Everyone has a theory about AI search. Almost nobody has a number.
So I went back through the Unscripted SEO Podcast transcripts looking for a narrow thing: guests who could attach a figure, a mechanism, or a shipped test to their claims about getting cited by large language models. Not a prediction about where this is all going — evidence about what already happened on a site they run.
Six could. They do not agree with each other, and the disagreement is the most useful thing here. One has real LLM traffic and is nervous about it. One has a technical lever that moved fast. One has been quietly compounding the right asset for a decade. One says you cannot reliably measure any of it. And one thinks the whole conversation is mostly noise.
If you want the step-by-step version, that lives in the SOP for getting cited by ChatGPT. This piece is the other half: what practitioners actually saw. It is also the sequel to 7 Quick Wins Hiding in Your Last SEO Crawl, because one of the levers below turns out to be crawl-export work wearing a new hat.
The only real traffic number anyone gave me
Paul Baterina works on SEO at Revolve, which is a large enough business that “LLM traffic is growing” is a measurable claim rather than a vibe. I asked him whether AI visibility was in his remit or somebody else’s, and he answered with the thing nobody else would put on the record:
Just for some context right now, we’re around 20,000 visitors per month with all of LLMs, whether it’s ChatGPT, Perplexity — of course ChatGPT probably is taking 90% of the traffic. And gross purchase revenues at roughly $58,000 a month, which is very low compared to the entire business. But is there opportunity? I think so.
Paul Baterina, Revolve
Sit with both halves of that. Twenty thousand visitors a month and fifty-eight thousand dollars a month in gross purchase revenue is not nothing. It is also, in his own words, very low compared to the entire business. That is the honest shape of AI search traffic for a brand at scale: real, attributable, and small.
What is more interesting is why he has not gone all in. His hesitation is not scepticism about the channel. It is scepticism about the tactics:
Maybe on LinkedIn or maybe on Slack or wherever it might be, we might see this new strategy that, hey, getting an increase of more comments or working on your brand reputation on Reddit, more citations, you’re going to show up. And then three weeks later, the number of citations have dropped with Reddit. And then later on, someone’s going say, do some FAQs on all of your product pages or your PLPs. And then soon enough, there’s a new algorithm that’s going to take place. And that’s not going to show up in these LLMs no more… we just want to be very careful if we want to go all in on something. Is it going to last in the long term?
Paul Baterina, Revolve
That is a durability test, and it is a better filter than most of the frameworks currently being sold. Before you adopt an AI-visibility tactic, ask whether it would still be worth doing if the citation behaviour it targets changed next quarter. Most of them fail that question. Two of the things below pass it.
The lever that moved fastest: schema
Brittany Trafis of Soarion Digital runs an AI-native agency, and her opening position is a useful corrective to a lot of the current positioning: “The number one piece I say is AI search is not SEO 2.0. And so SEO as a foundation is still important… Your backlinks, your keywords, that strategy still exists. And a lot of that is a foundation for AI search.”
What changed, in her account, is not the foundation but which parts of the technical stack now carry weight — the elements that mattered enormously twenty years ago, went quiet, and have come back:
Things that 20 years ago were so important from an SEO perspective, words like schema and meta tags and descriptions. Like this all was talked about all the time and then started to get a little bit more quieter. And now adjusting your schema can really improve your citation rates very quickly within AI search.
Brittany Trafis, Soarion Digital
Note the two words doing the work: very quickly. That is what makes schema the first thing to touch rather than the fifth. Almost every other AI-visibility play is a compounding, months-long investment in reputation. Schema is a deploy. It is also, conveniently, something your last crawl already told you the state of.
Her analogy for why people get this wrong is the sharpest framing I have heard for it: treating AI search like SEO would “almost be like when Facebook ads came out, if you were to say, I’m just going to run it like Google ads. Well, that wouldn’t work.”
The asset that compounded: mentions, not links
Jason Barnard of Kalicube has been arguing about entities since before it was fashionable, and he arrived with the receipt. In January he did not appear on any list of answer-engine-optimization experts. Then the back catalogue caught up with him — a fifteen-part SEMrush series, a TrustPilot webinar, a Search Engine Watch article from 2018. Lots of mentions. Not many links.
Number one is whoever 10 years ago was saying, don’t care about mentions, should be kicking themselves today.
Jason Barnard, Kalicube
He calls the system claim, frame, prove. You claim a thing, you frame it, and then you have to be able to prove it — and mentions across time are what constitutes proof to a machine. “I started this in 2018 and I can prove it. And here’s the proof.”
