Keyword research for AI search means researching two lists instead of one: the long questions people ask about your category, and the shorter searches AI engines run behind the scenes to answer them. Volume-based keyword tools only see the second kind, and only partly. The method below uses Search Console, your own customers and the engines themselves to fill in the rest.
Here is the problem in one test. Open Keyword Planner and type marketing dashboard. You get a monthly volume, a competition level and a suggested bid. Now type something closer to what an agency owner asks AI Mode today: which dashboard can I share with clients that pulls Google Ads and Meta without copying numbers into slides every Monday. Most likely you get nothing. That person exists, and so do thousands like them asking the same thing in their own words. The tool shows nothing because it counts identical strings, and nobody writes the same sentence twice.
What changed in the query, according to Google
Queries got longer, more specific and more tied to a situation, and Google has the numbers. On May 19, 2026, it published a report on its first year of AI Mode in the US, signed by Shivani Mohan, VP of Data Science and UX Research for Search.
- AI Mode passed one billion monthly users globally, and its query volume more than doubled every quarter since launch.
- Planning queries grew 80% faster than AI Mode overall over six months. Brainstorming queries grew 30% faster.
- Searches starting with where should I, where to and ideas for are growing.
Google had already framed it in December 2025, in its year-end search trends piece for advertisers on Think with Google: search is moving from keywords to conversations. And the 15% figure is older than AI Mode. Google first reported it in 2019, and in March 2025 John Mueller said it still holds.
Put together: a large share of queries are long, contextual ("I run a five-person agency and...") and have never been typed before. A keyword list built around best marketing dashboard describes a small and shrinking slice of the demand.
Why the usual tools miss it
The tools still work for the short, high-volume head of demand. They go quiet in the part of search that is growing, for three reasons, and none of them is that the tools are bad.
Volume is counted per string. Keyword Planner, Semrush, Ahrefs and the rest estimate how many times a given phrase is searched per month. For a 20-word question the real count is close to one, so the tool rounds it to zero or doesn't list it. Keyword Planner also merges close variants and, for accounts with low ad spend, shows ranges instead of numbers, so even mid-length queries come back blurry.
Search Console hides the long tail on purpose. Google anonymizes queries that few people search, to protect privacy. Ahrefs measured the effect across 887,534 properties and 22 billion clicks: in April 2025, 46.77% of clicks came from queries Search Console doesn't show. The longer and rarer the question, the more likely it falls in that hidden half.
Google's AI report shows pages, not questions. Since August 31, 2026, every site has a Generative AI performance report in Search Console, with impressions in AI Overviews and AI Mode by page, country, date and device. It has no query dimension: you can see that a page appeared in AI answers, but not what people asked to get there. AI Mode data has also counted toward the regular Performance report since June 2025, mixed into the Web search type, and each follow-up question is logged as a separate query.
What people ask vs. what the engine searches
When someone asks ChatGPT, Perplexity or AI Mode something that needs current information, the engine doesn't send that sentence to an index as is. It rewrites it into one or more shorter searches, called fan-out queries, retrieves pages for each, and writes the answer from what comes back. We go through the full retrieval, citation and mention chain in our guide to GEO and AEO.
That leaves you with two lists to research, and they look nothing alike.
- The prompt is the person's question: long, personal, full of context. It is what your page has to answer.
- The fan-out is what the engine actually searched to answer it. It is what your page has to rank for so the engine reads it.
Peec AI analyzed more than 20 million ChatGPT fan-outs between October 2025 and January 2026: their average length went from about 6 words to about 12, while the number of searches per prompt stayed between two and three. In a separate sample of about 5 million fan-outs from April 2026, the averages were 1.4 searches per prompt on Perplexity, 2.1 on ChatGPT and 6.8 on Grok.
The fan-out is the closest thing AI search has to a keyword list. It is shorter than the prompt, it repeats across different people asking similar things, and it decides whether your page gets retrieved in the first place. If your page doesn't show up for the fan-out, the best answer in your niche never makes it into the model's context, no matter how well it is written. That is also why traditional SEO is still the base.
Peec found that 43% of the fan-out steps triggered by non-English prompts ran in English. A question asked in Spanish or German may be answered with pages found through an English search.
