LLM SEO: How to Optimize Content for LLMs

LLM SEO is the practice of shaping your content so that large language models can use it as a source. When someone asks ChatGPT, Gemini, Perplexity, Copilot or Google AI Mode a question, the model writes an answer and credits a few sources. LLM SEO is the work of being one of those sources, and being described correctly when you are.

You’ll see other names for the same thing. Some people say LLMO (large language model optimization). Others say generative engine optimization, or GEO, or AEO. The labels differ, but the goal is the same: get your brand and your pages into AI-written answers.

What makes LLM SEO its own topic is the unit of competition. Classic SEO is a contest between pages. A page ranks, a person clicks, and the person reads whatever they want on it. LLM SEO is a contest between passages. The model reads the page for you and keeps one or two sentences. If those sentences are good, you get cited. If they aren’t, a competitor’s sentences get cited instead, even when your page ranks higher.

It also changes what counts as your content. LLMs don’t only read your website. They read reviews, forums, news stories and lists that mention you. So LLM SEO has an on-site half (pages that are easy to quote) and an off-site half (a reputation that other sites back up).

This guide focuses on the on-site half: how LLMs read and use your text, and how to write so they use it well. If you want the full playbook that also covers each platform, start with our guide on optimizing for AI search step by step. If you want the writing method in depth, our pillar on writing answer-first content is the companion to this page.

OpenAI. OpenAI runs three. OAI-SearchBot is for search, and OpenAI says sites that opt out of it “will not be shown in ChatGPT search answers, though can still appear as navigational links.” GPTBot collects content for training foundation models, and blocking it only signals that your content shouldn’t be used for training. ChatGPT-User fetches a page when a user’s action asks for it, and OpenAI notes robots.txt rules may not apply to it. OpenAI also says robots.txt changes can take about 24 hours to take effect for search.

Perplexity. PerplexityBot surfaces and links websites in Perplexity’s results and, per Perplexity, isn’t used to crawl content for foundation models. Perplexity-User fetches pages for user actions and generally ignores robots.txt. Changes can take up to 24 hours to show up.

Google. AI Overviews, AI Mode and Gemini’s search grounding use Google’s normal index. Google says a page only needs to be indexed and eligible to show with a snippet, with no extra technical requirements. If Googlebot can crawl and index you, you’re eligible.

Microsoft. Copilot’s answers are grounded in Bing’s index. In May 2026, Microsoft described a grounding index that builds on the same crawlers and quality signals as search but adds a new layer for answers (more on that below).

The practical rule is simple. Allow the search crawlers (OAI-SearchBot, PerplexityBot, Googlebot, Bingbot). Decide on training crawlers like GPTBot separately, as a business choice. Blocking a training crawler doesn’t remove you from AI search, but blocking a search crawler does.

Once a page is fetched, the system breaks it into units it can compare against the question. Engineers call these units chunks, passages or spans. Each platform does it differently, and none publishes its exact method, but the published research and platform statements describe the same basic pattern.

How big is a chunk? In its write-up on retrieval, Anthropic says chunks are “usually no more than a few hundred tokens”. That’s roughly a few short paragraphs. Microsoft goes further in its May 2026 post on grounding and says the unit of value is “discrete, supportable facts with clear provenance” rather than whole pages.

How are chunks matched? Two methods usually run side by side. Keyword matching (often an algorithm called BM25) finds chunks that share exact words with the question. Embeddings turn text into numbers that capture meaning, so a chunk about “setting up a business” can match a question about “registering a company.” Anthropic describes combining both and merging the results.

What goes wrong? Cutting text into pieces strips away context. Anthropic’s example is a chunk that reads “The company’s revenue grew by 3% over the previous quarter.” Out of context, that chunk doesn’t say which company or which quarter, so it’s hard to retrieve for a question about a specific company, and risky to quote.

Anthropic’s fix, which it calls contextual retrieval, adds a short explanation of where the chunk sits before indexing it. The results show how much missing context costs.

You can’t control how ChatGPT or Perplexity chunk your page. But you can control whether your passages need repair. A passage that already names its subject, its time frame and its numbers doesn’t lose anything when it’s cut out. That’s the single most useful idea in LLM SEO, and we’ll turn it into a writing method in a moment.

Microsoft makes the same point from the platform side. Its May 2026 grounding post says “chunking/transformations must preserve meaning and claims used in the answer.” Jordi Ribas, who leads search and AI at Microsoft, summed up the shift when the post went live.

