The prompts your team optimized against were made up
AI visibility strategies were built on guesses dressed up as data. Actual behavior changes what gets built and what gets cut.
Jonny Bentwood is an advisor, Ragan’s Center for AI Strategy, and global head of data and analytics, Golin Ketchum.
About 10 times a week, someone on my team gets asked the same question by a client. What are people actually asking ChatGPT about us, and how do we show up in the answer? It is a fair question. It is also one we could not answer honestly until recently.
Most generative engine optimization (GEO) has been built on guesswork. We sit in a room, we use our judgement, we write a list of prompts we think people might type, and sometimes we lean on Google Trends as a stand-in. Then we optimize against that list and report back with a straight face. The problem is obvious once you say it out loud. If the list is a guess, everything downstream is a guess measured to two decimal places.
There is a second flaw, and it is worse because nobody mentions it: We treat every prompt as if it matters equally. A question 10,000 people ask sits on the same line as one that 40 people ask. So we spread effort evenly across things that are not evenly important and we call it a strategy.
The list you optimized against was a guess
Over the past few months we have been fixing this with Similarweb. Instead of guessing what people ask AI, we read millions of the real conversations people are already having with it.
We cluster those conversations into topics, and we weight each topic by how much of the actual conversation it commands. For the first time, we can see what people genuinely bring to AI, and how much any of it matters.
This is better than a focus group. People tell AI things they would never tell a moderator or answer on a survey.
They ask it what is actually on their mind, their hidden agenda and their secret truth. They think out loud. They reveal the worry behind the question rather than the tidy version they would offer a stranger with a clipboard. At scale, those conversations are the most honest customer research I have seen in two decades in this business.
Real conversations reveal what audiences actually bring to AI
This changes more than measurement. It changes what we make. Once you know the handful of topics your audience actually cares about, you stop producing content for the ones you assumed they cared about. You build a content strategy around real demand rather than internal opinion, and you point your visibility efforts at the questions worth winning.
I am wary of big claims in this industry, because we make far too many of them. But I will make this one. Moving from proxies to observed behavior is the biggest leap in audience intelligence I have seen in decades. Not because the technology is clever, though it is, but because it removes the guess that was underneath everything we did.
Build against real demand, not internal opinion
We spent years getting very precise at measuring the wrong things. Now we can start with what people actually say, decide what matters and work on that. It is a simpler way to work.
It is also at last an honest one.
