Charles Floate: How to Rank #1 in Claude

So today we have Charles on the podcast talking about Claude AI SEO. Welcome Charles. Thank you for having me Julian. We can't move too much in this because the autofocus on the camera is totally messed up, but we'll try and go through this.

So you've been working on basically reverse engineering the system prompts for Claude to understand how to rank inside Claude, right? Exactly, yeah. So last year I figured out because I'm in a lot of the red teaming and like prompt injection kind of community discords, that a lot of them were talking about Anthropic. It's not just Anthropic, but Google Gemini and Tract2BT, their system prompts and how actually the system prompts can massively affect how your pages are retrieved, how they synthesize and how you actually get output.

And Claude's system prompts specifically, because they had a public version of that system prompt, a lot of people were just going by that. But actually this public version that Anthropic released wasn't the full actual system prompt. Like it didn't talk about retrieving anything. It didn't talk about like tool usage, like web runs, all of this kind of stuff that you'd expect inside of the system prompt, how it actually interacts with the web and how it interacts with data and how it interacts with the tools that are automatically available to all the default Claude models.

Inside of that system prompt, it actually started talking about targeting SEOs and how large amounts of SEO content can often be distrusted because they can essentially be forcibly optimized by SEOs and that to trust things like changelogs or Wikipedia or dev sites or the actual core brand itself, rather than large amounts of SEO content or pages. And even in that same line, if you went to the same paragraph, the next line in that paragraph was talking about conspiracy theories. So I thought it was really crazy that SEO had been sort of mangled in with essentially conspiracy theories. That's how much Claude was telling its model to distrust or Anthropix was telling Claude to distrust SEO information.

And that doesn't just mean literal like how to rank in Google. It means that its SEOs have created information for a certain niche. That's crazy. So literally Claude's algorithm is trained to not trust SEOs because SEOs ruin everything.

Exactly. Unfortunately. And so like, is it better to, for example, target social media sites where content is more likely to be more natural? So yes, it's, I would say it's more better to trust this or it's better to target the sites that Claude seems to overlappingly trust over and over again.

And it's not even necessarily the sites themselves. It's the way the information is formatted and the way the information is structured and presented as well. So one of the techniques that we started doing was creating these little dot wiki sites. And this was not my invention at all, by the way, this is another very clever SEOs invention that I reverse engineered.

But if you search up myself into Claude, it's going to come back with charlesfolk.wiki. And that site, that domain name is only a few months old. So the fact that it's pulling that over several other of my own sources that are maybe years or even a decade plus old is a very telling sign that it prefers that format and the structure of the Wikipedia style data than it would over maybe a blog post. So does that mean like, if you're trying to rank inside Claude, do you need like a totally separate strategy to rank inside Claude versus say ChatGPT and Google and everything else?

It is getting to that point where you, because essentially you have two factors, where ChatGPT, Gemini, Anthropic, Claude, all these different models tend to have preferential sources anyway. So there's going to be types and specific domains and things that they're going to cite higher than others will anyway. So that kind of targeting means you either have to be kind of omniscient, but you're everywhere anyway, or you have to be hyper-focused down the specific kind of sites and things that you can get. And bear in mind, some of those sites, it might be thousands of dollars per post or something that you actually have to go on to, to be able to get onto those sites in the first place.

Yeah, that's it. I can imagine like some of those sites just charging crazy prices. Yeah, exactly. Yeah.

What about like, what are the differences would you say in ranking in ChatGPT versus Claude? ChatGPT, A, it tends to pull a lot more sources. So I think from our tracking, and this tracking is a couple months old, so it might be different as of right now. We were looking in last time, Claude is pulling between five and six sources as a step one.

ChatGPT is pulling between 18 and 30. So just from the get-go, if you have a much, much larger amount of data to kind of sift through and actually extract from, then it's less likely to maybe pull your entity from the one out of 30 pages than it is from the one out of five or six, right? So you just have like a higher probability of being cited if you're on a smaller retrieval source. And then secondarily, it's the format and the structure of the content.