A method that works with long queries
1. Start with the long queries you already get
Search Console doesn't show everything, but what it shows is real. In the Performance report, add a Query filter, choose Custom (regex) and try two patterns:
- Questions:
^(who|what|when|where|why|how|which|can|does|do|is|are|should)\b - Five words or more:
^(\S+\s+){4,}\S+$
Sort by impressions, not clicks. A long query with impressions and almost no clicks means Google considers your page relevant to it, and the page isn't answering it well enough to win the click. The interface caps tables at 1,000 rows; if you have more, pull the data through the API or a Looker Studio connection.
Two newer features help. If your site has enough volume, the Insights report shows Query groups, launched in October 2025: Google clusters similar queries with AI and flags which groups are trending up or down. And the branded queries filter, introduced in November 2025, separates people who already know your name from people asking about the problem.
2. Group by task, not by keyword
A list of 300 long queries is useless as a list. What you need is the handful of jobs behind them. Google's report points at the categories that are growing: planning, brainstorming, deciding (which tool should I..., where should I...) and getting something done.
For a reporting tool, how do I show clients ROAS across Google and Meta, client report template for agencies and why do Meta and Google report different conversions are three different jobs: build a report, find a template, understand a discrepancy. Each deserves its own page. Twenty rewordings of how do I build a client report are one job and get one page.
3. Collect the phrasing from people, not from tools
The way customers describe their problem lives in places no keyword tool indexes: sales call notes, support tickets, chat transcripts, your site search, onboarding forms, Reddit threads in your niche, G2 and Capterra reviews, the People also ask box. Copy the sentences literally.
What you're looking for is the context people add: for a store under $50k a month, without a developer, for a team of three, that my client can open without a login. That context is what makes AI queries long, and your page needs to address it explicitly, in a sentence the engine can lift out on its own.
4. Capture the fan-out
Take 20 to 30 of those real questions and run them in the engines you care about. Do it several times, on different days, because both the answers and the searches behind them vary. For each run, log three things: the searches the engine ran, the pages it cited, and whether your brand appeared.
Some interfaces show part of the fan-out: Perplexity lists its steps, and ChatGPT sometimes shows what it searched. Browser extensions built for this, our free GEO Observer among them, record the queries and cited sources automatically and export them.
However you collect them, a pattern shows up quickly: the same few sub-queries come back across different prompts. Those recurring fan-outs are your new head terms. Unlike the prompts, many of them have measurable search volume, so the classic tools become useful again for that layer.
5. Prioritize without search volume
For the long layer, monthly volume doesn't exist or can't be trusted. Rank each task with three signals instead:
- Recurrence: how many of your test prompts produced the same fan-out, or cited the same competitor page.
- Trend in Search Console: whether impressions for that query group grow quarter over quarter, even if each query is tiny.
- Business value: how close the task is to a purchase. A which tool should I use for... question is worth more than a what is... question, even at a fraction of the volume.
What not to do
Don't create one page per long query. Twenty near-identical pages answering variations of the same question is what Google's spam policies call scaled content abuse, and an AI engine doesn't need twenty pages to find one answer. One page per task, with sections that make sense on their own, covers the variations.
Don't invent FAQs. A list of questions nobody asks, written to catch long queries, reads exactly like what it is. Use the questions you collected in step 3, in your customers' words.
And don't throw away the short keywords. Fan-outs are short and run against a classic index. The page that ranks for client reporting dashboard in organic search is the one most likely to be retrieved when an engine searches that phrase on someone's behalf.
How to tell if it's working
Give each new page six to eight weeks, then look at four things together:
- Impressions from long and question-style queries on that page, using the same regex filters in Search Console.
- Impressions for that page in the Generative AI performance report.
- The split between branded and non-branded clicks. Growth on the non-branded side means you're reaching people who didn't know you.
- Whether your brand starts appearing when you re-run the same test prompts from step 4.
None of these proves much alone. Together, checked against the same set of prompts every month, they tell you whether you're showing up for the questions people actually ask, and not only for the keywords a tool can count.
Figures come from Google, Ahrefs and Peec AI publications between March 2025 and September 2026 and may change. The Keyword Planner test and the agency and reporting-tool queries are illustrative examples, not customer data.