agency" and click the first blue link; they ask "how do I choose between a local SEO agency and an AI content agency?" and the AI synthesizes from whoever's cleanly answering that specific angle.
Those conversational questions carry real intent but tiny, unreliable volume readings in classic tools. A keyword tool will show you "seo services" at 22,000 searches a month and effectively nothing for "how do I know if my seo agency is actually doing work" — and the second one is closer to a decision than the first. So everyone ignores the real demand, which is exactly why it's under-priced.
There's a deeper reason the old model breaks down: LLMs don't rank pages, they concatenate sources. A query isn't a keyword auction anymore — it's a synthesis. The question isn't "which page ranks #1" but "which sources does the AI decide to include in its answer." Your keyword research needs to find the inputs to that synthesis, which means mapping not just phrases but the angles, entities, and answer structures the AI prefers. Our guide on what AI visibility is and how to measure it explains why that reframing matters, and why volume-scored keywords are a shrinking share of the demand you're actually competing for.
The method that actually finds them
The process below is field-tested in our AI Visibility Alerts work — here it is in plain steps:
I'm going to give you a process that mixes data with a little field work, because no single tool does this end to end.
Step 1 — Start with what you already know you're asked. Pull your own Search Console queries and your sales team's most repeated live questions. Every inbound "how does this work" is a piece of demand. That alone is a goldmine most teams never mine. On your analytics you'll also find the questions people typed into your own site search box — that's unfiltered, unreported intent, and it's free. Log every sales call question for two weeks. You'll have 30–50 long-tail queries that no keyword tool in the world could have shown you.
Step 2 — Decompose your pillar topic into every angle. Take one service. List every question a buyer has at each stage:
- Problem-aware: "why does my roof leak in the spring," "signs my furnace is oversized"
- Solution-aware: "how much does roof repair cost in [city]," "rafter repair vs truss replacement"
- Comparing: "metal roof vs asphalt for [province] winters," "local studio vs online fitness coaching"
- Ready to buy: "same-day roof patch service [city]," "quote for furnace replacement [neighborhood]"
Write them long-form, naturally, the way a real person would type them. Don't sanitize them into keyword phrases — the whole point is ratcheting up the natural-language coverage, which is what AI engines reward and what classic tools undercount.
Step 3 — Interrogate the AI engines themselves. This is the part people are still embarrassed to admit we do, but it works. Ask ChatGPT, Perplexity, and Google's answer to enumerate what sources they consider for your topic. Ask them what they'd want to see in a trustworthy answer. Ask them what information gaps exist in the current answers. They'll hand you a coverage map in minutes. Do it repeatedly, because the answers shift as the models update — what we call the engine's "answer drift" is exactly the opportunity window where you can be the source they add next. Our guide on how to rank in ChatGPT Search covers how to turn what you learn into a strategy.
Step 4 — Check against the Q&A mines. People Also Ask, Reddit, Quora, and the comments under your competitors' posts are where unpolished demand lives. AI engines train on this stuff, so the phrasing that shows up there is the phrasing AI answers inherit. The strongest research move here is Reddit: search your topic, sort by top and by new, and read the actual questions people ask in defeat before they even know service categories exist. That's your long tail, pre-phrased.
Step 5 — Watch what the AI is not citing. Run your top-ten buyer-intent questions through ChatGPT and Perplexity. Note which sources dominate the answers and which topics get weak, incomplete responses. An answer that reads "there are several factors to consider" without specifics is an invitation — the AI is asking for a better source. That gap is your shot. This is the closest thing to a free AI competitiveness report you can generate.
How to pick which ones to chase
Don't get drunk on volume. Weight by two things:
- Commercial intent. Does this question put money in motion? "Can a dentist rank locally without reviews?" is closer to a buyer than "history of SEO." Map each question to a specific offer you sell. If you can't draw the line between the question and a paid service, it's awareness content at best — which has its place, but it's not your priority.
- Entity fit. Can your business answer it better than the fifty generic roundups? If you have a real opinion or data, it's winnable. If not, the AI will pick someone who does. This is where your experience edge matters — a question your shop can answer from 400 completed projects is an asset; a question that only a national magazine can answer well is not.
A handful of intent-laden questions you can dominate beats a bucket of high-volume terms you'll never crack. That's the entire philosophy: precision over volume, coverage over rank. Our guide on how to get recommended by AI explains the mechanics of becoming the default answer for a specific, defensible query — and it starts with choosing queries you can actually own.
The scoring shortcut: give each candidate question a 1–5 for commercial intent and a 1–5 for entity fit. Multiply. Target the 9–25s first. Everything below a 4 is a "maybe later." Yes, it's that simple, and it prevents you from spending months on questions that were never winnable or never worth winning.
Turn it into coverage, not screenfuls
The output of AI research isn't a spreadsheet of 10,000 keywords. It's a short list of questions you can answer authoritatively, mapped to pages and schema. Every question gets a page that answers it in the first two sentences, states your position, and stays consistent with everything else you publish. If someone lands on a page that answers their exact question with a specific number in the opening lines, that's a page an AI can quote — and an AI only cites what it can cleanly quote.
How to structure a question-led page:
- Answer the question directly in the first sentence (this is the part AI extracts).
- Give the specifics the answer demands — numbers, steps, timelines, or price ranges.
- Add the experience layer — data from your own work, a named process, an honest caveat.
- Map to a related offer or resource — the next logical step in the buyer's journey.
For a deeper dive, see our guide on running an AI visibility audit.
For a deeper dive, see our guide on understanding your AI visibility score.
Do that for sixty questions and you'll have more genuine, extractable coverage than most sites have in total. That's the kind of depth AI engines reward. A site that answers sixty real buyer questions better than anyone else isn't just a site with good keyword coverage — it's a site that becomes a source, which is the only currency that matters in AI search. Our guide on content that ranks in AI search goes deep on this structure, including the specific formats — tables, direct answers, FAQ blocks — that engines extract most reliably.
Track what you can actually see. The tools landscape has moved faster than most teams have adapted. The right approach is a mix: classic ranking platforms for the legacy picture, AI-visibility scanners for the answer-engine picture, and their free tiers as a starting point. Our guide on AI visibility tools: free vs paid lays out which ones are worth money and which free options get you surprisingly far. Whatever you pick, measure the same thing: whether your brand shows up in answers for the questions you've decided to own.
We structure this research for our clients in Montreal and Vancouver every month. If you'd like to see what a month of LLM research turns up for your niche instead of guessing, book a free website teardown and we'll map your first ten opportunities.