What is LLM optimization (LLMO)?

LLM optimization (LLMO), usually spelled LLM optimisation in Australia, is the work of making your business easy for large language models, and the AI assistants built on them, to find, understand, trust and mention, so you are named, linked or recommended inside AI-written answers. Large language models are the technology underneath assistants such as ChatGPT, Gemini, Claude, Copilot and Perplexity, and the AI Overviews and AI Mode features in Google Search. When someone asks one of those tools who to hire or what to buy, the answer names a handful of sources, and LLM optimisation is how you give your business a fair chance of being one of them.

This guide is for business owners and marketers who want a plain explanation before spending time or money. Its claims come from what Google, OpenAI, Anthropic and Perplexity publish about their own systems, the robots.txt standard, two research papers, a university survey and our own reporting.

What is LLM optimisation?

Start with what the model does. Google's machine learning crash course describes a language model as a system that works out how likely a token, or a run of tokens, is to appear within a longer passage of text, where a token is a word, part of a word or a single character.

Google's own example: given "When I hear rain on my roof, I _______ in my kitchen," a model might rate "cook soup" at 9.4 percent and "warm up a kettle" at 5.2 percent, then pick one of the likeliest. Scale that up and the same pattern-matching lets a model generate, summarise and translate text.

When an assistant names a business, it is either repeating a pattern it absorbed in training or lifting the name from a page it has just read. LLM optimisation works on both routes and aims at three outcomes:

• A citation. Your page is linked as a source inside the answer.
• A mention. Your brand is named in the answer, with or without a link.
• A recommendation. You are suggested when someone asks who to hire, what to buy or which options to compare.

The term overlaps with generative engine optimisation (GEO), coined in a research paper first posted in November 2023 and accepted to the KDD 2024 conference, which described generative engines as systems that gather information from several sources and summarise it with large language models. LLMO is the plainer, broader label. It covers those live search assistants and also what a model says from memory with no search at all.

How an AI assistant decides what to say

An assistant has two sources of knowledge: what the model learned in training, which researchers call parametric memory because it lives in the model's parameters, and what it retrieves at the moment you ask.

The 2020 paper that introduced retrieval-augmented generation (RAG) explains why the second source exists. Pre-trained models store factual knowledge in their parameters, but the paper notes three limits: that knowledge is hard to access precisely, hard to attribute to a source and hard to update. RAG pairs the model's memory with documents fetched from an outside index, and the authors found it produced more specific, diverse and factual language than memory alone. An answer from memory cannot show a link or be updated quickly; an answer from retrieval can do both.

ChatGPT search was introduced by OpenAI on 31 October 2024 to give "fast, timely answers with links to relevant web sources." ChatGPT decides whether to search based on the question, or you can trigger it, and a Sources button opens a sidebar of linked references. OpenAI says any website or publisher can choose to appear, and the feature has been open to everyone, with no sign-up, in regions where ChatGPT is available since 5 February 2025.

Google AI Overviews and AI Mode may use what Google calls query fan-out, issuing multiple related searches across subtopics and data sources to build a response while its models identify supporting pages to show as links. AI Overviews only appear when Google's systems judge them to add something beyond classic Search, so they often do not trigger. Perplexity and Claude work on the same pattern, each with a crawler for search and a separate fetcher for user requests; the names and controls are in step one below.

So for live answers your page has to be reachable by that engine's crawler and clear enough to lift; for answers from memory your brand has to be described consistently across the web the model learned from. Nothing the vendors publish explains how a brand ends up in a model's memory, so treat that part as reasoning rather than a rule.

LLMO, SEO, GEO and AEO: how they relate

These are overlapping labels for one goal: being used in AI-generated answers. GEO is the research term from the 2023 paper. AEO, answer engine optimisation, looks at the same goal from the answer box rather than the model, and LLMO names the technology. Our guide to generative engine optimisation and our guide to answer engine optimisation cover each in full. The beginner's guide to AI SEO ties them together.

Underneath all of them sits ordinary SEO. Google says the same foundational SEO best practices apply to AI Overviews and AI Mode as to Google Search overall: to be eligible as a supporting link, a page must be indexed and eligible to appear in Search with a snippet, and there are no additional technical requirements.

The emphasis shifts once a page is indexed. The GEO researchers found that keyword stuffing offered little to no improvement in generative engine responses, and in their Perplexity test it performed about 10 percent worse than the unchanged page. Adding citations to credible sources, relevant quotations and statistics lifted source visibility by more than 40 percent across queries, and by up to 37 percent on Perplexity. These are lab results on rewritten content, and the authors said the effect varied by domain, so read them as a direction rather than a forecast.

