How to Optimise WordPress for AI Search (llms.txt and GEO)

This guide is a plain-language, deliberately honest tour of how to make a WordPress site work well with AI search: the answers that ChatGPT, Google's AI Overviews and AI Mode, Perplexity, Gemini and Copilot now put in front of people before they ever reach a list of blue links. It covers what an llms.txt file is and how to add one, the practices grouped under Generative Engine Optimization (GEO), how structured data helps machines understand your pages, and how to control which AI crawlers may read your site. This field goes by several names you will meet online: GEO, Answer Engine Optimization (AEO), and LLM optimisation (LLMO) all describe roughly the same goal, which is being the source an AI answer draws from and cites. It is a fast-moving, half-formed area, so this guide is careful throughout to separate what is genuinely established from what is still marketing and guesswork. It is written for Noiz clients who run their own WordPress site, and it names specific tools only as examples, never as endorsements.

Last reviewed: 27 July 2026, against WordPress 7.0.2 (latest stable). This guide is written for Noiz hosting and is kept current against WordPress. It complements, and does not replace, the official documentation linked below. AI search is changing month to month, so treat any specific figure or tactic here as a snapshot of mid-2026, and weight the durable principles over the fashions.

Official Documentation Reference

• The /llms.txt proposal (llmstxt.org): the original specification by Jeremy Howard, defining the file's exact format and intent.
• Schema.org: the shared vocabulary of structured-data types (Article, Organization, Product, FAQ and so on) that search and AI systems read.
• Introduction to structured data (Google Search Central): how to mark up pages with JSON-LD, and which types Google supports.
• AI features and your website (Google Search Central): Google's own guidance on how AI Overviews and AI Mode use your content, and how to control it.
• Google crawlers overview (Google Search Central): the full list of Google's user agents, including Googlebot and Google-Extended, and exactly what each one controls.
• Reading Settings (WordPress Documentation): the built-in Search engine visibility control and how WordPress signals crawlers.

Prerequisites

• You can log in to your WordPress admin dashboard as an Administrator, which you need to install plugins and change settings.
• The ability to upload a file to your site's web root (the folder that holds wp-config.php, usually named public_html or httpdocs). On Noiz hosting you reach it over SFTP or through the panel File Manager. If you are on a managed plan, you can ask the Noiz support team to place a file for you.
• A site with real, substantive content. None of this helps a thin site. AI answers draw on genuine, useful material, so the single biggest lever is having something worth citing.

What "AI Search" Actually Means, and the Two Ways Your Content Reaches a Model

Before any tactics, it helps to understand the machine you are optimising for, because one distinction explains almost everything that follows. Your content can reach a large language model by two very different routes, and they are governed by different controls.

• Training. A model is trained, once, on an enormous snapshot of the web. Anything learned this way is baked in until the next training run and has no live link back to your page. Crawlers such as GPTBot, ClaudeBot and Google-Extended gather this training data.
• Retrieval (also called grounding). When you ask a modern assistant a current question, it usually runs a live search, fetches a handful of pages, and writes an answer from what it just read, with citations. This is where being "cited by AI" actually happens, and it is driven by a search index and by retrieval crawlers such as OAI-SearchBot, PerplexityBot and Google's own index behind AI Overviews.

Almost all the practical wins in 2026 are about the second route. Retrieval means an AI answer is, underneath, a search result that has been summarised and attributed. That is why the old craft of being findable and trustworthy in search has not died so much as changed shape: if a system cannot find and parse your page at answer time, it cannot cite you.

The Shift You Are Optimising For: Zero-Click and Citation

The reason this topic exists at all is a measurable change in how people get answers. A large and growing share of searches now end without anyone clicking through to a website, because the answer is shown on the results page itself. Across 2026, credible studies put the proportion of Google searches that end without a click at somewhere around two-thirds, up from roughly half a few years earlier, and the figure climbs higher still on the queries where an AI Overview appears. Exact numbers vary a lot between studies, because they measure different query sets in different ways, so treat any single headline percentage with caution. The direction, though, is not in doubt: the click is becoming scarcer, and the answer box is becoming the destination.

This reframes the goal. For informational content especially, you are no longer only competing to rank and win the click. You are competing to be the source the answer is built from and names. There is a genuine silver lining worth holding onto: the visitors who do still click after reading an AI answer tend to arrive better-informed and more ready to act, so the traffic you keep is often higher-intent than before. The honest cost is that raw visit counts to informational pages have fallen for many sites, and no amount of GEO fully reverses that. The realistic aim is to protect your share of a changed pie, not to pretend the pie is the same.

Honest Ground Rules: What Is Established and What Is Speculation

This area attracts a great deal of confident advice, much of it selling something. It is worth fixing a few reference points before spending any effort.

