Hub
AI search optimization
The hub: how to optimize a whole site for AI search.
AI search
LLM optimization: how to make your site easy for AI to find, read and cite
LLM optimization (LLMO) is the work of making your website and your brand easy for large language models such as ChatGPT, Claude, Gemini and Perplexity to find, understand and mention. Nobody can promise you a place in an AI answer. You can control whether these systems are allowed in, what they read, and how clearly your pages answer the question.
What is LLM optimization?
LLM optimization (LLMO) is the practice of making a website and a brand easy for large language models to find, read, understand and mention when they answer someone's question. It covers who may crawl your pages, what those pages contain without JavaScript, how clearly they answer questions, the structured data on them, and how your brand is described elsewhere.
The term has two meanings, and the search results mix them. Machine learning engineers use "LLM optimization" for making a model faster, cheaper or more accurate: quantization, batching, caching, fine-tuning and prompt design. Marketers and site owners use it for being visible in AI answers. This guide is about the second meaning. If you run a model, the first meaning is an engineering topic with its own literature.
For a site owner, the useful way to think about LLMO is as a chain. A model can only mention you if something it relies on contains you: its training data, a page it fetches while answering, or a search index it queries. Each link in that chain has parts you control and parts you do not.
How do LLMs find and choose content?
Large language models get information about your site in three main ways: from training data collected by crawlers before the model was built, from live searches and page fetches made while answering a question, and from what other sites say about you. You can influence the first two directly through crawler access and page content, and the third only slowly.
| Route | What happens | What you control | What you do not |
|---|---|---|---|
| Training data | Crawlers such as GPTBot, ClaudeBot, Google-Extended and CCBot collect pages that may be used to train future models | Whether your robots.txt allows those crawlers, and what your pages say | Whether, when and how your pages are used in training |
| Live retrieval | While answering, an assistant searches or fetches pages, for example with OAI-SearchBot, ChatGPT-User, Claude-SearchBot or PerplexityBot | Crawler access, page speed, content in the HTML, clear answers, structured data | Which pages the assistant picks and how it words the answer |
| Third-party mentions | Reviews, directories, forums, press and comparison pages describe your brand | Your profiles, your facts being consistent, earning genuine coverage | What other people write |
The crawler names above are the ones each company documents for its own bots. Check each operator's current documentation before you write rules, because names and purposes change. Some operators separate a training crawler from a search crawler, so you can decide on each one rather than allowing or blocking AI as a whole.
LLM optimization vs SEO and GEO: what is the difference?
SEO aims to rank a page in a list of search results. LLM optimization aims to have a page read, used and mentioned inside a generated answer. GEO, generative engine optimization, and AEO, answer engine optimization, describe the same goal from different angles. The groundwork is shared, so a site with sound SEO is most of the way there.
The differences are in emphasis:
- The unit is the passage, not the page. An answer may use two sentences from the middle of your article. Each section needs to make sense on its own.
- The crawlers are different. AI operators run their own bots, and a robots.txt written years ago may block them without anyone deciding to.
- JavaScript matters more. Some AI fetchers read the HTML your server sends and do not run scripts, so content built in the browser can be missing for them.
- Reputation is read, not counted. What forums, reviews and articles say about you can shape how an assistant describes you, beyond any link count.
- There is no rank to check. Answers change with the wording, the user and the day. You measure samples and trends, not a position.
What parts of LLM optimization can a site owner control?
A site owner controls five things that affect LLM visibility: which AI crawlers robots.txt allows, whether the main content is in the HTML before JavaScript runs, how directly each section answers a question, whether structured data states the key facts accurately, and whether the brand is described the same way on its own site and its profiles elsewhere.
| Lever | Why it matters | How to check |
|---|---|---|
| Crawler access | A blocked crawler cannot read the page at all | Read your robots.txt rules for each AI user agent, or build new rules with the robots.txt generator |
| Content in the HTML | Fetchers that do not run JavaScript only see what the server sends | View the page source, or run the crawlability test, and confirm prices, answers and key copy are there |
| Clear answers | A model lifts passages that answer the question in the first sentences | Read each heading as a question and check the next two sentences answer it alone |
| Structured data | States who you are, what you sell and what a page is, without guessing | Paste your JSON-LD into the schema markup validator |
| Consistent brand facts | Conflicting names, prices or descriptions make a summary less reliable | Compare your about page, your profiles and your directory listings line by line |
How to do LLM optimization in 7 steps
To do LLM optimization, check which AI crawlers can reach your site, make sure your main content is in the HTML, list the questions your buyers ask, answer each one clearly in its own section, add accurate structured data, make your brand facts consistent everywhere, and then measure crawler visits, AI referrals and sampled answers over time.
