Site architecture is how your pages are organised and connected: which pages exist, how their URLs are named, which pages link to which, and how many clicks it takes to reach each one. It decides what crawlers find, how Google understands which page answers which query, and increasingly which single page an AI engine picks to quote.
This post covers the principles, a practical way to plan a structure, the AI-search angle most guides skip, and what we learned restructuring our own site this month.
What makes a site architecture good for SEO?
Five properties, in order of how often they are the problem.
One owner page per question
Every question your buyers ask should have exactly one page on your site that answers it best. When three blog posts, a glossary entry and a landing page all half-answer "what is X", search engines have to guess which one to show, and each page competes with the others. Pick the owner, make it the best answer, and merge or link the rest into it.
Hubs that group owners
Group owner pages under a small number of hub pages: a features hub, a guides hub, a compare hub. A hub lists its children, says in a sentence what each one answers, and links to all of them. Hubs give crawlers a short path to everything and give readers a map.
Shallow depth
Important pages should sit within about three clicks of the home page. Depth is not a ranking factor by itself, but deep pages get crawled less, linked less and found less. If a page you care about is five clicks down, link it from its hub and from related pages.
Descriptive URLs and folders
Use readable URLs with words that describe the page, and group similar topics in folders. Google says grouping similar topics in directories can help it learn how often the URLs in each directory change. Keep URLs stable: a renamed URL loses its signals until the new one earns them back.
Links that explain themselves
Internal links carry meaning. Google says anchor text tells users and Google something about the page being linked to, so "see our pricing" beats "click here". Link where a reader would naturally want to go next, not to every page from every page.
How do you plan a site structure?
Work from questions, not from your org chart.
- List the questions. Pull them from Search Console queries, sales calls, support tickets and the questions you track in AI answers. Group duplicates, and note what each searcher wants from the page (search intent decides whether the owner is a definition, a guide, a comparison or a pricing page).
- Assign one owner per question. For each group, decide which page owns it, existing or new.
- Group owners into hubs. Usually by job: what the product does, who it is for, how it compares, how to learn the topic, tools and reference.
- Map the URLs. Hub at /{hub}, owners at /{hub}/{owner}. Avoid more than two levels for anything that matters.
- Draw the links. Every owner links up to its hub, across to three to eight related owners, and on to the next step a reader would take.
- Decide what to retire. Pages that answer nothing and earn nothing go. Pages that duplicate an owner merge into it. Questions with no owner at all are your gaps; a content gap analysis ranks them.
What changes for AI search?
Two things, and both push in the same direction as good classic architecture.
Engines cite passages from pages that clearly answer one question. When an AI engine looks for a source, it wants a page whose opening answers the question directly. A structure with one owner per question produces exactly those pages. A structure with five overlapping posts produces five weak candidates and often none gets cited.
Engines reach beyond page one. Ahrefs found that in March 2026 only 37.9% of URLs cited in Google AI Overviews also ranked in the top 10 for the same query. The engine fans a question out into related searches and picks pages that answer the sub-questions. Owner pages for the sub-questions, linked from a hub, are what it finds.
There is one more practical point: make sure the main text and the links are in the HTML your server sends. Most AI crawlers do not run JavaScript, so navigation built only in scripts can hide your whole structure from them.
How did we restructure our own site?
We are rebuilding attensira.com right now, and the architecture decisions were made from our own measurements. The numbers below are our own workspace's page citations on ChatGPT and Perplexity, read over the Attensira MCP on 4 October 2026.
We are cutting the site to one owner per question. Every URL gets one of three fates: it owns a question and is rewritten, it repeats a question another page owns and is deleted, or a question has no owner yet and a new page is written. Deleted pages simply go away; their best content is absorbed by the owner page that answers the same question. Roughly a third of the site's URLs are being removed.
Most posts earned nothing. Of the 310 blog URLs in our page inventory, 3 received any ChatGPT or Perplexity citation in the 90 days to 4 October 2026, a separate, longer read than the month below. Many of the rest answered the same few questions with slightly different titles. That is the failure mode one-owner-per-question is designed to prevent.
The home page, hubs and comparisons earned the citations. Across 4,524 ChatGPT and Perplexity answers from 4 Sep to 4 Oct 2026 (answers read on 24 of 31 days), the most-cited pages on our site were the home page at www (72 citations), /compare (61), /alternatives/profound and /pricing (58 each) and the head-to-head /compare/attensira-vs-profound (47). Engines leaned on pages that answer one clear question: what are the options, what does it cost, how does it compare.
Crawlers come in through the hubs. The chart at the top of this page is the crawler side of the same picture. In the 30 days to 4 October 2026, AI crawlers made 143,601 requests to attensira.com, by the user agent each request claimed, and each section hub drew between 2,807 and 3,723 of them. No single blog post was among the 30 most-fetched pages. A hub is the short path a crawler takes to everything under it, which is why every owner page needs a link from its hub.
Some sections were never fetched at all. In the same 30 days, at least 121 of our pages drew no AI-crawler request, against at least 657 that drew one or more. The /api-platform hub and all 18 pages under it were in that unfetched list. When a whole section is missing from the crawler log, start with its hub: is it linked from the home page and the navigation, and is it in the sitemap?
Architecture problems showed up in the citations. In the same 4,524 answers, the home page was cited under two host names, 72 times at www and 47 at the bare address, splitting one page's citations across two addresses. And two pages we are retiring, /use-cases/saas-startups and /use-cases, were still cited 41 and 34 times. Our rule for that case: the owner page must carry the answer the old page was cited for, and we read the citations again after the change.
This rewrite is part of the same work: this post now owns "site architecture SEO", and it links to the owners of the neighbouring questions instead of repeating them.