SEO & AI Search
LLM SEO: How to Get Your Site Read by AI Models
LLM SEO is the work of making sure large language models can fetch your pages, parse them, and reuse what they find. Here is the part that surprises people: most sites that fail at it do not fail on content quality. They fail because the model literally could not read the page, either because the content only appears after JavaScript runs, or because the crawler was blocked at the door.
That makes LLM SEO less glamorous than it sounds. It is mostly plumbing, and plumbing is fixable in days, not quarters. Once the page is readable, a second layer decides whether you actually get quoted: clear structure and evident credibility. This guide covers both, in that order, because there is no point tuning your writing for a model that cannot open the door.
What LLM SEO actually means
An AI model can meet your content in three ways. It can be trained on it, which happens once and cannot be undone or repeated on demand. It can retrieve it live when a user asks a question, which is how ChatGPT search, Perplexity, and Google’s AI Mode work. Or a browsing agent can fetch your page on a user’s behalf.
Only the second and third are things you can influence week to week, and both depend on the same chain: a crawler reaches your URL, receives readable HTML, extracts a passage, and decides it is worth using. Break any link in that chain and nothing downstream matters. Most of the fixes below are about not breaking it.
Most AI crawlers cannot run JavaScript
This is the single biggest technical gap between traditional SEO and LLM SEO. Google renders JavaScript. It crawls your raw HTML, queues the page for a headless browser, runs your scripts, and indexes the result. Most AI crawlers skip that step entirely.
A Vercel and MERJ analysis of crawler traffic found that while OpenAI’s and Anthropic’s crawlers do download JavaScript files, they do not execute them. GPTBot fetched JavaScript on about 11.5% of requests and Anthropic’s crawler on about 23.8%, running none of it. Across more than 500 million GPTBot requests in a single month, the study found zero evidence of JavaScript execution. Their conclusion was blunt: client-rendered content is effectively invisible to roughly 70% of AI crawlers.
The practical consequence catches teams by surprise. A React or Vue page can rank at position one on Google and be a blank shell to ChatGPT at the same time. If your service descriptions, pricing, or FAQ content load client-side, they do not exist as far as most AI models are concerned.
To test it, disable JavaScript in your browser and load your most important page. Or run curl against the URL and read what comes back. Whatever is missing from that raw HTML is missing from the model’s view. If the important content disappears, server-side rendering is your highest-leverage fix, and it will help traditional SEO too.
Check which crawlers you are blocking
The second common failure is a robots.txt written years ago, or a security rule added by IT after a traffic spike. Many sites block AI crawlers without anyone having decided to.
| Crawler | Run by | Purpose | Renders JavaScript |
|---|---|---|---|
| GPTBot | OpenAI | Training data collection | No |
| OAI-SearchBot | OpenAI | ChatGPT search results | No |
| ClaudeBot | Anthropic | Training data collection | No |
| PerplexityBot | Perplexity | Search index | No |
| Google-Extended | Controls Gemini’s use of Googlebot-crawled content | Yes, via Googlebot | |
| Applebot | Apple | Apple search and AI features | Yes |
The rendering behavior above reflects the Vercel and MERJ measurements. Treat it as a working assumption rather than a permanent guarantee, and check your own server logs for which of these user agents actually reach you.
Blocking is a legitimate business decision, and some publishers make it deliberately. Just make it deliberate. Blocking GPTBot keeps you out of training data but does not remove you from ChatGPT search, which uses a separate agent. Those are two different choices, and most robots.txt files treat them as one.
Structure pages so a model can lift a passage
Once a crawler can read the page, the question becomes whether anything on it is usable. Models retrieve chunks, not whole documents, so each passage has to make sense on its own.
- Put a direct, complete answer within the first hundred words, with no setup paragraph in front of it.
- Write H2s and H3s as the questions people actually ask, not as clever labels.
- Keep paragraphs to three or four sentences, since long blocks get truncated mid-idea.
- Make definitions self-contained, so a sentence lifted alone still says something true.
- Use tables and short lists for comparisons, which models extract more reliably than prose.
