LLM SEO for WordPress sites

What LLM SEO means

LLM SEO is optimising so a large language model can find your pages, read them, and quote them without getting you wrong. It is not a different discipline bolted onto SEO so much as a different success condition: you are no longer competing for a position on a results page, you are competing to be the sentence inside somebody's answer.

A page can be crawled, indexed, ranking respectably, and still never cited. That gap is the whole subject.

Two ways a model can know about you

The first is training. Your text was in the data a model learned from, so it may describe you without looking anything up. You cannot schedule this, cannot verify it, and cannot correct it — if a model learned something wrong about you in training, you wait for the next one.

The second is retrieval. The assistant runs a search, fetches pages, reads part of each and writes an answer citing them. This happens live, responds to changes within days, and is the path you can actually influence.

Almost everything worth doing targets retrieval. Anyone promising to get you "into the training data" is selling you something they do not control.

Writing for extraction, concretely

"Write for extraction" is vague advice, so here is what it actually means when you are editing a post in WordPress.

• Use headings as structure, not styling. H2 and H3 are chunk boundaries now. Skipping from H2 to H4 because it looked better fragments the page in ways that show up in what gets extracted.
• One idea per section. A section covering three things gets chunked mid-thought and none of the three survives intact.
• Front-load each section too, not just the article. The first sentence under a heading is doing the most work.
• Expand pronouns near section starts. "It" is unresolvable when the previous paragraph is not included.
• Put recurring questions in a real FAQ, with the answer directly under the question and both visible on the page. Question-and-answer pairs are close to ideal retrieval shape because they match how people prompt.
• Write the excerpt on pages that matter. An auto-generated one truncated mid-sentence is sometimes the thing that gets read.

Schema that agrees with the page

Structured data does not make a model rank you higher, as far as anyone outside these companies can tell. What it does is remove ambiguity about the price, the author, the date and the organisation — facts a model would otherwise infer from prose, sometimes wrongly.

Four types cover almost everything: Organization for who you are, Article for anything written, FAQPage where questions are genuinely visible on the page, and Product or Offer if you sell something and want the number stated correctly.

The rule that matters most is that schema must describe what is actually on the page. Markup contradicting your visible text teaches a model that your structured data is unreliable, and it has no way to tell which half to trust. Most SEO plugins emit Organization and Article already, so check your page source before adding a second copy — two Organization blocks give a parser two competing answers.

Being one identifiable thing

Ask an assistant "what is [your company]?" in a fresh conversation. If the answer is vague, hedged, or describes someone else with a similar name, no amount of keyword work will fix it — the problem is that you have not resolved into a single thing.

Three fixes, in order of value. Write one canonical sentence describing yourself and use it everywhere, word for word. Pick one canonical name and stop using the abbreviation, the product name and the domain interchangeably without connecting them. Then corroborate: an entity is confirmed by independent sources agreeing, so a plugin directory listing, a repository or a well-kept profile is worth more than another page on your own site saying the same thing.

Where a third party has you wrong, fixing that usually beats writing something new. A stale directory entry with your old pricing will be read and quoted.

Measuring, honestly

Nobody has good measurement here yet, and it is worth saying so plainly rather than pretending a dashboard solves it.

• Ask the assistants directly, monthly. Same questions, answers recorded. Watch three things: whether the description is yours, whether the facts are current, and whether you get confused with anyone.
• Referral traffic from assistant domains. Real, and a large undercount, because the answer usually satisfies the person without a click.
• Crawler hits in your access log. Proof of reachability only. A bot you believed you allowed not appearing at all is a robots.txt problem, not a content problem.

Anyone selling a definitive AI visibility score is selling an estimate. Treat it as one.

Where the plugin fits

Generated drafts arrive with real heading structure and sections written to survive being read on their own, which is the shape retrieval wants. The fact-checking pass flags claims it could not verify rather than rewriting them into confident prose — the failure mode that turns a citation into a liability.

The rest of this page is site work rather than plugin work, and we would rather say that than imply otherwise.

Further reading: what LLM SEO is, how ChatGPT chooses sources, entity SEO and llms.txt.

Questions