AI search
LLM SEO: how AI finds and cites sources
The quick answer
LLM SEO is helping AI tools like ChatGPT find your business and show your website as a source. An LLM (large language model) is the kind of AI behind ChatGPT. Getting shown takes four steps: the tool reads your page, finds it, uses it and links to it. No step guarantees the next, so nobody can promise a recommendation.
On this page
Key takeaways
- An AI tool can answer from what it learned in training or from web pages it looks up while it answers. The two behave differently.
- Providers run separate page-reading programs, called crawlers, for training and for search, so blocking one doesn’t have to block the other.
- For Google’s AI features, a page must be stored in Google’s search system and allowed to show a short text preview. You don’t need a special AI file or any extra page-labelling code.
- In our Sydney trades check, directories were linked in 8 of 50 ChatGPT answers and 14 of 50 Perplexity answers.
- A citation shows that the tool displayed a page as a source. It doesn’t prove that every claim next to it is backed up.
Here’s an example. Someone in Marrickville asks ChatGPT for a plumber. It names three businesses and links to a couple of websites but yours isn’t one of them. LLM SEO is the work of helping AI tools like ChatGPT find your business and show your website as a source. An LLM (large language model) is the kind of AI behind ChatGPT.
This guide explains how those tools find and pick pages so you can tell which parts you can change. Our AI SEO agency uses the same map when we plan work for Australian businesses. For step-by-step website checks, read our AI SEO guide, because this page covers how the tools choose and show sources and not what to fix first.
What does this mean for your business?
You can’t control which businesses an AI tool names, though you can remove the common reasons a business gets missed. These are the practical steps that come out of what the providers say.
- Let the search tools read your site. Ask your web developer to check that you haven’t blocked them by accident and keep your pages public.
- Write pages that answer a real customer question. Say what you do, where you work, who you do it for and what it costs or roughly what it costs.
- Keep the same details everywhere. Directories and review sites sometimes show up as the source in an AI answer, so your listings should match your website.
- Skip the tricks. Google says you don’t need a special AI file or any extra page-labelling code to appear in its AI features.
- Ask the tools your own customer questions each month. Save the answer, the date and every link, then count how often your site is linked and how often a directory is linked instead.
The rest of this guide explains why each step matters.
What is LLM SEO?
A large language model is software that writes text by using patterns it learned from a huge amount of writing.9 ChatGPT and Claude are built on large language models. LLM SEO is about whether accurate information about your business is available when one of these tools answers a relevant question.
You may also hear the work called GEO, which stands for generative engine optimisation. It covers how AI-written answers describe and link to a business. Our guide on what is GEO explains that wider job. This page sticks to how the tools find and display sources.
Three words to know. Training is when an AI learns from a large body of text before it’s released. Retrieval is when a tool looks up current web pages while it answers and uses them in its reply. A citation is a link shown with the answer.
What are the four gates between your page and an AI answer?
Think of your web page as a little paper boat floating down a canal towards the sea. On the way, it has to get through four gates that can each shut on their own. The first gate decides whether the boat can get in at all and the second decides whether anyone finds it. The third decides whether the AI picks the boat up and uses it in an answer, while the fourth decides whether the AI shows a link back to it.
Two words come up in the table below. A crawler is a program that visits web pages by itself, a bit like a robot reading your site. A search index is a stored list of web pages that a provider keeps so it can look things up quickly.
Scroll across to read every column
| Gate | The question | Who decides | What you can change |
|---|---|---|---|
| Access | Can the tool read your page? | Your site’s settings and the provider’s crawler | Let the search crawler in and keep your pages public |
| Finding | Does the tool find your page for this question? | A search index kept by the provider or a search partner | Write clear pages that answer the question |
| Selection | Does the tool use your page in its answer? | The tool, at the moment it answers | Be relevant and clear and state facts that someone can check |
| Citation | Does the tool link to your page? | The tool’s own rules for what it shows | Not much directly, so check each answer |
Getting through one gate doesn’t guarantee the next. The picture below shows the paper boats going through the four gates, marked Access, Finding, Selection and Citation. Each gate is a separate check before a page is linked.
Does an AI answer come from training or a live search?
It can come from either and the tool decides which. Think of a person answering a question from memory or checking their phone first. A model can write an answer from patterns it learned during training, while Google describes this as predicting likely sequences of words from its training data.9 Tools that can search the web can also look up current pages, while the provider pages below show that search access and training collection are separate jobs.
OpenAI runs two crawlers. OAI-SearchBot finds sites to show in ChatGPT search, while GPTBot collects content that may be used for training. You can allow one and block the other.4
A small file called robots.txt tells crawlers which parts of your site they may visit. OpenAI says its search systems can take about 24 hours to adjust after you change that file.4 That timing is about access rules and isn’t a promise to cite your new content or to retrain a model.
Which questions are likely to trigger a live search?
