Get your WordPress site cited by ChatGPT: find the gap, fix the post, re-check

ChatGPT's search side relies on crawlers it runs itself. ThePlusAddons describes three OpenAI bots: OAI-SearchBot builds the search index, ChatGPT-User fetches a page when a question needs live data, and GPTBot crawls for training. You can allow the first two and block GPTBot, so a training opt-out does not have to cost you citations.

Windsor.ai joined 30 days of server logs with 90 days of Search Console data across 238 pages. Claude fetched every page with 1,000 or more Google clicks, but only 19% of pages with 1 to 9 clicks. ChatGPT showed the same pattern at larger volume, 6,415 citation fetches against Claude's 860. Their first finding is that pages that get cited are mostly pages that already rank.

A skipped post therefore has a short list of causes: it is blocked from the search crawlers, it does not rank, or it ranks but does not open with the answer a question needs.

The pages that rank for this query agree on a checklist: allow AI crawlers in robots.txt, add FAQ and Article schema, generate an llms.txt file, use an SEO plugin such as Rank Math to fill in schema, and write server-rendered, extractable answers. These are sound hygiene items and worth doing once for the whole site.

What none of them gives you is the list of questions where an answer currently names a competitor. Without it, you apply the checklist to every post, including posts that no buyer question points to. Windsor.ai's advice to open with the sentence you want quoted only pays off on the pages that match a question you actually lose.

That is the part Linkeddit supplies. answer_gaps_get returns the open gaps ranked by intent, evidence and confidence, and each gap carries the observations, the cited sources and the actions kept separate. Claude can then choose the three posts worth editing this week.

WordPress.com's tools reference lists posts.get, posts.update and block-level operations. post-sections.list returns a post's top-level blocks with an index and a block hash, and post-sections.replace swaps one block while leaving the rest untouched. A change to the opening block, which is the one a quotable answer lives in, does not need a rewrite of the post.

Every write goes through a confirmation step: the vendor says the agent must describe what it plans to do, ask for explicit approval and pass user_confirmed as true. It also warns that posts.update with a status of publish goes live immediately. Keep Claude on show-before-after, one post at a time.

Keep the URL and slug as they are. A backlink points at an address, and Unstale notes that WordPress adds a redirect for a published post's old slug but not for pages or drafts.

Re-run the same questions in Linkeddit after the edit and compare presence and recommendation rates across repeated samples. On Linkeddit's test project, ChatGPT named the brand in 16 of 18 admissible samples, while Perplexity returned no usable samples because all 21 errored. Linkeddit reports that as unmeasured, not as zero, and you should read an unmeasured engine the same way.

Measurement runs through official provider APIs with grounded web search, so it is a controlled proxy and not an exact copy of what a consumer sees. Windsor.ai also cautions that a fetch by a user-agent bot is a potential citation, not proof the answer used the page.

This connector is WordPress.com's. The vendor documentation says Jetpack-connected self-hosted sites use the same server and tool catalog, and does not describe a route for a self-hosted site with no Jetpack connection.