-
Decide which AI crawlers to allow
Open your robots.txt and look for rules that name AI user agents, and for broad rules that block everything except a few search engines. Decide crawler by crawler. Many site owners allow search and retrieval crawlers, so their pages can be fetched for answers, and make a separate choice about training crawlers.
-
Put the content in the HTML
Load your key pages with JavaScript switched off, or read the page source. If the product description, price, answer or pricing table is missing, an AI fetcher that does not run scripts will miss it too. Render that content on the server or in the static HTML.
-
List the questions your buyers actually ask
Collect them from sales calls, support tickets, site search, Search Console queries and People Also Ask boxes. Write them the way a person would type or say them. These questions become your headings, and they are also the prompts you will later test.
-
Answer each question in its own section
Use the question, or close to it, as a heading. Answer it in the first two or three sentences, name the subject instead of writing "it", then add detail, steps or a table. A passage that stands alone is easier for a model to quote correctly.
-
Add structured data that matches the page
Mark up the organization, products, articles and FAQs with JSON-LD, using only facts visible on the page. Validate it. Markup that contradicts the page, or claims things the page does not show, does more harm than none.
-
Make your brand facts consistent
Use the same name, description, prices and claims on your site, your social profiles, directories and marketplace listings. Earn genuine mentions on sites your buyers read: reviews, comparisons and articles. Do not plant fake reviews or forum posts; they can be traced back to you and damage the reputation you are building.
-
Measure, then repeat
Check your server logs for AI crawler visits, set up a channel for AI assistant referrals in analytics, and ask the main assistants your buyers' questions every month. Record what they say and which sources they cite. Then fix the next gap you find.
An optional extra is an llms.txt file, a proposed plain-text index of your most useful pages. It is quick to write and harmless, but do it after the seven steps, not instead of them.
How to write content that LLMs can use
Content that language models can use answers one question per section, puts the answer first, names things explicitly, and backs claims with specifics a reader can check. It avoids vague marketing language, because a model summarizing "the leading solution for modern teams" has nothing concrete to repeat.
- Lead with the answer. The first sentence under a heading should work as a reply on its own.
- Be specific. Say what the product does, for whom, at what price, and with what limits. Specific facts are easier to repeat accurately.
- Use real lists and tables. Steps as numbered lists, comparisons as tables. Both survive being lifted out of the page.
- Define your terms. If you use a name for a feature or a method, say what it means the first time.
- Show who wrote it. A named author and an about page give the content a source that can be checked.
- Keep it current. Update facts such as prices and policies where they change, and show a real updated date only when the content really changed.
How do you measure LLM optimization?
You measure LLM optimization with three kinds of evidence: AI crawler requests in your server logs, sessions referred by AI assistants in your analytics, and repeated manual checks of how assistants answer your buyers' questions. None of these is a ranking. Look for trends over months rather than reacting to a single answer.
| Signal | Where to find it | What it tells you | Limit |
|---|---|---|---|
| AI crawler visits | Server or CDN logs, filtered by user agent | Whether AI operators fetch your pages, and which ones | A fetch is not a mention; user agents can be faked |
| AI referral traffic | Analytics, with a channel for assistant hostnames (see AI traffic tracking) | Visits and conversions from people who clicked a link in an answer | Many answers send no click, and some referrers are stripped |
| Sampled answers | Asking assistants your buyers' questions on a schedule | How your brand is described and which sources are cited | Answers vary by wording, user, location and day |
| Search Console | Google Search Console performance report | Queries and pages that earn impressions in Google Search, which includes AI features | Does not separate AI features from ordinary results |
Be careful with tools that report a single "AI visibility score" from prompts they run. They can be useful for spotting a trend, but the number depends on which prompts they chose. Ask how the prompts were picked before you treat the score as a measure of your market.
Common LLM optimization mistakes
The most common LLM optimization mistakes are blocking AI crawlers by accident, hiding key content behind JavaScript, writing vague copy with nothing specific to repeat, adding structured data that does not match the page, chasing tricks instead of fixing basics, and buying services that promise guaranteed placement in AI answers.
- Blocking by accident. A blanket robots.txt rule, a firewall bot setting or a CDN challenge can stop AI crawlers without anyone choosing to.