• Established. AI Overviews and AI Mode are generated from Google's ordinary search index, so classic search fundamentals (crawlable, fast, well-structured, genuinely useful pages) still decide whether you are eligible to be cited. Clear, well-organised, quotable content is easier for a model to lift accurately. Structured data helps machines read your pages without guessing. Being mentioned favourably on other trusted sites raises the odds an assistant repeats you. None of this is exotic; it is good publishing.
• Plausible but unproven. Small formatting habits (leading with a direct answer, using question-shaped headings) probably help, and cost little, but the evidence is soft and the platforms do not confirm the mechanics.
• Speculative or oversold. The idea that adding an llms.txt file measurably improves your AI citations is, as of mid-2026, not supported by the available evidence. Any tool promising to "guarantee" placement in AI answers is selling certainty that does not exist. The systems are opaque and change without notice.

Read the rest of this guide with that ranking in mind: spend most of your effort on the established fundamentals, adopt the cheap plausible habits, and treat the speculative items as low-cost experiments rather than priorities.

llms.txt: What It Is, How to Add One, and What to Honestly Expect

What the File Is

The llms.txt proposal, published by Jeremy Howard in September 2024, suggests a single Markdown file placed at your site's root, at https://yourdomain.com/llms.txt, that hands a language model a clean, curated map of your most important content. The reasoning is that a model's context window is too small to swallow a whole website, so a short, hand-picked index of links (with a one-line note on each) lets it find the good parts quickly. It is, loosely, a sitemap written for machines that read prose rather than XML. The format is deliberately simple:

• A single H1 line with the site or project name (the only strictly required part).
• A blockquote giving a one-sentence summary of what the site is.
• Optional free-form Markdown describing the site in more detail.
• One or more H2 sections, each a list of Markdown links to key pages, with an optional short note after each link.
• An optional section literally named Optional, whose links a model may skip when it needs to keep things short.

A companion convention, llms-full.txt, goes further by concatenating the full Markdown text of your key pages into one large file, so a model can ingest everything in a single fetch. Here is what a modest business site's llms.txt might look like:

# Example Accounting

> Example Accounting is a Cape Town accountancy firm offering tax, payroll and advisory services to small businesses across South Africa.

## Core pages

- [Services](https://yourdomain.com/services): the tax, payroll and advisory services offered
- [About](https://yourdomain.com/about): company background, team and professional credentials
- [Contact](https://yourdomain.com/contact): office address, phone number and enquiry form

## Guides

- [Small business tax guide](https://yourdomain.com/guides/small-business-tax): plain-language guide to provisional tax
- [Payroll basics](https://yourdomain.com/guides/payroll): how PAYE and UIF work in practice

## Optional

- [Full article archive](https://yourdomain.com/blog): every article published on the site

Replace yourdomain.com and the example content with your own. Keep the list short and honest: the point is to highlight your best, most representative pages, not to dump your whole sitemap.

The Honest Status of llms.txt in 2026

This is where realism matters. Despite eighteen months of discussion, llms.txt sits on only around one in ten sites, and adoption is not climbing quickly. More importantly, it is a community convention, not a ratified standard: no standards body backs it, and, as of mid-2026, no major AI provider has publicly committed to reading it in their production search systems. Google has said plainly that it does not use it, likening it to the long-abandoned keywords meta tag, and log studies of hundreds of millions of AI-crawler visits find that the retrieval bots overwhelmingly ignore the file and simply read your normal HTML instead. Independent analyses of large domain samples have found no measurable lift in AI citations from having the file.

So why mention it at all? Because it does have one genuine, proven use: developer-facing documentation. AI coding assistants and the IDE tools built on them (and setups using the Model Context Protocol) do consume llms.txt to work with a product's docs more accurately. If your site is documentation for software, or an API, an llms.txt is worth having on those grounds alone. For an ordinary business or content site hoping for more AI search citations, set your expectations to near zero and treat it as tidy housekeeping, not an SEO tactic.

How to Add llms.txt to a WordPress Site

WordPress does not create this file for you, and there is nothing about it in core. You have two straightforward routes.

1. Upload a static file (simplest and most reliable). Write your llms.txt in any plain-text editor and upload it to your site's web root, alongside wp-config.php. Because the web server serves any real file that exists before it hands the request to WordPress, a physical llms.txt at the root is delivered directly and correctly, exactly as robots.txt works. On Noiz hosting you upload it over SFTP or through the panel File Manager. Managed-plan clients can send the file to the Noiz support team to place for them.
2. Use a plugin. Several SEO and dedicated plugins now generate and maintain an llms.txt for you, keeping it in step as you add pages. This trades a little control for convenience, and it means one more plugin to keep updated. If you go this way and also upload a physical file, the physical file wins, so pick one method.

