How to Do Keyword Research for AI Search in 2026

The honest answerDoes keyword research still matter for AI search?

Yes. But the goal changed, and Google itself drew the map. In its May 2026 AI optimization guide, Google states plainly that "from Google Search's perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO." AI Overviews and AI Mode run on the same crawl, the same index, and the same ranking systems as classic Search. Google is equally direct in its AI features documentation: "There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary." No llms.txt, no special schema, no AI file.

Here is the practical nuance Google leaves out. "It is still SEO" does not mean the research is the same. AI engines draw their answers from pages that already rank and clearly answer specific questions, then blend several sources into one response. So the work splits in two: you still need content that is crawlable, indexed, and authoritative, and you now need to research the exact questions and entities an answer should cover. The first gets you eligible. The second gets you cited. If you want the full picture of how the two connect, that is the core of how we approach AI visibility.

What actually changedHow is keyword research different for AI vs Google?

Five things changed, and they compound.

Query fan-out is the big one. Google's own documentation confirms that 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 prompt in, a dozen sub-searches out, synthesized into a single answer. Ahrefs analyzed 173,020 URLs and found you are 161% more likely to be cited in Google's AI Overviews if you also rank for the fan-out queries, not just the head term. Ranking number one for your main keyword no longer guarantees you show up at all.

Prompts are not keywords. A keyword is three or four words. A prompt is a paragraph with context, constraints, and often a competitor's name in it. And the engine rewrites it before searching. Profound, in an analysis of ChatGPT's behavior reported by Nick Lafferty, found 91% of ChatGPT's internal search queries were unique, with only 13% word overlap with what the user actually typed. The phrase you optimize for and the query the engine runs are different objects.

The unit of research is no longer the keyword. It is the question - and every sub-question the AI invents on its way to answering it.

Zero-volume keywords earn citations. Google has said for years that about 15% of daily searches are brand new and have never been seen before - a figure John Mueller reaffirmed in 2026. If a query returns zero volume in your tool, that often means the tool cannot measure it, not that nobody asks it. Those hyper-specific questions are exactly what AI systems get and cite.

The long tail got longer. Clearscope calls this the "long long tail" - the shift from being found to being understood. Where the long tail was about matching a specific phrase, the long-long-tail is about being one of the voices an engine references when it builds an answer across a topic.

Share of voice is the new rank. Being cited is measurably valuable. Seer Interactive, tracking 3,119 informational search terms across 42 organizations, found that brands cited inside an AI Overview earned 35% more organic clicks and 91% more paid clicks than uncited brands on the same queries (organic CTR of 0.70% versus 0.52%). Being in the answer beats ranking below it. That is why earning citations is now its own discipline.

Step 2 of the processHow to find the questions people ask AI

This is the single highest-return activity in AI-era research, and most of it is free. Pew's data shows why: question-based searches (who, what, why) triggered an AI summary 60% of the time, and 53% of ten-plus-word searches did, versus just 8% of one or two-word searches. Long, conversational, question-shaped queries are where AI intercepts demand. Go get them:

• People Also Ask and related searches - mine them for the exact questions Google already clusters around your topic.
• Reddit, Quora, and niche forums - people ask questions there in their own words, not keyword-optimized phrasing. This is raw prompt language.
• Your sales and support logs - the questions prospects ask on discovery calls map almost exactly to how they phrase a ChatGPT query.
• Google Search Console, filtered by length - apply a regex filter to your Performance report to surface queries of ten or more words. As Similarweb notes, these are queries where someone typed a real question and your page already earned an impression. That is a prompt candidate handed to you for free.
• The AI tools themselves - enter a seed topic into ChatGPT or Perplexity and read the sub-queries it generates while researching. That is fan-out you can observe directly.

The awkward truth about volumeWhy traditional search volume misleads you now

Traditional monthly search volume measures one thing: how many people typed a specific string into a Google box. It cannot see conversational demand inside ChatGPT, and it under-reports the long tail because tools report twelve-month averages that dampen emerging topics to zero.

The industry response has been "prompt volume" - modeled estimates of how often people ask AI about a topic. Use it, but know what it is. As Similarweb and Semrush have both acknowledged to Digiday, this is synthetic, clustered data - they group anonymized user prompts into topics and normalize a panel, so the number is directional, not a real count. Semrush president Eugene Levin put it bluntly: "We don't really go and check any one specific prompt, because there's no value in that. Everyone asks different questions differently." Some practitioners are harsher. FatJoe's team argues in "Tracking Prompt Volume - The New Mirage Metric" that "a high prompt count doesn't mean there's traffic or value behind it."

