What is LLM SEO?
Before changing your content, understand how an AI answer can find and use it.
Steen Stones · Reviewed
LLM SEO means large language model search engine optimisation. The aim is to make your business understandable and useful as a source in AI answers. Ross Simmonds describes that goal as earning citations and recommendations, while Google says ordinary SEO remains relevant to its AI search.
| Process | What happens | Relevant control or condition |
|---|---|---|
| Training: OpenAI | Learning patterns changes a model’s internal settings. | OpenAI separately identifies GPTBot for possible training. |
| ChatGPT search: OpenAI | Website access for summaries and snippets. | Allow OAI-SearchBot; decide training access separately. |
| Google AI search: Google | Reads retrieved pages and writes responses with supporting links. | Pages must be indexed and snippet-eligible, and the site included in Search generative AI features. |
What does LLM SEO try to achieve?
LLM SEO is about whether AI systems can understand, cite and recommend your business. Ross Simmonds’ definition concerns that visibility, rather than using AI to write your website copy.
Start with the questions buyers ask when making a shortlist. Read the answers, noting where your company appears and which sources support what is said about it.
Use that record to agree what needs improving and how you will judge the work.
How do training and web search differ?
Decide search access and training permissions separately
Training
- The model learns patterns from material.
- Learning changes the model’s internal settings.
- OpenAI identifies GPTBot for possible training.
Web search
- Pages are retrieved to answer a question.
- Information from those pages supports a response.
- OpenAI identifies OAI-SearchBot for summaries and snippets.
Blocking potential training is a different decision from allowing search access. These crawler names and controls are OpenAI’s.
Sources: OpenAI: model development; OpenAI: publisher controls; Google: retrieved pages
ChatGPT learns patterns from text, images, audio and video, according to OpenAI. Training changes the model’s internal settings to reflect those patterns.
Looking up web information is a different job. Google says its AI systems read retrieved pages and write responses with supporting links, using its core Search ranking and quality systems.
Ask which process the proposal addresses, because training a model and looking up a page for an answer are different things.
Which website controls should you check?
For ChatGPT summaries and snippets, OpenAI says to allow OAI-SearchBot. It says GPTBot can be disallowed to exclude pages from possible training, so check these permissions separately.
For Google’s AI search features, your page must be indexed and eligible for a Search snippet. Your site must also be included in Search generative AI features in Search Console.
Ask whoever manages your website to check those conditions before you commission a rewrite. Decide what access to allow once each crawler’s purpose is clear.
Should you write specifically for LLM visibility?
Google recommends useful content with clear headings and expert knowledge or experience beyond common knowledge. It sets no ideal page length for its generative search features.
Choose a buyer question, answer it directly and give readers the evidence they need. Keep important conditions beside the claims and check that the page reads well on a phone.
Google needs no special schema.org markup for its AI features. It also says special AI text files or Markdown are unnecessary for appearing in Search. Keep that advice scoped to Google.
What should an LLM visibility agency measure?
Repeat the checks, because OpenAI says the same question can produce different answers. Ross Simmonds makes the related point that prompt-panel tracking measures a sample, not a complete count.
Use Google’s Generative AI performance report in Search Console for its AI features. ChatGPT adds a referral marker to URLs, helping you identify visits from its search results.
Save the question, engine and date with each answer, and distinguish a citation from a recommendation of your company. Compare those observations with enquiries and revenue, making clear what your sample covers.
That’s a sample, not a census.
