Marginalia Search resurfaces by ranking the web Google leaves behind

Why it matters

Marginalia shows that an independent search index can run without venture scale, while exposing how much of search quality comes from ranking incentives rather than the query box.

Swedish software engineer Viktor Lofgren is getting another burst of attention for Marginalia Search, the independent search engine he started building in 2021 to surface personal sites, old blogs and other pages pushed below the commercial web.

A Reddit post on September 21st framed Marginalia as a one-man attempt to fight Google. That description oversells Lofgren's ambition and undersells the technical work. Marginalia Search operates its own crawler and index instead of pulling its results from Google or Bing, an expensive distinction that separates it from many alternative search interfaces.

The engine itself is five years old. Lofgren made the first commit on February 26th, 2021, during the pandemic, after growing frustrated with an internet that seemed smaller and increasingly difficult to explore. He envisioned a keyword search engine resembling Google in the late 1990s, built on PC hardware and tuned to find pages overlooked by mainstream ranking systems.

Marginalia now prioritizes non-commercial, text-heavy websites and lets users filter pages containing tracking, affiliate links, scripts and other attributes. Its current homepage describes the service as using "simple technology, no AI." An Explore tool sends users through sites in the index without requiring a conventional query.

That makes Marginalia useful for discovery rather than as a universal answer machine. Lofgren's own description of the project warns that it may struggle to locate a page a user already knows exists and calls it the wrong tool for finding trusted facts. Its purpose is serendipity: finding material a user did not know to search for.

A search index built around different incentives

The ranking choices follow from Lofgren's complaint about modern search. Commercial engines have incentives to serve pages designed around advertising, transactions and search optimization. Marginalia can downrank those characteristics because its business does not depend on selling ads beside the results.

Lofgren told NLnet in a 2023 interview that ranking algorithms inevitably encode the values and perspectives of their designers. A small number of search indexes therefore leaves developers with a narrow set of source material, even when numerous search brands appear to offer competition.

Marginalia applies its own answer to that problem. Its crawler collects pages directly, while its ranking system favors older-looking, text-oriented sites and limits the weight given to heavy modern web design. The project has focused primarily on English content, reflecting the constraints of an index maintained by one developer.

The project is open source, though self-hosting a meaningful internet index still requires substantial storage and memory. The Marginalia Search repository is predominantly available under the AGPL 3.0 license and had about 2,100 GitHub stars as of September 21st. Its documentation says a production-like installation benefits from enterprise SSDs, several terabytes of storage and substantial RAM. Smaller deployments can reduce the index size.

Lofgren studied physics and spent much of the 2010s working on high-reliability Java systems in financial technology, according to his own coding history. His earlier projects included a trinary computer emulator with an assembler, operating system, development environment and compiler. Marginalia extends the same habit of rebuilding difficult infrastructure to understand what assumptions are hiding inside it.

Grants bought the crawler time

The renewed social-media framing says Marginalia has no investors or loans. The project has avoided a conventional venture-backed model, though it has received material grant support.

NLnet funded Marginalia through the European Commission-backed NGI0 Entrust program from December 2022 through November 2024. Lofgren began working on the engine full time in June 2023, using a combination of savings, grant money and donations. FUTO awarded the project a $15,000 grant in September 2023 to help move the search engine from consumer PC hardware to a server. Marginalia received a second NLnet grant in March 2025, which Lofgren said extended its development runway. The amount was not stated.

The project remains active. In a July 22nd, 2026 engineering update, Lofgren said he had moved production services from Docker Compose to systemd, added a cheaper path for unranked queries and separated large hosting domains into their own crawler partition. The change reduced the main crawler's run from nearly two weeks per partition to roughly five days, while sites hosted on services such as Substack, WordPress and GitHub Pages are processed separately.

Those operational details explain why independent search remains rare. A crawler must continually revisit a web full of dead links, rate limits, spam and dynamically generated pages. The index then needs enough computing power to answer queries quickly, without giving the most commercially optimized documents automatic control of the first page.

Marginalia does not offer Google's breadth, freshness or utility for routine navigation, and Lofgren has consistently presented it as a complement rather than a replacement. Its value comes from maintaining an index built around a different definition of relevance. Five years after Lofgren began the project, that difference is why people keep rediscovering it.