GEO for WordPress: how AI search finds your site

This page is the overview. The same topics are covered as eight common mistakes for diagnosing an existing site, and llms.txt has a separate guide.

All five parts can be implemented on a standard WordPress installation without a new platform or a subscription, and each part can be checked afterwards with the steps on this page.

How generative search differs

A search engine ranks pages and sends a visitor to one of them. A generative engine reads several sources, writes an answer, and selects which sentences to use and which sources to credit. The relevant questions are whether a page contains a sentence suitable for quoting, and whether its source can be identified.

This places weight on three factors that ranking reports do not show: whether an answer stands on its own, whether the facts on the page also exist in machine-readable form, and whether the crawlers that cite sources are allowed to access the site.

Part 1: Self-contained answers

Place the answer near the top of the page in two or three self-contained sentences that remain clear when the surrounding text is removed. Avoid references such as “as mentioned above” and answers that depend on several paragraphs of preceding context.

• One page, one question. A page that answers several questions has no single clear answer to extract.
• Put numbers, prices, durations and limits in the text, not in images or on pages several clicks away.
• Phrase headings as the question each section answers, in the form a user would ask it.

Check: hide the heading and read the first two sentences alone. They should answer the page title.

Part 2: Structured data that matches the page

Structured data is the machine-readable copy of what a visitor sees: a product and its price, an article and its date, questions and their answers. Missing or incomplete markup causes no visible change on the page, so it can remain unnoticed for months.

• Product, Article, FAQPage, Organization and BreadcrumbList cover most of what a business site needs. Adding more types does not improve the result.
• Include every required property with accurate values. A missing required property disables the rich result without any visible error.
• The markup must match the page. A price in the markup that differs from the price on screen is worse than no markup.

Check: run the page through a structured data test and review the warnings as well as the errors. On AumCreate’s own product pages, a missing image property kept the rich result disabled for three weeks while every page displayed correctly.

Part 3: Consistent organization information

To answer a question such as “who makes X”, a model assembles an entity from the sources it can find. If the name, description, location and links differ between the site, wordpress.org, GitHub and LinkedIn, each version reduces confidence in the others.

• Define Organization once on the site and reuse it, rather than writing it per page.
• Use the same short description on every profile. Copy it rather than rewriting it.
• Link to the profiles the organization controls so the sources can be connected.

Check: search for the brand name and read the top five results. They should describe one organization consistently; if they do not, address this before the other parts.

Part 4: llms.txt as an index

llms.txt is a plain-text list at the site root that points to the main pages. It is a proposed format, support varies between AI providers, and as an index it does not change the content of the pages it lists. Generate it from published content rather than writing it by hand, and make sure only one source on the site generates it.

Which plugins provide it, which takes precedence when two are active, and how to confirm what is served are covered in the llms.txt guide. The AumCreate file is published at /llms.txt.

Part 5: Crawler access policy

AI crawlers fall into different groups. Some fetch pages so that answers can cite and link them: OAI-SearchBot, PerplexityBot, Claude-SearchBot. Others collect content and send no traffic back: GPTBot, ClaudeBot, CCBot, Bytespider. Blocking every crawler with an AI-related name also removes referrals from the first group. Google-Extended is not a crawler: it is a robots.txt token that controls whether content Google has already crawled may be used for Gemini training and grounding, and it does not affect Google Search.

• Review which crawlers actually visited before changing any rules.
• Each control has a different strength: turning a crawler away at the server blocks the request; a robots.txt rule is a request that compliant operators honour and others ignore; the noai header has limited support.
• Do not block Googlebot or Bingbot, and keep the crawlers that generate link previews allowed.

Checklist

The following five checks take about an hour. Each has a clear pass or fail result, and none requires a paid tool.

• Open three pages and read only the first two sentences. Check that they answer the title.
• Run the same three pages through a structured data test. Resolve all errors and review the warnings.
• Search for the brand name. Check that the top results describe one organization.
• Open /llms.txt on the site. Check that it exists and that three random entries lead to relevant pages.
• Review a month of crawler activity. Check that citing crawlers can access the site and that blocked crawlers are no longer arriving.

Scope

These measures do not guarantee that an AI system will cite a site. They address the common reasons a page is passed over: an answer placed too deep to quote, facts available only in human-readable form, inconsistent organization information, and blocked access for crawlers that cite sources.

Each part can be implemented by the site owner using this page. AumCreate also provides a fixed-price setup service: $199, three working days, checklist included.