Research on retrieval units points the same way. A 2023 study, Dense X Retrieval, found that indexing text as propositions, which the authors describe as “atomic expressions within text,” significantly outperformed passage-level units in retrieval tasks. In plain words: the smaller and more self-contained a statement is, the easier it is for a system to find the right one.

Google has been unusually direct on this topic, and its position is worth knowing word for word, because a lot of LLM SEO advice contradicts it.

In July 2025, Gary Illyes told a Google Search Central event that “to get your content to appear in AI Overview, simply use normal SEO practices. You don’t need GEO, LLMO or anything else.” In December 2025, Danny Sullivan and John Mueller spent a podcast episode making the same case: SEO for AI is essentially SEO, and Sullivan suggested GEO is best treated as a subset of it.

Then, in May 2026, Google published a full help document called “Optimizing your website for generative AI features on Google Search”, announced on the Search Central blog. It includes a mythbusting section that lists what you don’t need to do. Search Engine Roundtable’s coverage called out the same items.

The positive advice is just as useful. Google says its AI systems look at a variety of sources, so a unique viewpoint helps. It warns against recycling what others have already said or what “could easily be produced by a generative AI model.” It separates commodity content, based on common knowledge anyone could write, from non-commodity content that “provides unique expert or experienced takes that go beyond common knowledge.”

That’s a sharper instruction than it looks. If an LLM can already write your paragraph from its training data, it has no reason to cite you for it. The passages that earn citations add something the model can’t produce alone: your prices, your process, your test results, your local knowledge.

Google also warns that creating separate pages for every query variation, mainly to manipulate AI responses, violates its scaled content abuse policy. So “one page per prompt” isn’t LLM SEO. It’s a spam risk.

Google Search Central · YouTube

Google Search Central · SEO, AIO, GEO, your site, & third-party support to optimize for LLMs · watch on YouTube

In that Search Off the Record episode, published in January 2026, Danny Sullivan addresses the most common LLM SEO tip directly, which brings us to the biggest debate in this field.

“Chunk your content” is probably the most repeated LLM SEO tip of the last two years. The idea is that LLMs like bite-sized text, so you should break pages into many tiny sections. Google has pushed back hard.

In the January 2026 podcast, Danny Sullivan said he keeps hearing advice to “turn your content into bite-sized chunks, because LLMs like things that are really bite size”, and responded: “we don’t want you to do that.” He added: “We don’t want people to have to be crafting anything for Search specifically.” Google’s May 2026 guide put it in writing: “There’s no requirement to break your content into tiny pieces for AI to better understand it.”

On the other side, Mike King of iPullRank published a detailed rebuttal in January 2026. His argument is that retrieval systems select passages rather than whole pages, so passage structure affects visibility whether Google likes the word or not. In a test with his own tool, splitting a paragraph that mixed two topics into two paragraphs raised the similarity score for one topic by 19.24%, and adding a heading raised the other by a further 17.54%. Those are similarity scores from his own tool, not measured changes in AI citations, so read them as a demonstration rather than proof.

Our position: both sides are describing the same good page

When you look past the headlines, the two camps agree on more than they admit.

Google’s objection is to fragmentation: pages chopped into dozens of thin, two-line sections written for machines. That makes pages worse for people, and Google says its systems “improve… to reward content written for humans.” Nobody serious defends that.

King’s point is about focus: a passage that mixes two ideas matches neither well. That’s not a machine trick. It’s what editors have always called “one idea per paragraph.”

Microsoft’s own guidance lands in the same place. It warns that long walls of text “blur ideas together and make it harder for AI to separate content into usable chunks”, and that headings “act like chapter titles that define clear content slices.” That’s an argument for structure, not for fragments.

So here’s the rule we use with brands: write for a busy human reader who skims headings and reads the first sentence of each section. If that reader gets a complete, accurate answer from every section, the page is already well chunked. You never need to write a second version for robots.

Headings do two jobs for LLMs. They mark where one passage ends and the next begins, and they carry meaning that gets attached to the passage under them.

Microsoft’s guidance says headings act “like chapter titles that define clear content slices” for AI. It also notes that question-and-answer formats work well, because assistants “can often lift these pairs word for word into AI-generated responses.” King’s test above found that adding a heading raised a passage’s similarity score on its topic.

The pattern is simple: question heading, answer sentence, supporting detail.

Back to our hypothetical company registration firm. A heading like “Timeline” says very little. “How long does company registration in Thailand take?” says exactly what the section answers, uses the words people type into ChatGPT, and gives the retriever a strong match for fan-out searches on timing.

