We Rebuilt This Site for AI Search. Here Is Everything We Changed

This is a first-hand account. We recently audited and rebuilt this website's entire search setup for the era of AI answers, and this post documents what we actually changed and the evidence behind each call. Copy freely.

First, the reality we optimised for

Search split in two. Classic Google rankings still matter, but a growing share of questions now get answered directly by ChatGPT, Perplexity, and Google's AI Overviews, which cite a handful of sources instead of listing ten links. One widely-cited analyst projection put traditional search volume down roughly 25% this year. Optimising to be cited is called GEO (generative engine optimisation), and the good news is that it is mostly excellent SEO with sharper structure.

What we changed, in order of impact

1. We removed markup that could have gotten us penalised

Our course pages carried aggregate star ratings in their structured data from placeholder content. Google's guidelines are explicit that ratings not backed by real visible reviews count as spammy markup, and the penalty can strip rich results from an entire site. We deleted it the day we found it. If your site was scaffolded quickly or generated with AI tools, audit your schema before Google does; this failure mode is common in generated sites and almost nobody checks for it.

2. We made every page state its identity

Canonical URLs on all pages (there were zero), unique titles and descriptions, one consistent domain (we discovered our host served www while every tag claimed the bare domain, so crawlers got contradictory signals), and honest lastModified dates in the sitemap instead of dates from before the site existed.

3. We added the schema AI engines actually read

Organization and WebSite markup site-wide, Course markup with real properties (syllabus sections, workload, pricing), ItemList on the catalogue, BreadcrumbList everywhere, FAQPage on our FAQ. Structured data is how machines confirm what your page is. Every claim in our markup is verifiable on the visible page, which is the rule that keeps schema safe.

4. We restructured content to be quotable

AI answers lift self-contained passages. So: headings phrased as the questions people actually ask, a crisp definition paragraph near the top of key pages that survives being quoted out of context, and specific facts attributed to sources. While doing this we found our course listing page had silently become client-side rendered, meaning its content was invisible in the initial HTML. If your site is built on a modern JavaScript framework, view source on your money pages; if your content is not in the HTML, neither is your ranking.

What we deliberately did not do

We skipped llms.txt, the proposed standard file for AI crawlers. The evidence convinced us: a study of 300,000 domains found only around 10% adoption, monitoring of hundreds of millions of AI bot visits found the file directly requested only a few hundred times, and Google has said on record it does not use it. Costs little, does roughly nothing. We would rather spend the effort on content structure, which every engine demonstrably reads.

The honest caveat

None of this ranks a site by itself. Technical correctness opens the door; content that deserves citations walks through it. That is the slower project we are working on now, this post included.

If you want the systematic version of this playbook, it is what our AI SEO & GEO course teaches.