The tools, honestlyWhich AI keyword research tools are worth it?

The AI visibility market is crowded and well-funded - Profound alone raised a $96M Series C in February 2026 at a $1 billion valuation. That does not mean you need to spend a fortune. Here is the honest breakdown.

Start free. Google Search Console now reports AI Overview impressions and, with a length regex, surfaces your conversational queries. Google Trends is the most underrated tool for velocity: its "Rising" and "Breakout" queries (Breakout means growth over 5,000%) and its ten-minute Trending Now refresh let you catch demand while it is building, before it ever registers in a keyword database. AlsoAsked and AnswerThePublic mine People Also Ask and question phrasing cheaply.

Bolt-on if you already pay for a suite. Ahrefs Brand Radar (built on a 240M-plus prompt database, with the advantage of historical data most rivals lack) and the Semrush AI Visibility Toolkit (a $99/month add-on) are good enough to start and live inside tools you already use. Be aware both have been caught confusing a brand name for a common word, and neither reports zero-click rate per keyword - a real gap for prioritizing citation plays. Clearscope's Topic Explorer is strong for mapping the entities and questions around a topic.

Dedicated tools if AI is a core channel. Profound leads on depth (real prompt-level querying across 10-plus engines and its own conversation dataset). Peec AI is the mid-market favorite for clean reporting and agency-friendly dashboards; Otterly is the cheapest credible entry point at $29/month. Scrunch and Similarweb serve specific niches - crawler behavior and market-level prompt data respectively. The honest caveat: most of these are measurement, not action. A dashboard that watches your number go up and down is not the same as work that makes it go up. Buy one only if someone will act on the data.

The playbookA step-by-step process you can run this week

Here is the workflow we use. It works for a solo founder with a spreadsheet or a team with a full stack.

1. Seed from your buyers, not your tools. List the real problems your customers are trying to solve, in their words. Pull ten to twenty phrases from sales calls, support tickets, and onboarding questions. This is your ground truth - tools expand it, they do not replace it.
2. Mine the questions. Run each seed through People Also Ask, Reddit, Quora, AlsoAsked, and your Search Console ten-word regex. Collect the full set of questions around each topic, not just the head term. Aim for a messy, comprehensive list.
3. Estimate demand directionally. Keep traditional volume for the head terms where it exists. For the long tail and conversational queries, use prompt-volume tools as a directional signal and your own judgment on business relevance. A zero-volume question from a real buyer beats a high-volume term with no intent.
4. Map the fan-out. For your priority topics, generate the sub-queries an AI would fire - either by reading them in a tool like Brand Radar or Profound, or simply by asking ChatGPT your seed question and watching it research. Do not try to write a page for each sub-query. Note which ones your existing content already answers, and where the gaps are.
5. Plan the cluster, not the page. Build one authoritative hub plus supporting pages that answer the sub-questions directly, each with a 40-to-60-word answer near the top. Cover the topic deeply enough that the engine can pull from you no matter which branch of the fan-out fires. This is the long-long-tail in practice.
6. Track share of voice by prompt. Pick your ten to twenty most important prompts. Record whether each one cites you, a competitor, or neither today. Set a 90-day target to move from zero to a top-three cited source on a handful of them. Report citations and mentions - not rank.

Getting citedWhat the research says actually works

Once you know the questions, the content itself has to earn the citation. The best evidence here is the Princeton and IIT Delhi study "GEO: Generative Engine Optimization" (Aggarwal et al., KDD 2024), which tested nine content tactics across a 10,000-query benchmark spanning 25 domains. The headline: their methods boosted visibility in generative-engine responses "by up to 40%," with the best single method improving the baseline by 41% on the paper's Position-Adjusted Word Count metric.

What worked: adding relevant quotations, citing sources, and adding statistics were the top three tactics, each delivering a 30-to-40% relative improvement. Fluency and clarity edits helped too. What failed: keyword stuffing actually scored below the do-nothing baseline. In the authors' words, "techniques effective in search engines may not translate to success in this new paradigm." The study also found effectiveness varies by domain - citations win for factual and legal topics, quotations for narrative and historical ones - which is one more reason to research your specific space rather than copy a generic checklist. If Google AI Overviews are your priority surface, our AI Overviews optimization approach is built around exactly these signals.

One reality check on scale before you rebuild everything: Google is not gone. Ahrefs' Patrick Stox estimated in early 2026 that search-like interactions on ChatGPT are roughly 12% of Google's volume. The move is real and accelerating, but the right play is to research for both surfaces at once - which, conveniently, the process above does.