Whichever route you take, verify it by visiting https://yourdomain.com/llms.txt in your browser. You should see your plain Markdown text, not a styled WordPress page and not a 404. If you get a themed "page not found", the file is not at the true web root, or a plugin is intercepting the path.

Generative Engine Optimization: The Practices That Genuinely Help

If llms.txt is the overhyped part, this is the part that actually earns citations, and reassuringly little of it is new. GEO is mostly good writing and good structure, aimed at being easy for a machine to find, parse and quote accurately.

Lead With the Answer, Then Explain

Models building an answer look for a passage they can lift that directly settles the question. Put a clear, self-contained answer in the first sentence or two under a heading, then expand beneath it. A paragraph that opens with "Provisional tax is paid twice a year, in August and February" is far more quotable than one that warms up for four sentences before reaching the point. In the WordPress block editor this is simply a matter of habit: state the conclusion first in each section, then support it.

Write in Question-Shaped Sections

People ask assistants full questions, so headings phrased as the questions your readers actually type give a model an obvious match between the query and your content. Use the heading blocks (H2 and H3) to structure a page around real questions, keep one idea per section, and let the answer sit immediately below its heading. This also happens to make the page clearer for human readers, which is the point.

Be Quotable: Specifics, Statistics and Named Sources

Vague, hedged prose is hard to cite; concrete claims are easy. Where you can, include specific figures, dates, named sources and direct statements of fact, and attribute anything you have borrowed. Original material a model cannot get elsewhere (your own data, a genuine case study, first-hand experience, clear definitions) is disproportionately valuable, because it gives the assistant something to cite that is uniquely yours. Confident, plain writing beats padded, keyword-stuffed writing here.

Build Entity and Topic Authority, and Keep Your Naming Consistent

These systems reason about "entities", meaning the people, organisations, products and concepts your site is about. Help them by being consistent: refer to your business, products and key terms the same way every time, rather than drifting between synonyms. Cover a topic thoroughly across several linked pages rather than in one thin post, and use internal links so a crawler can see how your pages relate. A clear About page, consistent contact details, and structured data (below) all reinforce who you are and what you are authoritative about.

Earn Mentions Where the Engines Already Look

This is the uncomfortable but important one, and it happens off your own site. Assistants lean heavily on sources they already trust: established publications, reputable directories, community discussions and review sites. Being mentioned, accurately and favourably, on those external sources raises the chance an assistant repeats what they say about you. That means the unglamorous work of real public relations, genuine listings, honest reviews and being part of your field's conversation matters more for AI visibility than any on-page trick. There is no file you can upload that substitutes for being talked about elsewhere.

Keep It Fresh

Retrieval systems favour current information, and a page last touched three years ago competes poorly with one updated this quarter. Revisit your important pages periodically, correct anything stale, and let genuine updates show. Freshness is not a licence to churn out filler; it is a reason to maintain the pages that matter.

Structured Data: Helping Machines Read Your Pages Without Guessing

Structured data is a block of machine-readable code, added invisibly to a page, that spells out what the page is about in the shared Schema.org vocabulary: this is an Article with this author and date, this is an Organization with this address, this is a Product with this price. It removes guesswork, letting search and AI systems understand your content precisely rather than inferring it from the words. The modern, preferred way to add it is JSON-LD, a small script in the page's head, and you almost never write it by hand.

On WordPress, most structured data comes from your theme and plugins rather than core. WordPress core outputs some basic markup, but the rich, connected schema that describes your organisation, articles, breadcrumbs and products is typically added by an SEO plugin. Several capable options exist, free and paid, and they generate the JSON-LD automatically from content you have already entered. The categories most worth having in place are:

• Organization or LocalBusiness, describing who you are, which anchors your entity.
• Article or BlogPosting on your content, carrying author, headline and dates.
• BreadcrumbList, describing where a page sits in your site.
• Product and offer details for shops, and FAQ on genuine question-and-answer pages.

One honest caveat so you calibrate your effort: since 2023 Google has stopped showing the eye-catching FAQ and how-to rich results in ordinary listings for most sites, so structured data is no longer the guaranteed route to a fancier search snippet it once was. Its value now is quieter and more durable: it helps machines, including AI systems, understand your pages accurately, and it is a low-cost foundation rather than a magic switch. After setting it up, check a few pages with a structured-data validator (Google's Rich Results Test or the Schema.org validator) to confirm the markup is valid and matches the visible page, since invalid or misleading markup does more harm than none.

Controlling Which AI Crawlers May Read Your Site

You get a say in whether AI systems may use your content, exercised mainly through the robots.txt file at your site's root. This is a decision with a real trade-off, and one widespread misconception that is worth clearing up before you touch anything.