Like I was saying, Claude really, really prefers content that isn't just claims, but it can back up those claims. It can actually have factual bits of information to reinforce those. And to be fair, ChatGPT has been moving more towards that, but for the longest time you could basically make up anything and ChatGPT and Google AOVs and Google Gemini would spit it out as fact. I feel like it's almost still at that point where it's so easy to manipulate Google, so easy to manipulate ChatGPT.

But from what I've heard, I was speaking to Jabez Rubin about this recently. He was saying that actually if you cite yourself consistently as the number one on all your listicle posts, all your blog posts ranking the top 10 products or services, typically Claude will mark you down or ignore you for that reason. Yeah, so I think that's moving into this over-optimization world that we used to have in SEO anyway. You could easily over-optimize your links, your assets profile, your blog posts, your sites, all that kind of stuff.

And it's the exact same thing for ASEO. Essentially, you're over-optimizing the reality of the situation. If you go and look up the best fast food chains in America, there's going to be different restaurants at number one for whoever it is. And bear in mind, it might be different depending on what state you're from even, because you might have access to certain restaurants in certain states that aren't in other states.

But there is no necessarily tier one, the ultimate kind of fast food chain in America. And because those lists are all different, the same applies to every other niche, right? Unless there is one monopoly that actually just does completely dominate that niche. And again, that's not necessarily allowed in most business environments and things, right?

There isn't actually a number one. So you want to be differentiating yourself. And maybe you want to be making sure that you're associating yourself with other companies and other entities as well. So you shouldn't necessarily put the exact same companies in every single list.

You should be diversifying yourself, connecting yourself, and building relationships between your competitors. This is another thing that AI will start understanding over time as well, especially from its training data, which companies have relationships. And that doesn't necessarily mean literal relationships like maybe Julian and I have, but it's a metaphysical relationship, which is the relationship that we are in the same industry as each other, basically. And what do you use for rank tracking and tracking your rankings yourself?

We use Amplitude. And then we also have our own tracking one. And then we also use AHFs and SEMrush data anyway. But AHF, SEMrush data tends to be lagging behind the reality of user searches.

And obviously, it's also not properly tuned towards personalization. And so in terms of our own tracking, when we're using an anti-theft browser with profiles, a dedicated proxy and TypeScript to kind of automate it all, it ends up, you can actually create persona profiles. So what we'll do is we'll pretend and engage with Chattripty over the course of several months and say, hey, look, we're a middle-aged man in New York state looking for X, Y, and Z, really interested in cars and really interested in whiskey and things like that, right? And we'll build the persona of our exact ideal customer profile.

And then we'll reverse engineer what Chattripty is pulling back for that profile versus this profile that might be the complete opposite, right? It might be a lady in Thailand or something that doesn't speak any English at all or something, right? But they could potentially have overlapping markets on some kind of niches. Nice.

Nice. Awesome. Any other tips you'd give before going? I would go back to where your audience is, 100%.

So we use SparkToro. SparkToro by Ran Fishkin, and I think he has another business partner as well, but I'm not quite sure her name. It's a fantastic tool for identifying where your customers are. So what you do is you create that IDP or the ICP, the initial or the customer profile, ideal customer profile, and it will tell you exactly where your customers hang out.

So they'll say 92% of them are in Chattripty, 5% of them are in Claude, 3% of them on this. And it will say the same for social media. It will say 80% in TikTok, 20% in Reddit, things like that. So you can then immediately know, okay, if 80% of my clients are going to be on Reddit, then I should be running Reddit ads, right?

Rather than Facebook ads, because maybe only 18% of them have Facebook. Yeah, it makes total sense for sure. I feel like it's changing so much as well right now. You go back five or 10 years ago, people middle-aged weren't using TikTok at all.

No. And I think they found it kind of cringy as well to some extent, right? A lot of the internet for the longest time was very cringy, and now my grandma's friends are doing TikTok dances and things like... It's crazy.

Yeah, it is pretty crazy. Even my dad watches it. Exactly, yeah. I would never expect him to.

No, not at all. All right, well, thanks so much for being on, Charles. Fantastic. What's the best way to get in touch?

We'll leave a link in the description. Twitter, x, at Charles underscore SEO. Awesome. Cheers.

Thanks for having me.