The question you get measured on decides the answer you see. Most “AI visibility” tools test your business against prompts built from Google search data — keyword databases and People Also Ask. But people don't talk to AI the way they type into Google. Test the wrong question, and you optimize for a customer who was never going to ask it.
People search Google in fragments and ask AI in full sentences. Into a search box you type best crm small team — three keywords, no context, because you're about to scan ten links and decide for yourself. To an assistant you say, in your own voice, “We're a five-person agency switching off spreadsheets — what CRM should we actually use?” You expect one answer, tailored to you. Different intent, different grammar, different surface.
Proof: in a study of 15,000 prompts across four assistants, only about 8–12% of the links ChatGPT, Gemini and Copilot cite appear in Google's top 10 results for the same prompt. Tellingly, this comes from Ahrefs' own research — the engines simply aren't reading from the same shortlist as the search page.
Bottom line: optimizing for keywords optimizes for the old surface. AI is a new one, with its own language.
Tools built on a search index generate AI prompts from Google behavior — because that's the data they have. A keyword-and-backlink platform like Ahrefs is, at its core, a giant index of how people use Google: search volumes, keyword variants, People Also Ask. Turning that into “AI prompts” is convenient, but it inherits a hidden assumption — that people ask AI the same way they query Google. The overlap figure above — from Ahrefs' own data — is exactly the evidence that they don't.
The catch: a keyword-derived prompt is a proxy for Google demand, not a sample of how anyone actually talks to an assistant.
The questions that decide AI recommendations are first-person, conversational and specific to a situation. They carry context a keyword never does — company size, industry, constraint, the job to be done:
Notice how each one already narrows to a shortlist — and how none of them would ever appear in a keyword report. This is where your business is won or skipped.
Proof: this isn't a hunch. Academic and industry analyses consistently find AI prompts are longer and more conversational than search queries — full sentences with context, where Google gets three-word fragments. See the comparative study ChatGPT vs. Google (arXiv) and Profound's gap analysis.
Reflexa generates the questions from your business, not from Google's keyword ghosts. The engine reads what you do, who you serve and the category you compete in, then writes the buying questions a real customer would put to an assistant — first-person, conversational, situation-specific — and asks the engines those. You get measured on the conversations that actually decide your sales, and you see, per question and per engine, whether you're named, cited or invisible.
Proof: that's what runs on every audit. The free check covers a starter set of buying questions; paid plans widen the set — 25 on Starter, 100 on Growth — so more of your real buying intent is covered.
Why it matters: a narrow, well-aimed set of the right questions tells you more about your AI visibility than a thousand keyword-shaped prompts about the wrong one.
Neither approach can read the engines' minds — and we won't pretend to. No tool sees the full distribution of what every user asks ChatGPT; the engines don't publish it. What we can do is stop guessing from Google and start from your business, generate questions the way your buyers phrase them, and report only what we can actually observe in each engine — with the evidence attached. That's a sharper aim at a narrow target, not a crystal ball, and we'd rather be honest about which.
Do this: when you compare AI-visibility tools, ask one question — where do their questions come from? Google's keywords, or your business?
Sources: Ahrefs — AI search vs Google overlap (15,000 prompts) · Profound — the ChatGPT vs Google gap · arXiv — ChatGPT vs. Google comparative study
The free check runs them on your domain — 3 minutes, evidence included.