- Content only in scripts. Tabs, accordions and prices injected after load may not exist for a fetcher that reads the HTML.
- Hidden text for bots. Content shown to crawlers but not to people is a risk in search and adds nothing a visible answer would not.
- Markup that lies. Ratings, prices or FAQs in JSON-LD that do not appear on the page.
- Prompt stuffing. Instructions aimed at AI systems hidden in a page are a manipulation attempt, and AI operators work to ignore them.
- Trusting guarantees. No one can promise a mention in ChatGPT or any other assistant.
- Forgetting the visit. A visitor who arrives from an AI answer still lands on a page that has to convert.
Audit and rebuilds
Where Convertica fits
The free AI visibility checker runs the same app as the free CRO audit. It reads your page and your site files: the robots.txt rules for AI crawlers, llms.txt, structured data, and what the HTML says without JavaScript. It does not ask ChatGPT or any other assistant anything, and it does not track mentions or citations. It shows which of the levers on this page are in place.
The audit scores eight checks for people and AI agents and gives you every finding and why it matters, for free. If you want the fixes built for you, the one page rebuild has Convertica's team build every fix the audit found on that page, and the full site rebuild covers your whole site, whatever platform your site runs on.
In this series
More on AI search
AI search
ChatGPT SEO
How ChatGPT finds pages, and what you can change on yours.
Free tool
Schema markup validator
Check your JSON-LD against Google's requirements, in your browser.
Short answer
LLM optimization (LLMO) means making your website easy for large language models like ChatGPT, Claude, Gemini and Perplexity to reach, read and mention. You cannot control their answers, but you can allow their crawlers, put your content in plain HTML, answer questions clearly, add accurate structured data and describe your brand consistently wherever it appears.
- LLMO is mostly good SEO groundwork plus attention to AI crawlers, passage-level answers and your brand's footprint on other sites.
- Start with access: a robots.txt rule that blocks an AI crawler, or content that only appears after JavaScript runs, can keep your pages out of reach.
- Write sections that answer one question each, in the first sentences, so a model can lift the passage without the rest of the page.
- Nobody can guarantee a mention in an AI answer. Measure crawler visits, AI referral traffic and sampled answers instead of chasing a rank.
- Visitors who arrive from an AI answer still need a page that converts, so LLMO and CRO belong together.
Questions about LLM optimization
What is LLM optimization?
LLM optimization (LLMO) is the practice of making a website and a brand easy for large language models to find, read, understand and mention in their answers. It covers crawler access, content that is readable without JavaScript, clear and specific answers, structured data, and a consistent description of the brand across the web.
Is LLM optimization the same as SEO?
No, but it overlaps heavily. SEO aims at a ranked list of links on a results page. LLMO aims at being read and mentioned inside a generated answer. Most of the groundwork is shared: a crawlable site, fast pages, clear content and a trusted reputation. LLMO adds attention to AI crawlers, passage-level answers and how your brand is described elsewhere.
What is the difference between LLMO, GEO and AEO?
They are three names for closely related work. LLMO names the technology (large language models). GEO, generative engine optimization, names the AI search products built on them. AEO, answer engine optimization, names the goal of being the answer. In practice the same changes to a site serve all three.
Can you guarantee my brand appears in ChatGPT answers?
No. Nobody outside the companies that run these models controls what they say, and answers change from one question, user and day to the next. Be wary of anyone who promises placement. What you can control is whether AI crawlers can reach your pages and whether those pages answer questions clearly.
Do I need an llms.txt file for LLM optimization?
It is optional. llms.txt is a proposed plain-text file that points language models to your most useful pages. It is cheap to add and does no harm, but no major AI search product has said it relies on it. Crawler access in robots.txt and content in your page HTML matter more.
Does structured data help LLM optimization?
Structured data such as Organization, Product and FAQPage markup states facts about a page in a format machines read without guessing. It does not force an AI answer to use your page, but it removes ambiguity about who you are, what you sell and what a page is about. Validate it before relying on it.
How do I measure LLM optimization?
Measure what you can see: whether AI crawlers fetch your pages in your server logs, visits from AI assistants in your analytics, and manual checks of how assistants describe your brand for the questions your buyers ask. Treat any single answer as a sample, not a ranking, because answers vary.
What does LLM optimization mean for machine learning engineers?
In machine learning, LLM optimization means making the model itself faster, cheaper or more accurate: quantization, batching, caching, prompt design, retrieval and fine-tuning. That is a different job from this guide, which covers how a website gets found and mentioned by language models.
