How an answer engine reads a site
By Ali Abedi ·
An answer engine, the AI assistant a prospect now asks before they search, reads a website differently from a search engine: it wants the answer in the first sixty words, questions as headings, structured data that names who you are, and a file that says how to describe you. We audited our own site this week. Here is what it read, and what it missed.
What is the difference between a search engine and an answer engine?
Search engines rank pages. Answer engines quote them. When someone asks an assistant which studio in West Palm Beach builds on a design system, the assistant does not send them to ten links; it reads the pages it can reach and composes one answer, naming its sources. That changes what a page has to do. It has to say the answer plainly, early, in words that survive being lifted out of context.
What does an answer engine actually read?
Four things carried the weight in our own audit. The opening paragraph on every page states what the studio is in the first sixty words, so a crawler that stops early still leaves with the answer. Headings on the blog are questions, because an engine matches a question to a heading before it reads the paragraph under it. Structured data, the JSON an engine reads beside the words, names the organisation, the founder, the services, and on each post the article and its questions and answers. And a plain text file at llms.txt tells an assistant what the studio is, what it does not do, and what not to guess on our behalf: no prices, no client revenue figures.
What did our own audit find?
What the audit found missing was smaller and just as real. The sitemap said when nothing: no last-modified dates, so an engine could not tell a fresh post from an old one. The posts declared themselves web pages rather than articles, with no publication date in the share metadata. The blog index published no schema at all, though the builder for it had been written. And every page shared one share image, so a post pasted into a chat showed the studio's tagline instead of the post's title. Each was a morning's work once named.
The one we could not settle by reading is worth writing down. Our animated headings render every word twice in the HTML: once as the pieces that animate, hidden from screen readers, and once as an intact copy for them. A text extractor that ignores the hiding reads the heading doubled. Whether search engines and assistants keep both is unverified, so it is a measurement to make, not a fix to ship.
What should a business do about it?
The habit that comes out of this is simple. Once a quarter, crawl your own site the way an engine would and read what comes back: the title, the description, the first sixty words, the headings, the structured data, the dates. If a stranger could answer what this company does from that alone, the page is doing its job.
Questions
- What is llms.txt?
- A plain text file at the site's root, like robots.txt, that tells an AI assistant what the site is and how to describe it. Ours also says what not to guess: no prices and no client figures.
- Does structured data matter for AI answers?
- Yes. It is the one place a page names its organisation, author, dates and questions in a form a machine reads without interpreting prose.
- Do we need to rewrite our pages?
- Usually not. Most of the work is the first paragraph, question headings, dates that mean something, and the files an engine reads beside the page.
Ali Abedi is the founder of Magenta Web, a brand identity, web design, and digital marketing studio in West Palm Beach, Florida. He has designed and built websites for architecture, hospitality, legal and technology clients for more than ten years.
Read next
- One decision, a hundred screens · the next note
- Zafiro · the newest case study
- What each engagement includes · services