LLM search optimization puts an expertise-led business inside AI answers, and the research so far cannot show that a citation produces a customer, so the work has to be measured on inquiries rather than on a vendor’s visibility score.
I sell a report that measures AI visibility. I can’t prove it makes anybody money.
Neither can anyone else selling one. In July a researcher named Olivier Martinez published a critical survey of 45 studies on generative engine optimization going back to November 2023, looking for which tactics actually hold up. His finding, in his words: no reviewed technique shows a stable, longitudinal, cross-platform causal effect on organic discoverability or downstream behavior. On whether any of it reaches the bottom of the funnel, he rates the evidence very low. One suggestive quasi-experiment and a handful of industry claims. It’s a preprint and it hasn’t been through peer review, so treat it as the best available reading rather than a settled one. It is still more rigour than any vendor deck you’ll be shown this year.
So the claim that a citation score predicts clicks, conversions or revenue is not supported. By anybody. Including me.
The question I get asked is whether SEO is dead now. It isn’t. What went missing is the proof, and SEO is taking the blame for it.
Pew Research Center looked at 68,879 Google searches from 900 American adults over a month in 2025. When Google’s AI summary appeared above the results, people clicked a regular search result 8 percent of the time. With no summary, 15 percent. They clicked a source inside the summary itself 1 percent of the time. Around a quarter of them closed the browser after reading it and went and did something else.
That is Google AI Overviews, measured on Google. Nobody has run the same study on ChatGPT or Perplexity, so I’m not going to tell you the number holds there.
What it does tell you is that roughly half the visits you used to get from a page like that stopped arriving, on the subset of searches where a summary shows up. Not all of them. Half of some of them. That’s a dent, not a death, and it’s enough to break what you were using the number for.
The traffic was never the thing. It was the part you could see.
Traffic was a receipt. It told you interest existed because you could watch it arrive. What I can’t tell you is that the interest is still there in the same volume, because click rate doesn’t measure whether anybody got what they needed. It measures whether they came to your house to get it.
So the industry sold a replacement receipt. A citation score. How often a model names you across a set of prompts somebody else chose.
Martinez is clear that a high citation rate among the documents that get retrieved does nothing for you if you’re never retrieved at all. Retrieval comes first, and retrieval is close to binary. Either the model can reach a usable description of your business or it can’t. That part is checkable today, it has a yes or no answer, and it is the only part of this that isn’t an approximation.
That’s what a visibility check is for. Not predicting revenue. Finding out whether you’re in the reachable set, and which specific thing in your own material is keeping you out of it. Usually that’s inconsistent business information across the web and nobody else describing you, which is the same problem that makes a model recommend your competitor when someone asks for a business like yours.
Past that point, stop shopping for a number and go collect your own.
Ask people how they found you. On the intake form, in the first five minutes of the call, out loud. An agency owner in an online forum pointed out that people don’t volunteer ChatGPT yet. They say “online” unless you probe. So probe. Ask what they typed and what came back.
That data is also unreliable. People misremember, and some of them will give you the easy answer to make the question go away. The difference is that it’s unreliable about your actual buyers rather than reliable about an average of somebody’s prompt list, and you can ask the next one a better question.
Then watch the shape of what arrives, not the volume. Someone who turns up already knowing roughly what you charge and roughly how you work got briefed by something before they reached you. That won’t show up on any visibility dashboard, because it already happened somewhere a dashboard can’t see.
And notice where your evidence is coming from. The Pew numbers are American, and so is nearly every AI visibility statistic being quoted at Canadian business owners this year. Your buyers may behave differently. No report is going to tell you that.
If you want one number to run the quarter on, count the inquiries where the person already knew something true about you before you said it.