Recall the two kinds of AI crawler. Training crawlers (such as GPTBot, ClaudeBot, Google-Extended and the Common Crawl bot CCBot) gather data to teach future models. Retrieval crawlers (such as OAI-SearchBot, ChatGPT-User, PerplexityBot and Perplexity-User) fetch pages to answer questions live, and these are the ones that produce citations back to you. Blocking a training crawler protects your work from being absorbed into a model; blocking a retrieval crawler removes you from that assistant's answers, references and all. That is the core tension: the same openness that lets an assistant use your content without a click is what lets it cite and link you.

The misconception concerns Google, and it catches people out. Google's AI Overviews and AI Mode are built from Google's ordinary search index, gathered by Googlebot. The separate Google-Extended token controls only whether your content trains Google's Gemini models; it does not control whether you appear in AI Overviews. So blocking Google-Extended will not take you out of AI Overviews, and the only way to leave AI Overviews is to block Googlebot, which would also remove you from Google Search entirely. For nearly every site that is far too high a price, so most owners keep Googlebot fully allowed and accept that AI Overviews come with ordinary search visibility.

A common middle path is to allow the retrieval and search crawlers, so you remain eligible to be cited in AI answers, while blocking the pure training crawlers if you would rather your content not feed future models. A minimal robots.txt in that spirit looks like this:

User-agent: GPTBot
Disallow: /

User-agent: CCBot
Disallow: /

User-agent: Google-Extended
Disallow: /

User-agent: *
Allow: /

That example blocks three training crawlers and leaves everything else, including search and retrieval bots, free to read the site. Adjust it to your own choice; there is no universally right answer, and reasonable owners land in different places. On WordPress you can edit robots.txt either by uploading a physical file to the web root over SFTP or the File Manager, or through the robots.txt editor most SEO plugins provide. Note that robots.txt is a request, not a lock: reputable crawlers honour it, but it is not a security control, so never rely on it to hide anything sensitive. Genuinely private material belongs behind a login, as covered in the Noiz WordPress Security Checklist.

How to Tell Whether Any of This Is Working

Honesty requires admitting that measuring AI-search success is genuinely hard, because the assistants rarely tell you when they used you. A few imperfect methods together give a usable picture.

• Watch your referral traffic. When someone does click through from an AI tool, it usually shows in your analytics as a referral from a domain such as chatgpt.com, perplexity.ai, gemini.google.com or a Copilot address. Tracking that segment over time is the closest thing to a direct signal. If you have not yet set up analytics, the Noiz guide to setting up Google Analytics 4 is the place to start.
• Spot-check by asking. Periodically ask the main assistants the questions your content answers, and see whether they cite you, a competitor, or no one. It is manual and unscientific, but it tells you plainly where you stand on the queries you care about.
• Track branded and comparison queries. Growth in people searching for your name, or asking assistants to compare you with rivals, is a sign your entity is registering. Ordinary search-console data still helps here.

Expect noise and incomplete attribution, and judge trends over months rather than reacting to any single day. Anyone promising you precise, real-time measurement of AI citations is overstating what the platforms currently expose.

Troubleshooting

• Symptom: visiting yourdomain.com/llms.txt shows a styled "page not found" instead of your text. The file is not at the true web root, so WordPress is handling the request. Confirm you uploaded it into the same folder as wp-config.php (commonly public_html or httpdocs), not into a subfolder or the theme.
• Symptom: your llms.txt is live but nothing about your AI visibility changed. That is the expected outcome in 2026, not a fault. The file is not a citation booster for general sites; put your effort into the GEO fundamentals and structured data instead.
• Symptom: you blocked Google-Extended to escape AI Overviews, but you still appear in them. This is by design. Google-Extended governs only Gemini model training, not AI Overviews, which are drawn from the normal index. Leaving them would require blocking Googlebot, which also removes you from Google Search, so most sites should not.
• Symptom: your structured-data validator reports errors, or "markup does not match visible content". The schema is describing something the page does not actually show, which search systems penalise rather than reward. Fix the plugin settings so the markup reflects the real, visible page, and never mark up content that is not on the page.
• Symptom: you see no AI referral traffic at all. Referrals from AI tools are genuinely low in volume because most AI use ends without a click, so a small or empty number can be normal. Confirm your analytics is installed correctly first, then judge the trend over a longer period rather than a single week.
• Symptom: assistants describe your business inaccurately. They are likely repeating stale or wrong information from elsewhere on the web. Correct your own site's facts and structured data, keep your naming consistent, and work on getting accurate mentions on the external sources the assistants trust, since that is where the wrong information usually originates.

AI search is one of the least settled areas in web publishing right now, and it is easy to waste effort on tactics that sound impressive and change nothing. If you would like help separating the worthwhile from the hype for your specific site, placing an llms.txt or robots rules on Noiz hosting, or setting up structured data and analytics so you can actually see what is happening, open a support ticket with the Noiz support team. Include your domain and a note of what you are trying to achieve, and a technician can help you focus on the parts that genuinely move the needle.