Schema Is Useless for LLMs: The SEO Lie That’s Finally Exposed

Category: N/A

Transcript

00:00

People hated me for the hard stand I took on this. I said despite what lots of other SEO influencers were saying, schema was not used by LLMs.

And these SEO influencers, especially the ones with agencies who went and said these

00:16

things, who said that LLMs need schema, you need to have schema to get covered more in LLMs. I said that these people were huers and scammers and people hated me for saying that.

It was a very hard

00:32

stand. But now we have evidence that chatpt is not using schema.

And this is going viral in the SEO community. And I want to say before I get into this, the thing about schema is it takes a lot of work to implement, especially if you're implementing granular schema, not

00:49

standard schema, which a lot of page builders and site builders, they do that automatically. And that's important.

But for the more niche schema takes a lot of work to implement, to maintain and it can be quite expensive. And something that I say all the time on this show is it is not worth it.

And especially for

01:05

LLMs, it is not necessary. It is a lie perpetuated by people who want to charge a lot of money to implement it.

So this is what is going viral. Now I'm going to share the intro from search engine roundtable.

Structured data does not help with visibility in AI search. And

01:20

then I'm going to share the actual study. And after that, I'm going to share things that you can do right now to get covered more in AI search.

This article says, "There are a lot of folks in the community saying that implementing structured data/s schema on your pages will help you with AI search

01:36

visibility, but few have really tested it until now. And those few tests show that adding structured data/ schema does not help with your visibility in AI search, at least not yet." The first to test this was Mark Williams Cook who

01:51

posted on LinkedIn an experiment he conducted where he posted a visual explanation of why your favorite LLM does not use schema in their core training data. And this is what he said.

LLMs work by tokenizing content. That means taking common sequences of

02:08

characters found in text and minting a unique token for that set. The LLM then takes billions of sample windows of sets of these tokens to build a prediction on what comes next.

And he shares an image. He says, "The image below is some

02:23

example schema that has a color change applied which represents that set of characters is a unique token as made by the GPT40 model. What you will notice is that the schema gets destroyed.

For instance, the schema at type colon

02:39

organization gets broken down. So there are separate tokens for type and organization which means that in terms of tokenization the regular words type and organization are not distinguishable from schema.

If schema was included in this training data all it would do in

02:55

reality is say there is a slightly likely insignificant probability of tokens such as at appearing before the word content. Schema is useful because it is explicit.

This is lost during tokenization. The entire point of the LLM approach is to build some kind of

03:12

knowledge from free form text without the need for structured data. Of course, LLMs could be using schema in some kind of RI process, but this would likely be very similar to what we are seeing with search engines.

It's just it's interesting because this image that Mark

03:27

shares, it really does show that the schema is getting destroyed, not being used. Now, if that's not good for you, somebody named Julio C ran a test and he starts the test.

This is on LinkedIn. He starts the test by saying, "I have seen this advice going on since the beginning

03:43

of the geo hype, the generative engine optimization hype, and it is just factually wrong." Structured data does not directly help you get mentioned in AI channels. It does not help LLM understand your pages better.

This is by design. When ingesting your page, LLM

04:00

tokenize the text they find. They chunk it up into little pieces roughly the size of a word.

This also applies to your information put within the structured data. So, it gets chunked up and is no longer structured nor readable.

What is worse is that LLMs usually use only text information

04:16

available to answer textbased questions. Since most structured data is in JSON, LLM will rarely register this as part of a pertinent answer.

To them, it looks like code. How do I know?

Here we go. We tested it.

We set up two product pages of the same madeup product that both

04:32

Gemini and ChachiPT had never seen before. One page had all content visible in the HTML as text plus structured data.

The other page had only structured data and nothing else visible as text visually empty. We tried different

04:48

extraction prompts hundreds of times to see if the LLM could give back information like price, colors, skew numbers. Surprise, surprise.

This only worked on the page with information visible as text. If your agency is telling you that structured data is the

05:04

way to be mentioned in AI channels, they are either uninformed or giving charlatan advice. Somebody else says it.

I'm so happy. And before you get into comments citing correlation studies, structured data correlates with citations because it correlates with normal SEO rankings since it helps their

05:20

systems. Correlation does not equal causation.

All right, so here's some things to get mentioned more by AI. I had Steve Toth on the show a few episodes, episode 798, AI SEO versus normal SEO, winning LLMs

05:36

with brand first strategy. And Steve Toth talks about how lots of people are doing due diligence with AI with Chacht, with AI overviews, with AI mode, with Perplexity, with Claude.

they're thinking about going with a company or

05:51

thinking about getting a a product or a service and they're using AI to make research on this easier. And what you want is you want to make it really really easy for the AI to pull up the best things about your company.

Not only do you want to shorten sales cycles by

06:07

having AI know all the stuff that is going to help you close a sale, but you also want to make AI recommend your product over other products. So, this is what you do.

This is just a very easy thing that you can do right now. Do this either on your homepage or on your about

06:23

page or on both. Write a ton about why your company is great.

Lots of people have thin homepages. Thin about pages or the about page is irrelevant or there's there's not enough relevance on any of these pages.

Talk about why you're great. Give as many positive

06:38

testimonials and reviews as you can. Share contact information.

Give prices. have an FAQ that is already dropped down a frequently asked questions that is already dropped down with a text visible.

So this actually a test that

06:53

was run. It was the hidden text experiment.

Search pilot looked to test the performance of text within tabs versus text that is fully visible upon page load. They split tested two versions across many product pages of their clients website and measured the

07:09

change in organic traffic over time. The version with the text fully visible got 12% more organic sessions than the text hidden within tabs.

So have an FAQ that's fully visible. Even comparisons, you can make comparison pages or mention

07:24

this on an about page or even on a homepage. Though the homepage isn't the best, but talk about comparisons on why you are better than competitors.

Think about ChachiPT. ChachiPT is always doing comparisons between products in a niche.

So you control that narrative. And I

07:41

talk about this in compact keywords, my SEO course about making alternatives pages, targeting alternatives keywords. People are searching for an alternative to one of your competitors.

You want them to find you and you want Chachi PT or Perplexity or these LLMs to recommend you. And then put out a press release

07:58

saying all of this stuff using different language. You don't want it to be verbatim language.

Using different language. And I talked about that on episode 768 of this show.

$6 press releases are hacking chachi PT perplexity in Google. Press releases influence AI like crazy and they're not

08:15

very expensive. On that show, I talk about how you can do it for six bucks.

Anyway, schema, especially niche schema, takes a lot of work to implement and there are way better things for getting results. And if somebody tells you to use schema to get better results with

08:30

LLMs, cut that person out of your life. That's it.

All right, I'm done. If you want to learn what actually what really gets sales with search engine optimization, my 13 and 1/2 hour SEO course, compact keywords, is about this exactly.

I'm getting just the craziest

08:46

testimonials. I'm constantly adding new testimonials, new video testimonials to the landing page.

If you want to see what everybody's talking about, that's at compactkeywords.com. And this is episode 86 of my daily digital marketing show.

806 days in a

09:02

row without missing a day doing this podcast. Thank you so much for watching.

Thank you so much for listening. I will talk to you again tomorrow.

Bye now.