How Search Engines Classify Websites by Source Type

Website classification is the site-level complement to document understanding. A Website Classification System generates representations for websites and then classifies them — determining, in the documented implementation, whether a site resembles an "expert, apprentice, or amateur source" using "visual and layout-related embeddings and features," and distinguishing among "affiliate sites, aggregators, service providers, ecommerce sites, and SaaS platforms."

The reason this layer exists is the same reason behavioral signals exist: an engine's direct understanding of documents is minimal. Classifying a whole site once — by layout, components and function — is a cheap, stable substitute for re-understanding every page from scratch, and it supplies the context in which each page is read: what the site can do, whom it serves, and how much trust its statements start with.

The core mechanism is the website representation vector: a composite representation assembled from feature vectors derived from the site's pages. The inputs named in the processing literature are mostly structural — "HTML tokens," the "DOM tree," "page layout," "hierarchy," "visual relationships," "annotations," and the "functional meaning of each structured information block." Models such as WebFormer incorporate the HTML layout into the document representation, and site-wide vocabulary features round out the picture.

• Layout and hierarchy: how pages segment into regions and which region dominates.
• Functional components: the interactive elements that say what the site does — purchasable, bookable, comparable, calculable.
• Site-wide vocabulary: the nouns and phrases that recur across the domain, which make a site labelable within its knowledge domain.
• Crawl attention: the crawl rate itself is read as a measure of how much the engine values the site.

Classification is comparative: a new site's representation is measured against composite representations of known classes, and the closer it sits to a class, the more that class's expectations transfer to it. There is no application form — the class is inferred from what the site shows.

The sharpest documented consequence: "the same content can rank differently on an affiliate website than it does on an ecommerce website," because engines "classify websites by their type rather than their content quality." Different types carry different quality thresholds and different intent expectations. An ecommerce site is a task-completing commercial resource built for transactional intent; an affiliate site typically aggregates information to direct traffic outward, and it earns its place only by demonstrating real effort, originality and added value over the sources it draws on. One documented move — the same content and context relocated from an affiliate site to a commercial one — improved rankings for exactly that reason:

"The content itself didn't change. What changed was the function, context, and source type surrounding it."

— Semantic SEO source-type case study

Intent alignment completes the logic. E-commerce pages are expected to serve "Do" queries — purchasing, comparing, ordering — while affiliate-style content more naturally serves "Know" queries that precede the transaction. When the page's demonstrated capability matches the intent its source type implies, the engine spends less effort deciding whether the result fits.

Classification decides eligibility — whether a source is "even in the game" for a query class — before any position is contested. It feeds the site-level trust scores (siteAuthority and the broader Q*), helps the engine interpret queries by associating them with knowledge domains and site types, and it is integrated with the Helpful Content System, whose classifier leans on the same function-and-layout evidence. Generative systems use it too: classification informs which sources are appropriate to cite in AI answers.

The classification is also stable but not frozen: layout changes "can produce different vector representations, which may affect how a document is understood, classified, and retrieved." A site can drift between classes when its presentation drifts — which is why classification is a standing property to maintain, not a one-time verdict.

None of the levers are exotic; they all make the site's type unambiguous:

• State and show the purpose: the page layout and its components should make the business model and the user task obvious without reading a word of marketing copy.
• Ship the function the type implies: commerce means working purchase, compare and filter paths; a service means booking and contact that work.
• Keep appearance honest with function: never imitate a function the page does not provide — misleading functionality is classified as spam.
• Hold one coherent focus site-wide: consistent topics and vocabulary make the site labelable; scattered topics are described as a self-inflicted eligibility problem.
• Stay out of the abuse classes: scaled production, borrowed reputation and recycled expired domains all end in negative classifications and demotion.