The other difference is scope. ChatGPT, Claude, Perplexity and Gemini each sit behind their own crawlers and controls, so a page can be visible in one and absent from another for reasons unrelated to ranking, which is why AI SEO services in Australia bundle the same jobs whatever the label: crawl access, answer-first content, structured data, brand mentions and measurement.

Why LLM optimisation matters for your business

The habit behind all of this is simple. A buyer who once scanned ten blue links now asks an assistant for a shortlist. If the answer to "who does commercial fit-outs in Brisbane" names three firms, a fourth is invisible to that buyer however well it ranks.

Google frames this as a feature. It reports that clicks from results pages with AI Overviews are higher quality, meaning visitors are more likely to spend longer on the site, though that is a vendor describing its own product.

There is a particular reason for smaller businesses to pay attention. In the GEO study, lower-ranked pages gained more from the content changes than top-ranked ones: adding citations lifted visibility by 115.1 percent for sites ranked fifth in the underlying results, while the top-ranked site lost 30.3 percent on average once every source was optimised. That is a benchmark rather than a guarantee, but an encouraging one.

Adoption is already broad. The Australian National University's AI Adoption in Australia report (July 2026, a survey of more than 3,500 adults) found that 48.6 percent of Australian adults have used generative AI; it is a university survey rather than an official statistic. The same logic applies across the Tasman, which is why AI SEO for New Zealand businesses is built on the same three outcomes: citations, mentions and recommendations.

Step one: let the right crawlers in

If a crawler cannot fetch your page, nothing else in this guide applies, so start here. Crawling is governed by robots.txt, standardised by the IETF as RFC 9309 in September 2022. The standard calls its rules requests that crawlers are asked to honour and says plainly that they are not a form of access authorisation.

OpenAI, Anthropic and Perplexity separate search from training, and Google separates Search from Gemini training, so you can usually stay in search answers while opting out of training. Here is how each describes its crawlers at the time of writing.

• OpenAI. OAI-SearchBot surfaces websites in ChatGPT search; sites that opt out are not shown in ChatGPT search answers, though they can still appear as navigational links. GPTBot crawls content that may be used to train OpenAI's models. The two settings are independent. OpenAI recommends allowing OAI-SearchBot and its published IP ranges, and says its search systems take about 24 hours to reflect a robots.txt change.
• Anthropic. ClaudeBot collects web content that could contribute to model training. Claude-User fetches pages when people ask Claude questions, and Claude-SearchBot analyses content to improve search results. Anthropic says disabling either of the last two may reduce your visibility in Claude's search, and warns that blocking its IP addresses instead of using robots.txt may not reliably opt you out, because it stops the bots reading the file.
• Perplexity. PerplexityBot surfaces and links websites in Perplexity's results and is not used to collect content for training. Perplexity-User fetches a page when a user asks a question and generally ignores robots.txt because a person requested it. Sites behind a web application firewall such as Cloudflare or AWS WAF may need a rule that explicitly allows both bots.
• Google. Googlebot rules control crawling for Google Search and all its features, including AI Overviews and AI Mode. Google-Extended is a separate token for Gemini training and grounding; Google says it does not affect inclusion in Search and is not a ranking signal.

A site that wants to appear in answers but stay out of training would allow OAI-SearchBot, Claude-SearchBot, Claude-User, PerplexityBot and Googlebot, and add rules such as:

User-agent: GPTBot
Disallow: /

User-agent: ClaudeBot
Disallow: /

User-agent: Google-Extended
Disallow: /

Blocking Google-Extended does not affect Google Search, AI Overviews or AI Mode, but it also opts your site out of grounding in Gemini Apps and Vertex AI, so a site that wants to be used in Gemini app answers should weigh that before blocking it. A common failure is a firewall or CDN rule written for one purpose that quietly does the other, so check your server logs after any change rather than assuming the file did what you meant.

Step two: write pages an AI can quote

Once crawlers can get in, the question is whether your pages give them anything worth lifting. An assistant assembling an answer wants a short, self-contained passage that states a fact plainly. Google's list of practices that remain worthwhile for its AI features includes keeping important content in text form rather than only in images or video.