- Add Article, FAQPage, and Organization schema markup so the machine does not have to infer what the page is or who wrote it.
None of this requires writing differently for robots. It is the same clarity a rushed human reader wants.
Then earn the model’s trust
Readable and well-structured gets you into the running. Whether you are the source a model actually names comes down to credibility, and here the signals are less technical.
Models lean heavily on consistency. If ten credible sources describe your company the same way, the model treats that description as settled and repeats it with confidence. If your homepage, your LinkedIn, and your directory listings each say something different, you have taught it nothing it can safely quote. Aligning how the web describes you is unglamorous work that most businesses skip, and it moves the needle more than any single page edit.
Named authors with real bios help, because a byline attached to an accountable person signals that experience produced the page. So do earned mentions on reputable third-party sites, which build both link authority and the description consistency above. This is the slow part, and it is the part competitors cannot shortcut.
What about llms.txt?
Skip it for now. The file was proposed as a markdown map for AI systems, and the industry adopted it faster than the AI companies did. Ahrefs studied 137,000 sites and found that among domains with a valid llms.txt file, 97% received no requests for it at all in the month measured. Google has said its search systems do not use it.
It costs little to publish, so it is not harmful. It is just not the reason you are invisible. If you have time for one file, spend it on your robots.txt instead.
What is genuinely new, and what is just SEO
Be skeptical of anyone selling LLM SEO as a wholly separate discipline with its own retainer. Server-side rendering, clean crawl access, clear structure, real depth, and earned authority are simply good SEO, and they serve a human searcher and a machine reader at once. Do that work once and it counts twice.
The genuinely new part is a short list: making sure your important content survives without JavaScript, deciding deliberately which AI crawlers you allow, tightening the answer passages at the top of your pages, and auditing whether models name you and how they describe you. Treat it as an extension of your SEO program, not a replacement for it.
Find out what the crawlers actually see
You do not need a tool for the first check. Pick your five highest-value pages, load each with JavaScript disabled, and note what disappears. That list is your LLM SEO backlog, in priority order. Then, for the strategy layer above the plumbing, get a citation baseline: run ten questions a real buyer would type into ChatGPT, Gemini, and Perplexity, and write down whether you appear and how you are described. Our guide to finding AI traffic in GA4 shows how to separate AI referral traffic you are probably not tracking yet, and GEO vs SEO covers how the economics of visibility have shifted.
If you would rather someone else run it, OgreLogic will do a crawler-eye render test on five of your pages and send back exactly what GPTBot and ClaudeBot receive, alongside what a human sees. Email us with the five URLs, or call +1 (512) 861-8471. If everything comes back clean, we will tell you that and you can stop worrying about it. If it does not, our AI Search Optimization work is where the fixes live. No obligation either way.
FAQs
Is LLM SEO different from regular SEO?
It overlaps heavily but diverges on rendering. Traditional SEO can rely on Google executing your JavaScript, so client-side content still gets indexed. Most AI crawlers do not render JavaScript, so that same content is invisible to them. Server-side rendering matters far more for LLM SEO than it does for Google, and it is the most common reason a well-ranked page gets no AI citations.
Will blocking AI crawlers protect my content?
Partly, and at a cost. Blocking training crawlers such as GPTBot and ClaudeBot keeps your content out of future model training. It does not remove you from live AI search, which uses separate agents, and blocking those agents removes you from answers your buyers are reading right now. Decide each one separately rather than blocking on reflex.
How long does LLM SEO take to show results?
Faster than traditional SEO and less predictable. Retrieval-based systems can pick up a republished page within days once it is readable, because they query a live index rather than waiting on a ranking update. What they will not do is hold that position steadily, so expect visibility to move around week to week.
How do I know if AI models are citing my site?
Test it directly and check your analytics. Build a fixed list of 20 to 30 questions your buyers would ask, run them monthly across ChatGPT, Gemini, and Perplexity, and log whether you appear and how you are described. Pair that with referral traffic from AI domains in GA4 and a watch on branded search volume, which often rises before the traffic does.
