Imagine it's the year 2000, and you get the advice to invest heavily in link building to get higher search engine rankings.

If you took that advice, the odds are pretty good that your business would succeed for the next two and a half decades.

As of today, with this research, you're getting the same level of advice, except with AI and LLMs.

Previous Research into AI/LLM Ranking Factors

Many of you are familiar with the two AI Search Visibility reports I released in May1 and July2 of this year.

In them, I analyzed more than 3 petabytes of data, including 10 billion web pages, all of Reddit's more than 25 billion submissions and comments, the entirety of Wikipedia, and over a quarter trillion links, across 100 industries and over 1000 buyer segments.

And do you know what I was able to unequivocally conclude from that?

Nothing.

Sure, there were some interesting and useful correlations, but nothing was conclusive.

Today's research changes that.

Across 100 industries, 16 LLMs from OpenAI and Google, ChatGPT's web interface, and Google's AI Overviews, I found the three ranking factors that correlate every...single...time...with not only being ranked, but achieving higher rankings in the LLMs.

In this test, that's over 20,000 times the pattern was true, and not once where it wasn't.

How LLMs Work

To understand why this combination of factors works, you need to understand how LLMs are made. Because, at a very high level, this experiment is similar to reverse-engineering how LLMs work.

This is slightly technical, but I'm going to make it as simple as I can.

I guarantee you that as you understand this better and implement this advice, you and your marketing team can position yourself better in LLMs.

For those of you who are more technical, yes, lots of details are left out and a lot of things are over-simplified. If you want more technical details of all of this, I'm releasing an hour-long video about all of this next week on YouTube. Follow me now to make sure you see that.

To understand LLMs and this research, there are three simple parts.

Part 1: Text Embedding Models

First, when you "train" an LLM, you're basically just giving it a lot of stuff to read.

It takes that web page, or book chapter, or video transcript, or whatever it is, and it then converts it into a whole lot of numbers.

Those numbers are called an Embedding.

To go from the text to those numbers, you need a thing called a Text Embedding Model.

There's an interesting feature of embedding models: They all very roughly, "learn" similar things.

What I mean by that is that they all come to understand that dogs and cats are similar pets. Basketball and football are sports. Hotels and airplanes are related to travel. etc.

In short, through this black box called a text embedding model, you create a generic understanding of the world via words and text.

Part 2: How an LLM Uses Embeddings

The second piece to understand is that an LLM uses embeddings like these, inside the model itself, both to learn as it's trained and to create its response to you.

I'm not going into the details of that in this video.

But, what you need to know is that a text embedding - this generic understanding of language converted into numbers - is used to both train the LLM and also generate its response to you.

First, it converts text from your prompt into lots of numbers, the embedding.

Then it converts those numbers back into the text response you see.

Because the nature of embeddings leads them to all come to very similar understandings of language, you can find similar patterns in ChatGPT or Gemini, without needing ChatGPT's or Google's own embeddings.

Part 3: Understanding Vector Similarity

Here's the third and final thing to understand about this research:

It's possible to take two different web pages, turn them into two different embeddings, then see how similar they are, based on how close the numbers are to each other.

You may have heard of "cosine similarity". That is one possible calculation of similarity between embeddings.

If you take three pages and get their embeddings, now you can see which two are most similar.

If you take 100 of them, you might be able to imagine some of those being in clusters of similar pages, just like you might stack documents into groups on your desk, or categorize files with labels or tags.

When you get thousands, or even millions of them, things start to get interesting...

If you are able to figure out "this embedding here is about hotels. this one is about airlines. This one is about life insurance." and so on, then you can start doing the real magic...

You begin noticing how a new web page's embedding is closer to the hotels embedding, then you look at the content, and see that really is about hotels...

Similarity is the key concept in all of this.

So...

And what if you had embeddings that represented thousands of business categories?

Could you then categorize web pages with embeddings?

But if all it took was to create really optimized text on their site, everyone would do it...so then the LLM couldn't choose one to rank, right?

So then what?

Well...we're going to need a bigger dataset.

The Research

For this research, I stuck with the 100 industries I used in the previous AI Search Visibility report from July.

I then decided to ask 13 OpenAI LLMs, 3 Gemini LLMs, ChatGPT through the web interface, and Google AI Overviews for the top business recommendations for each of those 100 industries.

To get some variation in there, I asked for the top 25 recommendations, 25 times, in 25 slightly different ways.

eg:

• "What are the 25 best hotels and resorts?"
• "List the top 25 hotel and resort brands."
• "Which 25 hotels and resorts do people trust the most?"

I then took those results and combined the scores based on rank position and rank frequency. This gets me a final, overall score across an industry, for that LLM.

Choosing Taxonomies

Next, we need to categorize content using embeddings.

Thankfully, there are a number of these things called Taxonomies, where people have spent a long time deciding what a good set of business and content categories would be.

This research uses four different taxonomies, from three different organizations, to make sure that the research isn't skewed toward one taxonomy.

