Last month I sat on a call with a founder who was genuinely thrilled. His agency had just gotten him to position one on Google for a keyword with 1,900 monthly searches. Champagne emoji in the Slack channel, the works. So I asked him one thing: when did you last check what ChatGPT says when someone asks it for tools in your category? Silence. He never had. And his buyers ask that exact question every day, in a place his five-figure keyword strategy can't see.
That call is why I'm writing this. Keyword research, the discipline of picking what to publish based on a search volume column, is running on fumes. The replacement already exists. It just doesn't have a volume column, which is exactly why most marketers haven't adopted it yet. It's called prompt research, and if you sell software, you're late.
The Dial Reads Calm While the Ground Splits
Let's start with what the instruments say, because the numbers are genuinely wild.
Gartner projected traditional search engine volume would fall 25% by 2026 as people shift their questions to AI assistants. That prediction looked aggressive two years ago. Now it looks conservative. AI Overviews currently appear on roughly a quarter of all Google searches, close to double their share from March 2025. And when an AI Overview shows up, about 83% of those searches end without a single click. Even across all searches, nearly 59% of US queries now end clickless.
Meanwhile, the other side of the ledger: ChatGPT crossed 900 million weekly users this year, up from 400 million in early 2025, and processes about 2.5 billion prompts every day. Perplexity alone handled 780 million queries in a single month. During the last holiday season, AI referral traffic to US retail sites grew 693% year over year.
Here's the uncomfortable part. Your keyword tool registers none of it. Search volume tells you what people typed into Google a month or three ago, rounded to a bucket. It says nothing about the 2.5 billion daily questions flowing through a chat window. The demand didn't shrink. It moved somewhere your dial can't reach, and the needle just keeps drawing a flat, reassuring line. We covered where all this is heading in our 2026-2027 AI search outlook, and the short version is: the gap widens from here.
Why No Tool Can See Prompts (Including Ours)
Real talk, because this is the part the industry doesn't love saying out loud.
Keyword tools know Google volumes because of clickstream data. Panels of millions of browsers report the URLs they visit, and a Google search URL carries the query right in it. Parse the URL, count the occurrences, sell the number. That's the whole trick, and it worked for fifteen years.
Now look at a ChatGPT conversation URL. It's an opaque token, a random string pointing at a private conversation. The prompt never appears in the URL. It never leaves OpenAI's servers. There is nothing to parse, no matter how big your panel is. Same story with Claude, Gemini, and Perplexity threads.
Which means every product currently marketed as an "AI search volume tool" is doing estimation stacked on inference. Some of those estimates are thoughtful. None of them are measurements. Until the model companies expose prompt logs, and I wouldn't hold my breath, exact prompt volume is unknowable. Anyone quoting you a precise number is selling confidence, not data.
I'll be blunt about our own product here: RankControl can't tell you prompt volume either. Nobody can. What can be measured, reliably and repeatedly, is what the engines actually say: which prompts get your brand named, which get your competitor named, and how that shifts over time. That distinction shapes everything else in this piece.
A Prompt Is a Different Species Than a Keyword
A keyword looks like this: "crm small business".
A prompt looks like this: "I run a 6-person lawn-care company, we're drowning in spreadsheets, what CRM should I use that my crew can learn in a weekend without hiring a consultant?"
Those are different animals. The keyword is a fragment you decode. The prompt carries the team size, the pain, the constraint, and the buying criteria in one breath. People talk to AI engines the way they'd talk to a sharp friend over coffee, and then they follow up four more times in the same thread. The conversation refines itself toward a shortlist while your analytics see nothing.
And under the hood it gets stranger. When an engine receives that lawn-care prompt, it doesn't search it as-is. It fans the prompt out into dozens of hidden sub-queries: best simple CRMs, CRM onboarding time, CRM pricing for small teams, alternatives to whatever brand it already suspects. It retrieves against those synthetic queries, then synthesizes one answer. So the query that surfaced your content was machine-generated, invisible, and absent from every keyword database on earth. You didn't rank for a keyword. You got retrieved for a question no human ever typed.
One more wrinkle. Run the identical prompt across ChatGPT, Gemini, and Perplexity, and you'll get three different answers with partial overlap. Run it again next week and the shortlist reshuffles. Across our tracking, the same prompt routinely names different brands on different days. A single screenshot proves nothing. Answers are distributions, and you have to sample them like distributions.
200+ SaaS teams already track their AI citations.
They know exactly when ChatGPT mentions their brand, and when it stops. Do you?
See plans→
OK, Backing Up a Step Before Someone Yells at Me
I can already hear the objection: "keyword research still works, my traffic says so." Fair. Let me be precise about what died and what didn't.
Keyword research as a validation layer is fine. It confirms a market exists, shows the vocabulary people use, and still runs your paid campaigns. Google hasn't vanished, and pages that rank top three on Google get cited by AI engines far more often than pages buried on page two. Your old rankings are an asset in the new system.
