Guide 13 · Evidence review
How to write content for AI search
The same way you write for people. Google says you do not need to write in a special way for its AI features, because they understand synonyms and general meaning. Our data adds a reason not to try: even a one-character change to a question changes a fifth of the sources an engine reads, so there is no single phrasing to write for.
What is writing for AI supposed to mean?
Most advice under this heading asks for some mix of the following: repeat the exact question as a heading, answer in a fixed number of words, add FAQ blocks to every page, use “entities” and definitions heavily, and match the wording people type into chatbots. Some of it is ordinary clear writing under a new name. Some of it is keyword-matching carried over from old SEO.
What does Google say?
Google’s guide to optimizing for generative AI features: “You don’t need to write in a specific way just for generative AI search.” Its reason is that AI systems understand synonyms and the general meaning of what someone is looking for.
What the guide does ask for is content that is not commodity. Its example of commodity content is a generic tips list. Its example of the opposite is a first-hand review. So the thing that decides whether you are worth citing is what the page contains, not how it is phrased.
Why is matching one phrasing a losing game?
Because the engines do not return a stable answer to a phrasing. In Run 04 we asked Perplexity the same questions repeatedly. An identical query string returned an identical source list, 100% overlap, because the answer was cached. Adding a single character dropped the overlap to 79%. Asking a different engine dropped it to 12%.
A buyer’s question arrives in dozens of phrasings. When we rebuilt the 24-prompt set for this site from Google autocomplete, the variants of one question were not word-for-word matches of each other. Writing a page to mirror one of them optimises for the least stable thing in the whole system.
| Change to the question | Perplexity sources still the same | Effect |
|---|---|---|
| None, asked again | 100% | Cached |
| One character added | 79% | Partly new |
| Same question, other engine | 12% | Mostly new |
What does get cited, then?
In our leaderboard run, what the engines read for a plugin category was vendor articles, YouTube (on Google), Reddit (on Google, 7 of 15 answers) and a small number of comparison sites. The common thread is pages that say something specific about the category. Across 45 answers, install count barely predicted whether a plugin was named: All-in-One WP Migration has over five million installs and was named in 2 of 9 backup answers.
So the useful writing advice is the plain kind:
Answer the question the buyer asked, early, in words they would use. Add something they cannot get elsewhere: your own numbers, a test, a comparison you ran. Say where it stops being true, which a reader trusts and a generic page never does. That is what Google’s guide means by non-commodity content, and it is how the pages on this site are written.
What would change my mind?
Our query-variation data is Perplexity on WordPress plugin questions, measured in September 2026. It shows how sensitive sources are to phrasing; it does not test page wording directly.
A test where the same facts are published in two styles, one written in the “AI-optimised” pattern and one plainly, with citation rates compared across engines and several phrasings, would settle it. If the optimised style won consistently, I would publish that result here.
How this was made: figures come from our own runs in the engines’ public interfaces or from the studies named above. Drafted with an AI assistant (Claude) and reviewed by Alauddin Chowdhury. Method and AI use.
Want to see which questions your buyers actually ask?
Send me your wordpress.org slug. I will pull the real phrasings for your category from autocomplete and run them on all three engines. Free, no call, usually within a day.