Being Cited by AI Search: What Actually Works from WordPress

Most AI answers that cite sources are built in two steps. The system retrieves candidate pages from a search index or a live fetch, then a language model writes an answer from those pages and links some of them. Google's AI features draw on Google's own index. ChatGPT search uses OpenAI's crawler, OAI-SearchBot, together with third-party search providers. Copilot is built on Bing. Perplexity runs its own crawler and fetches pages on demand.

The practical consequence is that the first gate is the same as for classic search: the page must be crawlable, indexable and readable without running a large JavaScript bundle. Google states that there is nothing special to do for its AI features beyond being indexed and eligible to appear with a snippet. If a page never reaches the index, no amount of structure will get it quoted.

This is also where many sites fail without knowing it. Hosting firewalls and CDN bot settings increasingly block crawlers by user agent. I have seen a host return 403 to Bingbot and 429 to GPTBot while every browser, and the site's own monitoring, received a normal 200. Because Copilot and other assistants depend on Bing's index, a single CDN setting like that can remove a site from a large share of AI answers.

• Request key pages with each crawler's user agent and confirm a 200 response
• Use Bing Webmaster Tools and Google Search Console URL inspection to confirm what the real crawlers see
• Make sure the main content is present in the HTML, not only after client-side rendering
• Submit changed URLs through IndexNow so Bing hears about them quickly

llms.txt is a proposed convention: a Markdown file at the site root that summarises the site and links to its most useful pages, written for language models rather than for people. It is cheap to produce, and on a WordPress site it can be generated from the same list of pages as the XML sitemap so it never drifts out of date.

What it is not, at the time of writing, is a documented input for the major answer engines. None of the large providers has published that it reads llms.txt to decide what to cite. Treat it as a low-cost experiment that might help some tools understand your site, not as a ranking factor, and do not let it replace the work that is documented: crawl access, indexation and clear pages.

Several providers now separate crawlers by purpose. OpenAI documents GPTBot for collecting training data and OAI-SearchBot for search, plus ChatGPT-User for fetches a user triggers. Google uses the Google-Extended token to control whether content is used for training Gemini models, separately from Googlebot, which still governs Search and its AI features. Anthropic and Perplexity publish similar distinctions.

That makes the decision more precise than "block AI or not". A business that wants to be recommended in answers should allow the search and user-triggered crawlers. Whether to allow training crawlers is a separate judgement about how your content may be reused. Blocking the search crawlers removes you from the answers that might have sent visitors; blocking only training crawlers has no documented effect on being cited in search features.

Whatever you decide, write it in robots.txt and check that the hosting layer agrees. A robots.txt that allows a crawler means nothing if the CDN in front of it answers with an error.

When a model builds an answer from several pages, it favours passages that answer the question directly and can stand on their own. A page that buries its answer under a long preamble is easy to summarise and hard to quote. Put a plain answer to the main question in the first paragraph or two, then expand on it.

Use headings that say what the section answers, keep each section focused on one question and write specific claims: a price range, a time estimate, a named setting, a measured result. Tables are useful because they keep facts aligned and are easy to extract. FAQs that answer real questions in two or three sentences work well for the same reason. Vague marketing copy gives a model nothing to cite.

Keep the facts consistent across the site. If your services page says one thing about pricing and your FAQ says another, an assistant may present either, or neither. One clear statement, repeated accurately where it is relevant, is safer than several approximations.

Answer engines are cautious about recommending a business or a person they cannot identify with confidence. A name shared by many people, a business with an inconsistent description, or a site that nothing else links to is a weak candidate, however good its pages are.

Make the identity unambiguous: the same name, role, location and one-line description on your site, in Person or Organization structured data with sameAs links, and on the external profiles you control. Then earn mentions you do not control, such as client case studies, directory listings, conference pages, genuine reviews and articles on other sites. The E-E-A-T and author pages guide explains how to set up the on-site part on WordPress.

Some answer engines pass a referrer when a user clicks a citation, so visits arrive from domains such as chatgpt.com, perplexity.ai, copilot.microsoft.com and gemini.google.com. In GA4, build an exploration or a custom channel group that collects these referrers so you can see the trend rather than individual visits. Clicks from Google's AI features are counted as Google organic search; Search Console has begun adding reports that show impressions in AI Overviews and AI Mode, so check whether your property has them.

Complement analytics with a simple manual test. Keep a list of eight to ten questions a prospective client might ask, such as who builds a particular kind of website in your city, and ask them in each assistant once a month. Record whether you are mentioned, whether the facts are correct and which page is cited. It is not statistically rigorous, but it shows whether your changes are working.

Skip tools that promise to "optimise for AI" by rewriting pages into keyword-dense summaries, hidden text aimed at language models or special markup that no provider documents. Prompt-injection tricks hidden in pages are more likely to harm trust than help. Buying mentions or publishing fabricated reviews is risky for the same reasons it is risky in classic search.

The durable work is unglamorous: let the right crawlers in, make pages fast and readable without JavaScript, answer questions clearly and make your identity easy to confirm. The technical WordPress SEO guide, the XML sitemaps and robots guide and the schema markup guide cover the foundations, and the SEO services page explains how I review a site for this.