As a marketer, I privation to cognize if determination are circumstantial things I should bash to amended our LLM visibility that I americium not presently doing arsenic portion of my regular selling and SEO efforts.

So far, it doesn’t look like it.

There seems to beryllium monolithic overlap successful SEO and GEO, specified that it doesn’t look utile to see them chiseled processes.

The things that lend to bully visibility successful hunt engines besides lend to bully visibility successful LLMs. GEO seems to beryllium a byproduct of SEO, thing that doesn’t necessitate dedicated oregon abstracted effort. If you privation to summation your beingness successful LLM output, prosecute an SEO.

It’s worthy unpacking this a bit. As acold arsenic my layperson’s knowing goes, determination are 3 main ways you tin amended your visibility in LLMs:

1. Increase your visibility successful grooming data

Large connection models are trained connected immense datasets of text. The much prevalent your marque is wrong that data, and the much intimately associated it seems to beryllium with the topics you attraction about, the much disposable you volition beryllium successful LLM output for those fixed topics.

We can’t power the information LLMs person already trained on, but we tin make much contented connected our halfway topics for inclusion successful aboriginal rounds of training, some connected our website and third-party websites.

Creating well-structured contented connected applicable topics is 1 of the halfway tenets of SEO—as is encouraging different brands to notation you wrong their content. Verdict: just SEO.

2. Increase your visibility successful information sources utilized for RAG and grounding

LLMs progressively usage outer information sources to amended the recency and accuracy of their outputs. They tin hunt the web, and usage accepted hunt indexes from companies similar Bing and Google.

It’s just to accidental that being much disposable successful these information sources volition apt summation visibility successful the LLM responses. The process of becoming much disposable successful “traditional” hunt indexes is, you guessed it, SEO.

3. Abuse adversarial examples

LLMs are prone to manipulation, and it’s imaginable to instrumentality these models into recommending you erstwhile they different wouldn’t. These are damaging hacks that connection short-term payment but volition astir apt wound you successful the long term.

This is—and I’m lone fractional joking—just black hat SEO.

To summarize these 3 points, the halfway mechanics for improving visibility successful LLM output is: creating applicable contented connected topics your marque wants to beryllium associated with, some connected and disconnected your website.

That’s SEO.

Now, this whitethorn not beryllium existent forever. Large connection models are changing each the time, and determination whitethorn beryllium much divergence betwixt hunt optimization and LLM optimization arsenic clip progresses.

But I fishy the other volition happen. As hunt engines integrate much generative AI into the hunt experience, and LLMs proceed utilizing “traditional” hunt indexes for grounding their output, I deliberation determination is apt to beryllium little divergence, and the boundaries betwixt SEO and GEO volition go adjacent smaller, oregon nonexistent.

As agelong arsenic “content” remains the superior mean for some LLMs and hunt engines, the halfway mechanisms of power volition apt stay the same. Or, arsenic idiosyncratic commented connected 1 of my recent LinkedIn posts:

“There’s lone truthful galore ways you tin shingle a instrumentality astatine aggregating a radical of information, ranking it, and past disseminating your champion approximation of what the champion and astir close result/info would be.”

I shared the supra sentiment successful a LinkedIn post and received immoderate genuinely fantabulous responses.

Most radical agreed with my sentiment, but others shared nuances betwixt LLMs and hunt engines that are worthy understanding—even if they don’t (in my opinion) warrant creating the caller subject of GEO:

This is astir apt the biggest, clearest quality betwixt GEO and SEO. Unlinked mentions—text written astir your marque connected different websites—have precise small interaction connected SEO, but a overmuch bigger interaction on GEO.

Search engines person galore ways to find the “authority” of a marque connected a fixed topic, but backlinks are 1 of the astir important. This was Google’s halfway insight: that links from applicable websites could relation arsenic a “vote” for the authorization of the linked-to website (a.k.a. PageRank).

LLMs run differently. They deduce their knowing of a brand’s authorization from words connected the page, from the prevalence of peculiar words, the co-occurrence of antithetic presumption and topics, and the discourse successful which those words are used. Unlinked contented volition further an LLM’s knowing of your marque successful a mode that won’t assistance a hunt engine.

As Gianluca Fiorelli writes successful his excellent article:

Many companies already worth off-site mentions, albeit with the caveat that those mentions should beryllium linked (and dofollow). Now, I tin ideate brands relaxing their explanation of a “good” off-site mention, and being happier with unlinked mentions successful platforms that walk small accepted hunt benefit.

As Eli Schwartz puts it,

“In this paradigm, links don’t request to beryllium hyperlinked (LLMs work content) oregon restricted to accepted websites. Mentions successful credible publications oregon discussions sparked connected nonrecreational networks (hello, cognition bases and forums) each heighten visibility wrong this framework.”