Let's use the Google Ads Topics taxonomy. It's 2,558 categories like Hotels, Air Travel, Life Insurance, Golf Equipment, etc.

With those categories, and some OppAlerts magic, we can then generate an embedding from a new web page, and see how similar that web page is to each of the taxonomy categories in the Google Ads Taxonomy.

And that's great, but all it does is show you that a web page is more or less similar to over 2,500 categories.

We're trying to figure out why the sites that rank, rank...

But, Anyone Can Create Relevant Content...

So, I could have downloaded the home pages of each site, turned it into a vector, and seen how close it was to which categories.

But that wouldn't be conclusive and we all understand that. Anyone can create relevant content. So it can't just be that.

But what if we look at the embeddings of a site's backlinks? How would we get those?

The OppAlerts Indexes

Well...

OppAlerts has two search engines it maintains.

One for news, including over 100k news sites, 20 million RSS feeds, and podcasts. This gathers over half a million new pages each day, converts it into a vector, does a bunch of other analysis, and puts it into a search engine for its users to monitor brands, competitors, and anything else they find relevant.

OppAlerts also has a search engine for link and marketing opportunities like guest posts, product reviews, events, sponsorships, etc. It also gets more than half a million new pages added to it each day, also with an embedding of the main content.

These embeddings allow users to monitor for opportunities and mentions that may not have specific keywords like you expect from traditional brand monitoring tools (though OppAlerts can also monitor specific terms and phrases), but it also lets you monitor for semantically-similar content.

This helps a company like the "Four Seasons" find content similar to hotels and travel, while filtering out content that's related to Vivaldi's Four Seasons, or sports content talking about the "last four seasons" for a team.

It's over 60 million documents in there, with embeddings, could we use that?

It turns out, we can.

Backlinks, Mentions, and Pure Mentions

I took the more than 100,000 sites that were recommended by the 50,000 prompts searched those OppAlerts indexes.

I searched them and got each result's embedding in three ways:

1. All pages with backlinks to the sites.
2. All pages with mentions of the site's brand or domain.
3. All pages with mentions of site's brand or domain, where it doesn't also have a link. ie: pure mentions.

Now, the real test...do the LLM's recommendations have backlink profiles that match their targeted industry?

Well, let's look.

This is the hotel industry in GPT-6 Astra. Here's the top-recommended site, Marriott. Then the second, the third, the fourth, all the way down to number 100. The height of each bar is the average similarity score of all the backlinks to that site, for one taxonomy category. On the left is the category most similar to the industry, Hotels. As the categories become less and less similar, based on the vector distance, you can see the chart goes down.

Across every industry, every LLM, and even when you swap out the taxonomy used, the top-suggested sites in the LLMs have backlinks that are, on balance, more relevant to the industry, and less relevant to less-related industries.

And the correlation is so strong, I believe we can call it causality.

But...what about mentions, and mentions without links?

Yes.

Yes, they do.

It's the same pattern. In fact, when it comes to pure mentions without backlinks, the correlation often appears even stronger.

So, when we include the analysis of mentions with backlinks, and non-backlink mentions, that gives well over 20,000 instances where, every single time, the correlation is true.

Context of your backlinks matter. We might call it Zero-Click Marketing. But Mike King at iPullRank was right: what we're doing is relevance engineering3.

These ranking factors might not be unexpected for some people.

But what really matters is that we as marketers finally have something concrete to point to.

With SEO, we never had anything concrete. With LLMs, now we do.

We no longer have to act based on our intuition. We now have facts.

And those facts are exactly what you need to get the resources and budget you need from your company or clients.

Optimizing for AI

So how does this change how we optimize our marketing for AI, right now?

First, you need to get all of your marketing team on the same page, whether that's one person in your business, an entire team, or the agencies you work with.

From SEO, to reputation monitoring, to content marketing, to competitive analysis, to link building, to PR, you need to get your messages all both aligned and targeted to your market.

Next, you need to bring semantic similarity to your content marketing, brand monitoring, and link building.

After looking across the industry, I can say that OppAlerts is the only company doing this.

Go sign up for an OppAlerts account. You get daily monitoring and daily alerts for brand and competitor mentions, link opportunities, the marketing opportunities you're looking for, and more, based on semantic similarity.

You also get access to the content optimization tool that shows you which taxonomy categories your content is most relevant to.

With a paid account, you can search the indexes, get results a thousand at a time, and sort by PageRank or by how established a site is. And even with the LinkGap tools, you can constantly monitor semantically relevant pages that link to two or more of your competitors.

And, whether you're getting a free account or not, you're going to have MCP and API access for ChatGPT and Claude Code.

So go sign up at oppalerts.com.

More Technical Details

For more technical details and images, read the full research post on the blog post coming out tomorrow.

Finally, I'm going to record a much longer version of this, to help you understand exactly how LLMs work. It's the video I wish I'd had 25 years ago, when I really wanted to understand how search engines work.

So follow me, and follow the OppAlerts YouTube channel. I'll let everyone here know when the hour-long video is released next week.