Use the customer’s words, not yours. If customers ask “how much does it cost,” don’t title the section “Investment.” Your inbox, your sales calls and your LINE chats are the best source of heading wording you’ll ever find.

Don’t stuff every heading with the keyword. Google’s guide says AI systems understand synonyms and general meaning. Vary the wording the way a person would.

Answer in the first sentence, every time. This is the habit that ties everything together, and it’s the subject of our full guide to answer-first content. If the heading asks a question and the first sentence answers it, the passage is quotable even when it’s lifted out alone.

You can test this quickly with our free AEO content checker, which flags sections whose opening sentence doesn’t answer the heading.

Length is where LLM SEO advice gets the most contradictory, so let’s look at the only large dataset we have and then at what Google says.

SE Ranking analyzed 216,524 pages from 129,000 domains in 20 niches and compared them with how often ChatGPT cited them. Search Engine Journal reported the results in November 2025. Two findings are about length.

Sections of 120 to 180 words between headings averaged 4.6 citations, while sections under 50 words averaged 2.7. Articles over 2,900 words averaged 5.1 citations, while articles under 800 words averaged 3.2.

Notice what that says about the chunking debate. The tiny-section approach, with lots of headings and a couple of lines under each, is exactly what performed worst. Sections with enough room to answer and explain did better.

Now the caveats. Search Engine Journal’s write-up calls these “directional correlations rather than proof.” Long articles tend to come from sites with more links and more authority, and SE Ranking found referring domains were the strongest predictor of citations. Length may be riding along with authority rather than causing anything.

Google’s guide is blunt: “There’s no ideal page length.” Make pages for your audience.

So here’s how we translate it into practice:

A section should be long enough to answer and explain. For most questions that’s a one-sentence answer plus two to four short paragraphs. Something in the 120 to 180 word range is a reasonable default, not a rule.

A page should be as long as the topic. A pricing page can be short. A guide that answers 15 related questions will be long, and that’s fine, as long as each section stands on its own.

Don’t pad. Padding pushes your best passages further down the page, and the position data says that’s where citations thin out.

LLMs reason about entities: brands, people, products, places and concepts. A page that makes its entities obvious is easier to retrieve, easier to trust and easier to describe correctly.

Here’s how to make your entities clear without writing like a robot.

Use one name, consistently. Pick the official form of your brand name and use it the same way on your site, your Google Business Profile, your social profiles and your directory listings. Variants confuse matching.

Define what you are early. Your About page and service pages should say plainly what the business is, where it operates and who it serves, in one sentence. That sentence is often the one LLMs reuse when they describe you.

Name your people. A named author with a real bio and a team page gives systems a person to connect expertise to. This matters for the trust signals Google calls E-E-A-T, and it gives AI answers something concrete to say.

Connect related entities. If your firm handles company registration, work permits and accounting, say so on each relevant page and link the pages together. Internal links help crawlers find pages, and Google lists findability through internal links as a basic requirement for AI features.

Mark it up where it fits. Organization, Person, Product and LocalBusiness structured data describe entities in a form machines read directly. Google says structured data isn’t required for AI features, but it should always match the visible text, and it helps with rich results. Our guide to entity SEO goes deeper on this, and our schema generator builds the markup.

For Thai businesses, entity clarity has a bilingual twist. Make sure your Thai name, English name and any abbreviation all appear together somewhere official, such as the About page, so a model can connect “บริษัท” mentions on Pantip with English mentions on review sites.

Formatting is where LLM SEO advice often turns into box-ticking. Here’s what the platforms actually say.

Lists and tables. Microsoft says “bulleted lists, numbered steps, and comparison tables break complex details into clean, reusable segments.” That’s true when the content is a list or a comparison by nature: steps in a process, prices for three packages, features of two products. A table of three packages with price, duration and what’s included is about as quotable as content gets.

Q&A pairs. Microsoft also says assistants can lift question-and-answer pairs “word for word.” A genuine FAQ section that answers real customer questions can help. A wall of invented FAQs written to catch prompts won’t. Our guide to optimizing FAQs for AI covers how to write them well.

Schema. Google is clear that “structured data isn’t required for generative AI search, and there’s no special schema.org markup you need to add.” Microsoft is warmer and describes schema as data “machines can interpret with confidence.” The data is mixed. In SE Ranking’s ChatGPT study, pages with FAQ schema averaged 3.6 citations against 4.2 without, which the researchers put down to FAQ pages often being simpler support pages.