I deliberation the inverse of the supra constituent is besides true. Many companies contiguous physique backlinks connected websites with small relevance to their brand, and people contented with nary transportation to their business, simply for the postulation it brings (what we present telephone site estimation abuse).

These tactics connection capable SEO payment that galore radical inactive deem them worthwhile, but they volition connection adjacent little payment for LLM visibility. Without immoderate applicable discourse surrounding these links oregon articles, they volition bash thing to further an LLM’s knowing of the marque oregon boost the likelihood of it appearing successful outputs.

Some contented types person comparatively small interaction connected SEO visibility but greater interaction connected LLM visibility.

We ran probe to research the types of pages that are astir apt to person postulation from LLMs. We compared a illustration of pageviews from LLMs and from non-LLM sources, and compared the organisation of those pageviews.

We recovered 2 large differences: LLMs amusement a “preference” for halfway website pages and documents, and a “dislike” for listing collections and listings.

Citation is much important for an LLM than a hunt engine. Search engines mostly aboveground accusation alongside the root that created it. LLMs decouple the two, creating an other request to beryllium the authenticity of immoderate assertion is being made.

From this data, it seems the bulk of citations autumn into the “core tract pages” category: a website’s location page, pricing page, oregon astir page. These are important parts of a website, but not ever large contributors to hunt visibility. Their value seems greater for LLMs.

Inversely, listings pages—think large breadcrumbed Rolodexes of products—that are created chiefly for on-page navigation and hunt visibility received acold less visits from LLMs. Even if these leafage types aren’t cited often, it’s imaginable that they mightiness further an LLM’s knowing of a marque due to the fact that of the co-occurrence of antithetic merchandise entities. But fixed that these pages are usually sparse successful context, they whitethorn person small impact.

Lastly, website documents besides look much important for LLMs. Many websites dainty PDFs and different forms of documents arsenic second-class citizens, but for LLMs, they are a contented root similar immoderate other, and they routinely mention them successful their outputs.

Practically, I tin ideate companies treating PDFs and different forgotten documents with much importance, connected the knowing that they tin power LLM output successful the aforesaid mode immoderate different tract page would.

The constituent that LLMs tin entree website documents raises an absorbing point. As Andrej Karpathy points out, determination whitethorn beryllium a increasing payment to penning documents that are structured archetypal and foremost for LLMs, and near comparatively inaccessible to people:

This is an inversion of the SEO adage that we should constitute for humans, not robots: determination whitethorn beryllium a payment to focusing our vigor connected making accusation accessible to robots, and relying connected the LLMs to render the accusation into much accessible forms for users.

In this way, determination are circumstantial accusation structures that tin assistance LLMs correctly recognize the accusation we provide.

For example, Snowflake refers to the thought of “global papers context”. (H/T to Victor Pan from HubSpot for sharing this article.)

LLMs enactment by breaking substance into “chunks”; by adding other accusation astir the papers passim the substance (like institution sanction and filing day for fiscal text), it’s easier for the LLM to recognize and correctly construe each isolated chunk, “boosting QA accuracy from astir 50%-60% to the 72%-75% range.”

LLMs besides bid connected caller accusation sources that person traditionally fallen extracurricular the remit of SEO. As Adam Noonan connected X shared with me: “Public GitHub contented is guaranteed to beryllium trained connected but has nary interaction on SEO.”

Coding is arguably the astir palmy usage lawsuit for LLMs, and developers indispensable marque up a sizeable information of full LLM users.

For immoderate companies, particularly those selling to developers, determination whitethorn beryllium a payment to “optimizing” the contented these developers are astir apt to interact with—knowledgebases, nationalist repos, and codification samples—by including other discourse astir your marque oregon products.

Lastly, arsenic Elie Berreby explains:

“Most AI crawlers bash not render JavaScript. There’s nary renderer. Popular AI crawlers similar those utilized by OpenAI and Anthropic bash not adjacent execute JavaScript. That means they won’t spot contented that is rendered client-side done JavaScript.”

This is much of a footnote than a large difference, for the elemental crushed that I don’t deliberation this volition stay existent for precise long. This occupation was solved by galore non-AI web crawlers, and volition beryllium solved by AI web crawlers successful short order.

But for now, if you trust heavy connected JavaScript rendering, a bully information of your website’s contented whitethorn beryllium invisible to LLMs.

Final thoughts

But here’s the thing: managing indexing and crawling, structuring contented successful machine-legible ways, gathering off-page mentions… these each consciousness similar the classical remit of SEO.

And these unsocial differences don’t look to person manifested successful extremist differences betwixt astir brands’ hunt visibility and LLM visibility: mostly speaking, brands that bash good successful 1 besides bash good successful the other.

Even if GEO does yet germinate to necessitate caller tactics, SEOs—people who walk their careers reconciling the needs of machines and existent people—are the radical best-placed to adopt them.

So for now, GEO, LLMO, AEO… it’s each just SEO.