The problem isn't volume. It's invisibility.
We've been doing keyword research for 15 years. We have tools, lists, clusters, volumes, difficulty scores. And yet, in 2026, some of the questions our clients ask every day don't show up in any tool. Not Semrush. Not Ahrefs. Not Google Keyword Planner.
Not because the tools are bad. Because people don't search the same way anymore. When someone opens ChatGPT and asks "what digital marketing agency in Romania would be good for a SaaS startup?", that doesn't generate a search query Google can log. It's a private conversation with a language model. And we, as an agency, have no idea it happened.
Traditional keyword research relies on a simple principle: if enough people search for something, a search volume appears. But AI search works differently. A recent Search Console analysis shows that 57.7% of impressions fall into the anonymized query category. More than half the searches generating impressions don't appear in your reports at all.
And that's just for traditional searches. Questions asked directly in ChatGPT, Claude, Gemini, or Perplexity don't even pass through Google. They're complete conversations happening outside the traditional search ecosystem. We've written before about how audiences now search in full sentences. What we didn't say then is that many of those sentences never reach a search engine. They go straight to AI.
Where the real questions actually live
If they're not in your tools, where do you find them? The answer is simpler than it seems: in the conversations you're already having with clients.
Sales calls are a gold mine. Every discovery call contains 3-5 questions the prospect has probably also asked an AI assistant before calling you. "How do I choose between SEO and PPC on a 2,000 EUR budget?", "Is content marketing worth it for a local business?", "What realistic ROI can I expect from social media?"
The same questions appear in support tickets, niche forums, Reddit threads, Facebook and LinkedIn groups. People phrase these questions naturally, exactly the way they'd ask an AI. They don't type short keywords like "SEO agency Bucharest" or "Google Ads pricing". They ask full questions with context, nuance, and their specific situation baked in.
What we do concretely: we run call transcripts through Claude and extract question patterns. We group them, count them, and see which ones appear across multiple conversations. The frequency of a question in sales calls is a far more real demand signal than an estimated search volume from a third-party tool.
A concrete example from our practice: we found that the question "how do I know if my marketing agency is actually delivering results?" appeared in various forms in 4 out of 10 sales calls. Zero search volume in Semrush. But when we created content that directly answers this question, that content started appearing in Gemini and ChatGPT responses for similar queries. Not because we optimized for a keyword, but because we answered a real question that real people actually ask.
Search Console helps. But with serious limits.
Yes, Search Console can filter long or question-format queries using regex. But as mentioned, most impressions are anonymized. Search Console sees AI, but not always the clicks that actually matter.
Google's Gemini API offers an interesting alternative: the webSearchQueries field in the Grounding with Google Search function shows the sub-searches the model runs internally. It's not an exact record of what AI Mode does, but it's a useful approximation. Think of it as a coverage checklist: if Gemini searches for X, Y, Z when answering a question, do you have relevant content for each?
Combining Search Console, call transcripts, and the Gemini API gives you a much more complete picture than any keyword tool used alone. Not perfect, but substantially closer to reality.
You don't need more content. You need better content.
The natural instinct when you discover new questions is to create new pages. Resist the urge. Most sites already have dozens or hundreds of blog posts. The issue isn't quantity. It's how existing content responds to questions.
We've analyzed hundreds of these question clusters over the past year, both for clients and for our own site, and in most cases, the answer was already there. It just wasn't phrased the way the customer asks. The page talked about the service from the provider's perspective, not from the buyer's perspective of having a specific problem to solve.
And this is the difference that matters for AI: a language model looks for answers to specific questions, not pages optimized for generic keywords. If your SEO page talks about "professional SEO services" but doesn't answer "how do I know if SEO is working for my business?", the AI will cite someone else who does.
We've written before about how the keyword universe was always smaller than we thought. The keyword universe was always smaller than we thought. But the question universe is much larger. The difference is you don't need a page for every question. You need to make sure existing pages answer these questions clearly, directly, and with structure.
Three things that work in practice:
- Add FAQ sections to pages with strong organic traffic, using questions extracted from calls
- Rephrase existing paragraphs as direct answers to specific questions
- Use schema markup that helps AI identify answers (not just for rich results, but for AI citations too)
What to do Monday morning
Put a simple process in place. This isn't a 6-month project. It's a mindset shift that takes a few hours per month:
- Monthly, go through sales and support call transcripts. Extract recurring questions using AI.
- Compare the list with existing site content. Where are the gaps?
- Don't auto-create new pages. First, improve existing pages with direct answers to these questions.
- Test with the Gemini API whether your content appears in generated answers for the questions you identified.
Keyword research isn't dead. But it's stuck in a search model that no longer covers the full picture. The questions your customers ask AI are out there. You're just not looking in the right places.