• Put the answer first. Open each page and section with the direct answer in one or two sentences, then the evidence. Define your terms, since an assistant quoting a definition needs it on the page, and use numbered steps and short FAQs for the questions buyers ask.
• Give it evidence it can reuse. The GEO study's strongest results came from adding citations, relevant quotations and statistics. A figure with a source attached is easier to quote than an adjective.
• Say something only you can say. Google's people-first content questions ask whether a page provides original information, reporting, research or analysis, and whether it adds substantial value rather than rewriting other sources. Your own data, prices, case results and first-hand experience are what no competitor can copy and no model already holds.
• Show who wrote it. Google's Who, How and Why test asks whether visitors can tell who created the content, how it was produced and why, and it warns that fabricated author profiles are a form of deception that marks a page as low quality.
• Use structured data to restate what is visible. Google describes it as a standardised format that gives explicit clues about a page's meaning and says the markup must match content visible to the user. It helps machines read facts you already publish.

Step three: build a brand the web agrees on

AI answers are assembled from many sources, so the more consistently your site, your profiles and third-party coverage describe who you are and what you do, the easier you are to name correctly. If an assistant gets a fact about you wrong, fix the page it learned from.

Google's list of worthwhile practices for AI features includes keeping Merchant Center and Business Profile information up to date, and its helpful-content guidance calls trust the most important part of E-E-A-T, while noting E-E-A-T itself is not a specific ranking factor. Earned mentions in press, reviews and industry lists widen the set of sources that describe you.

Common LLM optimisation mistakes

• Keyword stuffing. The GEO study found little to no gain from it, and it breaches Google's spam policies.
• Mass-producing AI pages. Google's spam policies name generating many pages with AI tools without adding value as scaled content abuse, and count attempts to manipulate generative AI responses in Search as spam.
• Blocking a search crawler by accident. A robots.txt, CDN or firewall rule meant for training bots that also stops OAI-SearchBot, PerplexityBot or Claude-SearchBot.
• Treating robots.txt as security. The standard says it is not access authorisation, and user-triggered fetchers may ignore it.
• Expecting a stable scoreboard. Google says AI Overviews often do not trigger and that AI Mode and AI Overviews may use different models, so the links they show vary.

How to measure LLM optimisation

With no rankings to track, measure on four fronts.

• Search Console. Google counts AI Overviews and AI Mode traffic inside the Performance report under the Web search type, mixed with other search traffic; even after Google moved inline links higher in AI Mode and AI Overviews, it does not show which clicks came from AI results.
• Analytics. Google suggests tracking conversions and time on site in Google Analytics. Segment visits referred by AI assistants so they are not lost in direct traffic; a recent ChatGPT referral drop in GA4, which appears to have come from inline links losing their tracking tag, shows how easily those visits land in Direct.
• A prompt tracker. Write down the questions your buyers ask, run them on a schedule in ChatGPT, Gemini, Perplexity, Claude and Copilot, and record whether you are cited, mentioned or recommended. Answers vary run to run, so watch the trend.
• Server logs. Check that OAI-SearchBot, PerplexityBot, Claude-SearchBot and Googlebot are actually visiting, and verify them against the IP lists each vendor publishes; a search crawler that never appears is a sign something is blocking it.

Frequently asked questions

Is LLMO replacing SEO?

No. Google says SEO fundamentals remain worthwhile for its AI features, and a live AI answer can only cite a page the engine could crawl.

Will blocking GPTBot remove me from ChatGPT?

Not from search. OpenAI says the settings are independent: OAI-SearchBot controls whether you appear in ChatGPT search answers, and GPTBot governs training.

Does Google-Extended affect my Google rankings?

No. Google says it does not affect inclusion in Google Search and is not used as a ranking signal; it governs Gemini training and grounding.

How long does LLM optimisation take?

No vendor promises a date for a mention. The only published timings are mechanical: about 24 hours for OpenAI's search systems to reflect a robots.txt change, and several days to several months for a Google recrawl.

Do I need an llms.txt file?

Google says there are no additional technical requirements for its AI features, and none of the crawler documentation from OpenAI, Anthropic or Perplexity mentions one. Spend the time on crawl access and content.

How to get started this month

1. Open your robots.txt and confirm OAI-SearchBot, PerplexityBot, Claude-SearchBot and Googlebot are allowed, then check your CDN and firewall rules and your server logs.
2. Write down the questions buyers ask before choosing a business like yours, in the form they would type into an assistant.
3. Rewrite your three most important pages answer-first, with a sourced figure or two, defined terms and a short FAQ, and move key facts out of images and PDFs into text.
4. Check that your business details match across your site, Business Profile, directories and any structured data.
5. Run your buyer questions through the major assistants, record who is cited, and note the baseline in Search Console and your analytics.

None of this needs new technology, and all of it doubles as ordinary SEO work. The fuller method, including how generative engines pick sources, crawler access, structured data and a 90-day plan, is in our GEO playbook.