Our rule: format for the shape of the content. Use a table when you’re comparing things, a list when there’s a sequence, and paragraphs when you’re explaining. Add schema that matches what’s visible, using our guide to schema markup for AEO, and don’t expect it to rescue weak text.

Good writing can’t be quoted if the machine never sees it. These are the technical checks we run before touching a single paragraph.

1. Crawler access. Confirm robots.txt allows OAI-SearchBot, PerplexityBot, Googlebot and Bingbot, and that your CDN or firewall isn’t blocking them. Check server logs for 200 responses to those user agents. We covered the crawlers above.

2. Text in the HTML. Google lists keeping important content available as text among its best practices for AI features. You can’t count on every AI crawler running your JavaScript, so main content that only appears after scripts run is a risk. View the page source, not the rendered page, and check that your headings and key paragraphs are there.

3. Nothing important hidden. As Microsoft warns, content inside tabs, accordions and images may be skipped. Keep core answers visible.

4. Indexing basics. For Google’s AI features, the page must be indexed and eligible for a snippet. A noindex tag, a nosnippet directive or a canonical pointing elsewhere can quietly take a page out of AI answers.

5. Speed and structure. SE Ranking linked fast First Contentful Paint with more citations, so keep pages lean. Use one H1, logical H2 and H3 nesting, real HTML tables for tabular data, and descriptive link text. These help both retrieval and accessibility.

What about llms.txt?

llms.txt is a proposed file that lists a site’s key pages in Markdown for AI tools. It’s often sold as essential LLM SEO. The evidence says otherwise. Google’s guide says AI text files “will neither harm nor help your site’s visibility or rankings in Google Search, as Google Search ignores them.” SE Ranking found llms.txt had negligible impact on ChatGPT citations.

It’s cheap to add and some tools may read it, so it isn’t harmful. Just don’t expect it to fix anything on this list. You can build one in minutes with our free llms.txt generator.

Everything so far has been about your own pages. But a large share of what LLMs say about any brand comes from other websites, and that changes where your effort should go.

Profound analyzed 680 million citations from August 2024 to June 2025. Among each platform’s ten most-cited sources, Wikipedia made up 47.9% of ChatGPT’s top-10 citations, Reddit made up 46.7% of Perplexity’s and 21.0% of Google AI Overviews’.

Ranking in Google doesn’t guarantee a place either. Ahrefs ran 15,000 long-tail queries and found that, on average, only 12% of URLs cited by ChatGPT, Gemini, Copilot and Perplexity ranked in Google’s top 10 for the same query. Perplexity overlapped most at 28.6%. ChatGPT’s in-text citations overlapped just 8.0%.

SE Ranking’s ChatGPT study adds more weight. Referring domains were the single strongest predictor of citations, and brands with heavy Reddit and Quora mention volume were cited far more often than those with little presence.

For LLM SEO, this means two things.

Your passages compete with other sites’ passages about you. If a Reddit thread or a review site describes your service better than your own page does, the model may quote them instead. Make your own page the most complete, accurate source on your own business.

Earn mentions honestly. Answer questions on forums like Reddit and Pantip as a disclosed expert, get listed on the review sites your customers use, and give journalists real data. Google’s guide warns that “seeking inauthentic ‘mentions’ across the web isn’t as helpful as it might seem,” and its spam systems still apply to AI features. Paid listicles and swapped mentions are a risk, not a strategy. We cover the off-site side in our guide to getting cited by AI.

Here’s the process we use to optimize one existing page for LLMs. It works for a service page, a product page or a guide, and it takes a few hours per page, not weeks.

Step 1: Take a baseline. Pick 10 to 20 prompts the page should win. Run them in ChatGPT, Gemini, Google AI Mode and Perplexity, three times each. Note who gets cited and which passages get quoted.

Step 2: Check access. Make sure the search crawlers can fetch the page and that the main text is in the HTML. If this fails, stop and fix it first.

Step 3: Map questions to sections. List the questions from step 1, plus the fan-out questions behind them. Give each real question its own heading. Merge questions that have the same answer. Cut sections that answer nothing.

Step 4: Write the answer first. Under each heading, write one sentence that fully answers the question, ideally 6 to 20 words long, the range Shashko’s data found for almost all AI Mode and Gemini citations. Then explain.

