Insights · Patent history
Making websites searchable: a technical idea from 2005
In 2005, I was working on a problem at the intersection of website architecture and search: how information inside a website could be made easier for search engines to discover.
The resulting published patent application describes a method and system for improving website searchability. The publication number is WO2007073262A2.
Today, AI systems are changing how machines discover, interpret, and use information across the web. The technology is different. The architectural question is still familiar: how should information be structured so that machines can find it and make sense of it?
The original problem
The application starts from an existing website that search robots had trouble reading. The techniques it names are scripts, session IDs, frames, DHTML, image maps, Flash, and other plug-ins. A robot might fail to retrieve the information, so the pages ranked poorly or did not appear at all.
Rebuilding a large site in search-friendly techniques was expensive. The text also describes two practices it treats as risky: shadow or mirror domains on a separate domain, often controlled by an SEO vendor, and doorway pages stuffed with keywords. The search engines named in the description are Google, Yahoo, and AltaVista.
What the application describes
The method identifies keywords or themes for the information already on the site. For those themes it creates search-engine-friendly pages and publishes them on the same domain, including a subdomain. Each new page links to a page on the original site that carries the related information.
The application is explicit that these pages are not a one-to-one copy of the original URLs. They map the information content, so a large site can be represented by fewer pages. A preferred embodiment adds a sitemap of those pages, navigation by theme, and an internal ranking based on click rate. A described tool, built with ordinary server-page technology of the time, lets an editor write the page, check that the keyword appears in the title and meta tags, and publish it after a second person approves it.
Publication and inventorship
The record is international application PCT/SE2005/002025, filed on December 23, 2005 by TWO I SVERIGE AB. It was published on June 28, 2007 as WO2007073262A2.
The inventors named on the publication are Stefan Helgesson and Lars Anders Tufvesson. I am one of the two. The application does not describe my employment relationship with TWO I SVERIGE AB, and this page does not add one.
WIPO informed the European Patent Office that Europe was designated, on 26 September 2007. The application did not enter the German national phase (24 June 2008). It did not enter the European regional phase (28 January 2009). Google Patents lists the status as ceased, and says that listing is not a legal opinion. This page describes a published application. It does not describe a granted patent that is in force.
What has changed since then
The crawlers in the text are the search engines of that decade. Today’s discovery also includes systems that summarize pages, answer questions, and act on what they read. The application does not describe language models, embeddings, training data, or citations inside an answer.
Some of the obstacles it names, such as Flash and frames, are no longer the usual way a site hides its content. Script-rendered pages, session-bound URLs, and content that a crawler never receives are still reasons a machine can miss what a person sees.
What remains relevant
The useful comparison is architectural. The application asks how to put a machine-readable account of a site’s information on the same domain, tied back to the pages a person actually uses. An AI-readable site still has to answer a version of that question: what is on the page, which URL is the source, and how a machine gets there.
The comparison stops there. The application is a method for helping the search robots of 2005 index an existing site. It is not a design for a language model, for generative engine optimization, or for an agent protocol.
What the record proves, and what it does not
The publication shows that this problem was written down in 2005, that two people are named as inventors, and that the applicant was TWO I SVERIGE AB. It shows a proposed method. It does not show that the method was built, sold, or used at any particular scale. It does not show that the idea anticipated modern AI systems.