Providers don’t publish a full rule. Questions that depend on current facts, such as prices, opening hours, local businesses or recent changes, are the likeliest to need a live lookup, because a model can’t know them from training alone. That is our reading and not a rule the providers state. It’s why local service businesses have more reason to care about live search than about training.
How do ChatGPT, Google, Perplexity, Claude and Copilot find your pages?
Each provider explains its crawlers and the controls it gives site owners, yet none of the pages we read gives a full list of search partners. Here is the short version and the table further down has the detail.
- ChatGPT. OAI-SearchBot finds sites for ChatGPT search and GPTBot collects content for training.4
- Google AI Overviews and AI Mode. A page must be stored in Google’s search index and allowed to show a short text preview called a snippet.1
- Perplexity. PerplexityBot finds and links sites in its search results and isn’t used to collect content for training AI models.5
- Claude. Claude-SearchBot looks at content to improve search results and ClaudeBot collects content for training.6
- Microsoft 365 Copilot Chat. Copilot writes a short search query and sends it to Bing.7
Google also says its AI features need no special AI files or extra page-labelling code. You may have heard of a file called llms.txt, a proposed way to guide AI tools that Google Search ignores.12 Our guide to AI search engines in Australia covers which ones Australians use.
How does one question become several searches?
Some tools turn one question into several searches before they answer. Google calls this query fan-out and says its AI Overviews and AI Mode may run several related searches across subtopics and data sources.1
Here’s an example of the idea. Say someone asks, “Who can fix a burst pipe in Marrickville on a Sunday?” The tool might look for emergency plumbers in the Inner West and then for after-hours pricing and for licence information. A page that covers only one of those parts has fewer chances to be selected.
Microsoft shows the idea more openly. Copilot Chat builds a short query from a few words in your question and sends it to Bing. After a web-search reply a Sources button shows the exact query it sent and the sources it used.7 You can use it to see what the assistant searched for.
Why might an assistant cite a directory instead of your site?
Directories are websites that list many businesses in one place and they can answer a question well. Assistants link to what they looked up, so a business can be named in an answer, while the link goes to a directory, a review site or a council page. A name and a link are separate events.
Kavera note: In our Sydney trades AI search index we saved 197 completed answers from 200 planned checks across 50 questions. Directory links appeared in 8 of 50 ChatGPT answers and 14 of 50 Perplexity answers, while Hipages was linked in 12 of 50 Perplexity answers. We matched a business’s own website to 140 of 144 name appearances in ChatGPT. That’s a minimum, because links we couldn’t trace weren’t counted.
In the same index, ChatGPT linked a government page in 34 of 50 completed answers and Service NSW in 21 of 50. Government pages appeared in 8 of 49 completed Google AI Mode answers and 4 of 48 completed Google AI Overview answers. These numbers describe one set of questions and not what any tool prefers in general.
We haven’t tested whether joining a directory changes an answer, so treat directory listings as places to keep your details accurate and not as a trick. Backlinks (links from other websites to yours) and brand mentions are common advice in LLM SEO articles, yet none of the provider pages we read describe them as something that decides citations.
Is LLM SEO the same as GEO, AEO or AI SEO?
The names overlap and providers use them differently. Here is how we use them:
- LLM SEO is about how language models find and use sources.
- GEO (generative engine optimisation) is work on generated answers.
- AEO (answer engine optimisation) is work on direct-answer formats.
- AI SEO can mean using AI tools to do SEO work or improving your visibility in AI answers.
Our guides explain each one: what is GEO, what is AEO and AI SEO. Which label a provider uses matters less than what they will change on your website and how they will measure it.
What does an AI citation prove?
A citation proves that the tool showed that page as a source. It doesn’t prove that every claim next to it is backed up or that the page was important. Microsoft says its citation counts reflect how often pages are cited and not page importance, ranking or placement.8
Open the link and look for the sentence that supports the claim. If the page doesn’t say it, the citation is weak evidence.
Kavera note: In two checks about two months apart, our client SHORE went from absent to first in one Sydney builder shortlist. The later answer linked its website. We built SHORE’s website and run its SEO. It was one question and two answers, so we don’t claim it shows which change caused the result. The full comparison is in our ChatGPT SEO guide.
Answers also vary, so repeat a check before you treat it as a pattern and save the question, the date and the full answer each time.
Does blocking training also block search visibility?
Not necessarily, because the main providers use separate crawlers for search and for training.
- OpenAI. Blocking OAI-SearchBot stops your site being included in ChatGPT search answers, although it can still appear as a navigational link. Blocking GPTBot signals that your content shouldn’t be used for training.4
- Perplexity. PerplexityBot finds and links sites in search and doesn’t collect content for training AI models. Perplexity recommends allowing it in robots.txt.5
- Anthropic. Turning off Claude-SearchBot can stop your content being added to its search system, while ClaudeBot relates to training.6
Requests made for one person at a time behave differently. OpenAI says ChatGPT-User may not follow robots.txt, while Perplexity says Perplexity-User generally ignores those rules because a person asked for the page.45 Check each provider before you decide what to allow and ask your developer to test the file.