What died is search volume as the north star. The workflow where a content calendar is a spreadsheet sorted by volume descending, and anything under 500 monthly searches gets deleted as "not worth it". That workflow now systematically deletes your best opportunities, because the highest-intent questions in your category read as zero volume in every tool while getting asked constantly inside AI engines.
The practitioners quietly agree, by the way. Spend an hour in SEO communities and the emerging consensus sounds like this: keywords stopped being the strategy and became a validation layer. You start from customer questions and intent, then use keyword data to confirm demand and learn the phrasing. The volume column moved from the driver's seat to the glove box. Even the people who built careers on keyword research describe their current process as question research with a keyword check at the end, which is a polite way of conceding the argument.
The economics make this sting. AI search visitors convert at roughly 4.4 times the rate of traditional organic visitors. We've watched companies get over a tenth of their signups from AI referrals that made up less than 1% of their traffic. One hyper-specific question with "no search volume" can quietly outperform the 10,000-volume head term you spent six months chasing. Volume was always a proxy for demand. The proxy broke. The demand is fine.
How to Actually Do Prompt Research
So what replaces the spreadsheet? Question intelligence, gathered from places that were always available and mostly ignored. Here's the working process.
Mine Search Console for the messy queries. Filter your queries to anything over 30 characters. Those long, awkward, natural-language searches are the closest free proxy to prompts, because the same person phrasing a Google search like a sentence is phrasing ChatGPT prompts the same way. This is twenty minutes of work and almost nobody does it.
Raid your own customer conversations. Sales calls, support tickets, onboarding questions. The exact sentence a prospect says on a demo call is the prompt a stranger types into ChatGPT next month. Your CRM is sitting on a prompt database and calling it call notes.
Sit where the questions form. The threads where your buyers ask for recommendations, the People Also Ask boxes, and the follow-up questions Perplexity and ChatGPT suggest at the end of an answer. Those suggested follow-ups are the engine telling you, for free, what it thinks people ask next.
Build a prompt panel and run it on a schedule. Write down 20 to 50 questions a real buyer in your category would ask, from broad ("best tools for X") to brutally specific (the lawn-care CRM phrasing). Run them across the major engines weekly. Log who gets named, who gets linked, and how the answers drift. Remember the distribution point: sample repeatedly, never trust one run.
Cluster, don't chase. If you get access to fan-out style query data, group those sub-queries into topics and build one strong page per topic. Building a thin page for every query variant is the fastest way to recreate 2015-era SEO spam with a new coat of paint. Our AEO audit playbook covers how to structure the pages that come out of this process.
Full disclosure on cost: the setup runs 10 to 15 hours, and the weekly panel runs eat a few hours forever. That's the honest price of doing this by hand.
How often does ChatGPT mention your brand?
Most founders have no idea. The answer might surprise you.
Find out now→
The Metric That Replaces Search Volume
Once you have a prompt panel, a new number falls out of it, and it's the one I'd actually run a content strategy on: of the questions that matter in your category, what share of the answers name you?
That's share of voice in AI search, and it behaves like a market-share number rather than a traffic number. It tells you where competitors own answers you should own, which prompts nobody has claimed yet, and whether last month's content actually changed what the engines say. Our AI visibility tracking exists because this number moves constantly: models retrain, sources reshuffle, and a shortlist you sat on comfortably in February can drop you by May without a single ranking changing in Google.
Which is the real argument against one-off research of any kind. The question was never "what are the keywords". The question is "what are buyers asking right now, and who's in the answer". That question needs re-asking every week.
This Is Literally Why Radar Exists
Here's where I tell you what we built, briefly, because the whole article has been describing it sideways.
RankControl's Radar agent is prompt research running as a service. It scans six AI engines continuously, surfaces the high-opportunity buyer questions in your category that competitors haven't answered, and hands them off as briefs. A typical week turns up dozens of them. The content engine then writes the deep answers to the questions Radar found, published on your own domain. Discovery feeds creation, creation feeds citations, and the loop repeats monthly without you babysitting it.
You can absolutely run this whole discipline yourself with the process above: the GSC mining, the customer-call transcripts, the weekly panel runs across four engines, the clustering, the logging. Or Radar and its sibling agents do it every week while you build product. Plans start at $499/mo, which is less than the hourly math on 15 hours of setup plus every Monday morning forever.
The Spreadsheet Was Never the Point
Keyword research had a great twenty-year run, and it earned it. It was the best available answer to the only question that matters in content: what do buyers want to know? For two decades, a volume column was a decent proxy for that.
The proxy broke. The question didn't. Buyers want to know the same things they always did. They just ask a machine that answers in full sentences, keeps the conversation private, and names three brands at the end. Your job is to be one of those three names, and no amount of staring at a volume column will get you there.
The founders who switch to prompt research this year get a compounding head start: every week of panel data makes the picture sharper, while their competitors keep sorting spreadsheets by a metric that stopped describing reality. Dead disciplines don't announce themselves. They just quietly stop correlating with revenue. This one already has.
15 hours a month manually. Or 15 minutes with RankControl.
Track citations, monitor competitors, and fix content gaps across every AI search engine. Automatically.
See how it works→
