James Dooley Podcast

James Dooley: How to increase AI search visibility. Today I'm joined with Charles Floate.

Charles Floate: Thanks for having me.

James Dooley: How are we doing? For anyone who wants to get higher rankings and more citations in places like Claude, Perplexity, ChatGPT and Gemini, how do you increase visibility within AI?

Charles Floate: The first thing is you need to have a defined approach around where your audience is.

For most niches, the audience is probably going to be on ChatGPT. But if you're in B2B or SaaS, it is more likely to be Claude right now.

First, you need to figure out where your audience is. Then you need to figure out how to approach rankings for that specific niche.

Right now, listicles and content on your own site that can get picked up initially in grounded search queries are the most important.

Those queries change quite a lot depending on the model you are using.

OpenAI and ChatGPT tend to use very long, formal queries that are extracted from their training data.

For example, if you type “best CRM tools for accountants” into ChatGPT, it may modify the original query to something like “best CRM for small business accountants, HubSpot versus Trello versus Atlassian”.

You could end up with a 16 to 18-word query.

If you look at Claude, it tries to keep the query between one and six words, but then refines the query afterwards.

Depending on what information comes back from the first search, it will then do more searches based on that initial query context.

So first, you need to see where your audience is. Then you need to create content targeted around how the AI finds information to give to the user.

In general, there are some rules around lists. You cannot be too self-promotional. You need to include some drawbacks.

You cannot be overly positive and say these are the best in the world and nobody else can come close.

You need some level of honesty and facts as well. Bear in mind that Claude does additional search queries to validate some of those facts and findings.

James Dooley: With regards to listicles, then, to increase search visibility within AI, you're saying the sentiment value cannot just be positive, positive, positive. You need to include pros and cons and things like that?

Charles Floate: Exactly.

James Dooley: You mentioned grounded data. What do you mean by that for anyone watching this?

Charles Floate: Grounding is when the AI goes and searches for additional information through a grounded search.

It will go to Google if it is Gemini. It will go to Bing if it is ChatGPT. Claude will use a mixture of its own index and third-party sources like Google.

Grounding makes sure the AI is not relying on outdated training data that may have been trained from last November or earlier.

It makes sure the AI has the freshest and most relevant data.

Most AI model companies, especially for fresh queries or current events, are now forcing the AI to do grounded search first so it has information that is as up to date and factual as possible.

James Dooley: So let's break it down with regards to AI visibility. Obviously, you have the different LLMs.

With Claude versus ChatGPT versus Gemini, should you use different strategies for different LLMs?

Charles Floate: If you want to be really granular and only target one AI model or one AI company because that is where your user base is, then yes.

But if your user base is all over the place, which it can be, someone could be using Gemini on their phone, someone could be using ChatGPT on desktop, and someone else could be using Claude at work.

In that case, you want a broader approach that takes in the general overlapping factors and signals that AI models have.

Some models have specific grounding differences, but there are general practices that they all use.

They all look for official domains. They all look for documentation. They all look for PDFs.

There are other things they look at too, but almost all of them take some level of training data into account when coming up with the query in the first place.

They are not just basing the query on your input. They are basing the query on your input plus what they already know about that niche.

If you can understand what the AI model already knows about your niche, you can reverse engineer more effectively and place content on the sources it already trusts and continually recommends.

James Dooley: If you're looking to increase AI visibility, you're talking about live search.

Gemini is using Google. You are saying ChatGPT is using Bing.

But you have mentioned training data a couple of times. How can someone get into that training data? How long does it take? If someone has a brand and wants to form part of the training data, how can they do that?

Charles Floate: The first question is how long does it take? It depends on the company.

Some companies release a model every two to three months. Some only release once a year.

How long it takes depends on that training data ingestion cycle.

Also, bear in mind that if the last training date says November 7th, that does not mean the specific source you are on was last crawled on November 7th.

It may have been crawled months before that, so the actual window is normally further back than the training cycle date.

Most companies are still using Common Crawl. They are still using Internet Archive. They are using a range of PDF websites and downloading from sites like Scribd.

If you can get your content, PDFs and brand context across these overlapping file sources where AI companies are generating and downloading information from, then you will usually be able to get into the next training set.

They are still downloading from those same areas, and they will continually do that across cycles.

If you place your content into those datasets, then the next training iteration that uses that downloaded dataset should be able to include it.

How much information do you need in the training dataset for it to actually build into the AI model?

If you have one PDF file with one sentence about your brand, that probably is not enough information to be ingested into the training data and absorbed as an element.

If you have 100 pages, 10 PDFs, an entire website, guest posts, podcasts and other content that has all been ingested, there is a much larger chance of it being absorbed.

So it is more of a volume game, but there is definitely weighting when it comes to the sources themselves.

James Dooley: When you talk about a volume game, is that where people talk about consensus?

Charles Floate: Yes.

To get picked up by training data, it is a volume game in the sense that the more sources you are on, the more likely you are to be put into the training data.

You are also more likely to be downloaded by one of the AI company engineers, or even one of the AI agents doing the downloading on behalf of the company.

However, in terms of consensus, you want that data to say the same thing.

It needs to reinforce the same sentiment, the same people, opinions and locations related to that entity.

If all the different pages are saying completely different things, it will very quickly confuse the training data and it will not be able to trust it.

You are then more likely to be ingested in the wrong way than the right way.

That is important because the more you get ingested incorrectly at the beginning, the harder it becomes later to change the AI's opinion about your business or entity.

James Dooley: A lot of people talk about query fan-outs.

Should a brand try to cover all the different query fan-outs? Under the reputation dimension, for example, you might have awards, case studies, testimonials and reviews.

Should brands proactively try to get those on third-party sources, not just their own website, to build that consensus?

Also, what about video? Should brands be doing video? Can video be pulled into LLMs to improve AI visibility?

Charles Floate: Specifically for Gemini, we found that even if a video does not have public transcripts, Gemini still has access to the transcripts in the back end.

Gemini can pull from some private YouTube API or something similar to get transcripts from videos that do not have public transcripts available.

For Gemini specifically, 100%.

For other models, it is very dependent on whether that model is prioritising YouTube or not.

ChatGPT and Claude tend to prioritise video quite differently, and it also depends on whether the query itself prioritises video.

But I would always say that most people call them LLMs, when actually they are not really just LLMs anymore because they are multimodal.

An LLM is literally inputting language and text. But if it can understand input images, video and audio, then it is much bigger than just an LLM.

Most AI is moving towards that.

You definitely want to position your content not just through text, but also through audio, video and imagery.

You should build your entity, knowledge understanding and consensus across all of those different verticals.

James Dooley: Anyone watching this, I hope you liked the video and podcast about how to increase AI visibility.

Today I was joined by Charles Floate, and it has been an absolute pleasure. Make sure you check out the links in the description.

There are lots of other videos where we talk about how to rank in Bing search, which increases exposure in ChatGPT, parasite SEO and what is working in today's algorithms.

Charles, it has been an absolute pleasure.