Can you use an LLM to do SEO work?
Yes, for research, grouping and drafting, as long as a person checks the output. That is a different question from how LLMs choose sources for their answers. Our guide to AI SEO tools covers which tools suit which job. No single model is best at every task, so test any of them on one real page before you rely on it.
What each provider says, in detail
This table separates what each provider says about its crawlers, as at 19 October 2026, from the details its pages don’t share. None of the pages we read gives a full list of search partners. Don’t read a single row as covering a whole product family, because the Copilot row describes Microsoft 365 Copilot Chat only.
Scroll across to read every column
| Product | What they say about search | What they say about training or other use | Not shared on the pages we read |
|---|---|---|---|
| ChatGPT | OAI-SearchBot surfaces sites in ChatGPT search4 | GPTBot is for content that may be used in training. ChatGPT-User runs when a person triggers an action4 | Every index or partner behind a given answer |
| Google AI Overviews and AI Mode | A page must be stored in Google’s search system, called its index. It must also be allowed to show a short preview called a snippet1 | Google-Extended controls specified Gemini training and the use of retrieved Search information in replies, called grounding. It doesn’t control inclusion in Google Search3 | Which pages Google selects for an individual answer |
| Perplexity | PerplexityBot surfaces and links sites in its search results5 | PerplexityBot isn’t used to collect content for training AI models. Perplexity-User fetches pages when a user asks5 | How it ranks sources |
| Claude | Claude-SearchBot analyses content to improve search results6 | ClaudeBot collects content for training. Claude-User retrieves pages when a person asks6 | Partner indexes |
| Microsoft 365 Copilot Chat | Copilot generates a short query and sends it to Bing7 | Not covered on the page | How it ranks the results |
What should you check in the next AI answer?
Follow one citation and find the sentence that supports the answer’s claim. Write down the question, the linked page and any gap between the two. A link to a directory and a link to your own service page tell you different things about where the tool found its evidence.
For website changes, use our AI SEO guide. Our AI SEO agency can turn that review into a scoped page and search-access plan.
Our SEO audit is free. Implementation is separate unless retained. Compare the SEO packages for monthly scope and terms. No SEO package guarantees rankings, leads or revenue.
Good questions
Common questions Clear answers
How do language models find information about my business?
Some answers come from patterns the AI learned during training. Tools that can search the web also look up current pages while they answer, using search tools and crawlers (programs that read web pages automatically). Google keeps its own stored list of pages, called an index. Keep your business information accurate and remember that letting a tool search your site is different from letting it collect your site for training. Updating a page doesn’t retrain a model straight away.
What makes an AI assistant cite one website instead of another?
The provider pages we read explain what a site needs before it can be considered. Google needs the page to be stored in its search system and allowed to show a short text preview. The other providers describe their own crawlers. Those conditions let a page be considered but they don’t explain every choice. In our Sydney check the tools linked to business sites, directories and government pages.
Does my website need to be indexed by Google to appear in AI answers?
For a supporting link in Google’s AI Overviews or AI Mode, yes. Google has to store the page in its search system first, which is called indexing. The page must also be allowed to show a short text preview called a snippet. The other assistants describe their own search tools and crawlers, so check each provider separately.
Do backlinks and brand mentions help language models find my business?
None of the provider pages we read name backlinks (links from other websites to yours) as something that decides AI citations. Directories and review sites do appear as sources in some answers, so accurate listings can help a tool find the right details. We haven’t tested whether adding a listing changes an answer and we don’t promise that it will.
How can I make my business information easier for AI systems to understand?
Put plain facts on a page that answers a real customer question, such as what you do, where you work, who you do it for and what it costs or roughly what it costs. Keep the same details on your listings. Google says you don’t need a special AI file or extra page-labelling code, so spend your effort on clear text and accurate information.
How can I check whether AI assistants are citing my website?
Write a fixed list of customer questions and ask each assistant. Save the full answer, the date and every link it shows. Count how often your site is linked and how often a directory is linked instead, then repeat every month with the same questions. Google and Bing also offer reports for their own AI features.
Sources
- AI features and your websiteGoogle Search Central. Accessed .
- Optimizing your website for generative AI features on Google SearchGoogle Search Central. Accessed .
- Google’s common crawlersGoogle Search Central. Accessed .
- Overview of OpenAI crawlersOpenAI. Accessed .
- Perplexity crawlersPerplexity. Accessed .
- Does Anthropic crawl data from the web and how can site owners block the crawler?Anthropic. Accessed .
- How web search works in Microsoft Copilot Chat and agentsMicrosoft Support. Accessed .
- Introducing AI Performance in Bing Webmaster Tools Public PreviewMicrosoft Bing. Accessed .
- Google Search’s guidance on using generative AI content on your websiteGoogle Search Central. Accessed .