Sara Nay of Duct Tape Marketing lands in the same place from the demand side. Asked whether to invest in your own content or in getting mentioned elsewhere, her answer is that the question is malformed:
I think you need to be thinking about a combination of both, to be honest. I think if you’ve been producing content on your own site in a very helpful, authoritative way over time, then you’re set up for more success right off the bat with LLMs. But LLMs are pulling from websites, but they’re pulling from Reddit, and they’re pulling from all different sources.
Sara Nay, Duct Tape Marketing
Her read on what that does to link-building economics is worth stealing: “Maybe the link is a little bit less important these days, but I’m still getting exposure to other audiences by doing these things.” The placement stopped being about the link and went back to being about the audience, which is where it started.
The most portable idea in this whole cluster came from Barnard — the mapping between two systems most people treat as unrelated:
You look at an entity in the knowledge graph. It’s a thing. And little by little, you can reinforce the presence of that entity in the knowledge graph… But if you think of an LLM, the equivalent is a parameter. In an LLM, I’m a parameter. And if I reinforce that parameter, it’s the same effect as reinforcing an entity in the knowledge graph.
Jason Barnard, Kalicube
If that mapping holds, entity work you did for Google was never wasted and does not need redoing for LLMs. It also means the Wikipedia panic is misplaced. Barnard’s numbers: Wikipedia has around six million articles, Google’s Knowledge Graph has fifty-four billion entities. “It’s a seed source. And if you get added to Wikipedia, it definitely helps you. But LLMs are way beyond that as well.”
The measurement problem nobody has solved
Benas Leonavicius supplies the caveat that should be stapled to every AI-visibility report you receive. It is not that the tools are bad. It is that the thing being measured does not hold still:
Tracking is one of the major sort of problems to actually monitoring what is happening. Yes, there are tools out there, but I mean, probably half of them are not really providing you with any like real tangible results based on the way AI works.
Benas Leonavicius
The failure mode is specific, and if you have ever screenshotted a favourable ChatGPT answer for a client deck, it applies to you:
Even if you type in something and you appear, doesn’t mean that the next person in the next location with a next history is going to type in and you’re going to appear. But they just, they say, I don’t care. I still want to be there. Even if it’s like 25% of the time, I still want to be there 25% of the time.
Benas Leonavicius
That last sentence is the commercially honest position. Visibility in AI search is probabilistic, and the correct unit of measurement is share of answers across many runs, not a screenshot. Clients will accept that framing if you give it to them up front. They will not accept it after you have shown them one good screenshot.
The dissent: none of this buys you anything
Jeremy Moser of uSERP is the reason this is not a consensus piece. He is not arguing that AI search does not exist. He is arguing that the specific prize everyone is chasing — being mentioned in an AI overview — is worth close to nothing right now.
It’s like you’re not getting any brand awareness from being mentioned in an AI overview at this point. I mean you’re seeing maybe a few brands show up in there from an informational standpoint but it’s so little and so few and far between that it’s just not worth it from a content perspective in our opinion right now.
Jeremy Moser, uSERP
What makes his position hard to dismiss is the data underneath it. His clients’ top-of-funnel traffic is being eaten, and it turns out not to matter:
People will see our traffic and SEO has declined 30% over the last year but our metrics are the same in terms of leads that are coming through, in terms of revenue driven from organic. How can that be possible? And it’s really a lot to do with that. They were just ranking for content that wasn’t actually driving any results for them long term.
Jeremy Moser, uSERP
He is blunter still about the vendor layer that has grown up around all this: “I think there’s just a lot of kind of BS out there in the AI search space and that’s kind of the long and short of my rant.”
Where the six of them actually land
Read together, they are not contradicting each other so much as describing the same channel from six positions on it.
- The traffic is real and small. Baterina has the number. Treat AI visibility as a channel worth instrumenting, not one worth restructuring around.
- Do the cheap technical thing first. Trafis’s schema point is the only lever here that pays off in days rather than quarters, and it costs you a deploy.
- Then do the slow thing that was never wasted. Barnard and Nay converge on mentions and third-party presence. It survives the tactic churn Baterina is worried about, which is exactly why it passes his durability test.
- Report it probabilistically or not at all. Benas’s point kills the screenshot as a deliverable. Share of answers across repeated runs, or nothing.
- Do not confuse a citation with a customer. Moser’s traffic-down-thirty, revenue-flat pattern is the sanity check. If the mention does not show up downstream, it is a vanity placement.
The practical sequence that falls out: fix your schema this week, keep earning mentions the way you always should have, instrument LLM referral traffic so you have your own version of Baterina’s number, measure presence as a percentage across many runs rather than as a screenshot, and refuse to spend a content budget on the AI-overview slot until your own data says it converts.
The how-to for the first two of those is in the get-cited-by-ChatGPT SOP. Everything above is why it is worth running.
Every quote in this article is verbatim from Unscripted SEO Podcast transcripts and attributed to the guest who said it. Light cleanup has been applied to filler words only.