Step 5: Make passages self-contained. Replace pronouns with names, add time frames, remove “as mentioned above.” Paste each key paragraph alone into a blank document and check it still makes sense.

Step 6: Add evidence. Swap adjectives for real numbers, name sources inside sentences and add first-hand detail only you can provide. This is the non-commodity content Google asks for.

Step 7: Fix freshness and visibility. Update expiring facts, date them in the text, and move anything important out of tabs and images.

Step 8: Re-test and iterate. After the page is recrawled, run the same prompts. Compare quoted passages before and after. Keep what worked, and rewrite sections where a competitor’s passage still wins.

If you’d rather have us do this across your key pages, that’s what our AEO service covers, from the crawler check to the monthly prompt tracking.

Is it worth it, and where should you start?

For most businesses, yes. OpenAI reported around 900 million weekly ChatGPT users in February 2026, and that’s one platform. And almost every fix in this guide also makes your pages better for human readers and for classic search, which is why Google keeps saying good SEO is the foundation. You’re not building a second website for robots.

Where LLM SEO differs is the level of attention. Classic SEO lets a page be “good overall.” LLM SEO asks whether each passage is good on its own. That’s a higher bar, and most sites still fall short of it, which is exactly why it’s an opportunity.

If you’re starting today: this week, check that OAI-SearchBot, PerplexityBot, Googlebot and Bingbot can reach your site with our AI crawler checker. This month, run the eight steps above on your five most important pages. After that, track the same prompts monthly and keep facts current. Our guides on GEO and how AI search engines work fill in the wider picture.

LLM SEO rewards the same thing good writing always has: saying something true, specific and useful, clearly, in the place the reader looks first. The only difference is that the first reader is now often a machine, and it only has time for one paragraph.

1. Google Search Central: Optimizing your website for generative AI features on Google Search (updated 10 Jul 2026)
2. Google Search Central: AI features and your website (updated 10 Dec 2025)
3. Google Search Central Blog: A new resource for optimizing for generative AI in Google Search (May 2026)
4. OpenAI: Overview of OpenAI crawlers
5. OpenAI Help Center: ChatGPT search
6. Perplexity: Introducing the Perplexity Search API (25 Sep 2025)
7. Perplexity Docs: Perplexity crawlers
8. Microsoft Advertising Blog: Optimizing your content for inclusion in AI search answers (Krishna Madhavan, 8 Oct 2025)
9. Microsoft Bing Blog: Evolving role of the index, from ranking pages to supporting answers (6 May 2026)
10. Aggarwal et al.: GEO: Generative Engine Optimization (KDD 2024)
11. Liu et al.: Lost in the Middle: How Language Models Use Long Contexts (TACL)
12. Chen et al.: Dense X Retrieval: What Retrieval Granularity Should We Use? (2023, revised 2024)
13. Anthropic: Introducing Contextual Retrieval (19 Sep 2024)
14. Practical Ecommerce: Studies reveal AI citation clues (Kevin Indig and Daniel Shashko data, 9 Mar 2026)
15. Search Engine Journal: New data on the top factors influencing ChatGPT citations (SE Ranking study, 26 Nov 2025)
16. Ahrefs: Only 12% of AI-cited URLs rank in Google's top 10 (11 Aug 2025)
17. Profound: AI platform citation patterns (680 million citations, June 2025, updated August 2025)
18. Search Engine Roundtable: Google, don't turn your content into bite-sized chunks (9 Jan 2026)
19. Search Engine Roundtable: Google's mythbusting guide to optimizing for generative AI (18 May 2026)
20. Search Engine Roundtable: Google, normal SEO works to get into AI Overviews (24 Jul 2025)
21. Search Engine Roundtable: Danny Sullivan and John Mueller on SEO for AI (18 Dec 2025)
22. Search Engine Roundtable: Microsoft Bing on search indexing vs grounding indexing (7 May 2026)
23. iPullRank (Mike King): A refutation of misinformation about chunking (15 Jan 2026)
24. eWEEK: ChatGPT nears 1 billion weekly users (27 Feb 2026)
25. Google Search Central on YouTube: SEO, AIO, GEO, your site, and third-party support to optimize for LLMs
26. Jordi Ribas (Microsoft) on X, 6 May 2026

Keep reading

• Content & technicalHow to Write Answer-First Content for AI Search
• AI platformsHow to Optimize for AI Search: A Step-by-Step Guide
• AEO basicsWhat Is GEO? Generative Engine Optimization Explained
• Content & technicalShould You Block AI Crawlers?