Marginalia Search

By Bill Hartzer · Updated 19 August 2026

What Marginalia is, and why the verdict is an own index

Marginalia is a search engine built by one person in Sweden that crawls the web itself and ranks it more or less backwards from the way Google does. Heavy JavaScript, dense advertising, affiliate links, tracking scripts and modern optimised design push a page down. What rises instead is text: personal sites, old blogs, university department pages, hobbyist documentation — the parts of the web with no marketing budget and consequently no visibility in mainstream results.

The verdict at the top of this page is own index, and it needs no hedging at all. Marginalia runs its own crawler, stores its own index (the engine's stored copy of the web), applies its own ranking, and involves no Google or Bing feed anywhere. Its FAQ states flatly that the engine does its own crawling — and unusually, that claim can be checked from outside rather than taken on trust.

The project's framing is more modest than the coverage around it. It is not seeking to replace Google and Bing, it says, but to complement them — to provide a minority report that keeps them honest. It is a discovery instrument, not a general-purpose answer engine, and has never pretended otherwise.

The crawler, published in the open

Most engines that claim an index of their own are hard to verify. Marginalia is the opposite.

The crawler identifies itself as search.marginalia.nu — the same token in both the user-agent string and robots.txt. That is quietly clever: the token is a URL, so a webmaster staring at an unfamiliar line in a server log can visit it and find out what hit their site. Crawling runs from a small static IP range, 193.183.0.162 to 193.183.0.174, published as a machine-readable text file so operators can allow or deny it programmatically. Blocking takes one standard robots.txt directive, and there is a named human contact address for complaints. Cloudflare independently catalogues the bot as a Search-category crawler operated by Marginalia — external confirmation that it exists and behaves as described.

Refresh runs on a major cycle of eight to ten weeks, with daily RSS polling filling the gap for sites that publish feeds. Two consequences follow, both architecture rather than defect: Marginalia is structurally bad at news and anything time-sensitive, and because the crawl is partly submission-driven, the index is shaped by what people bring to it as much as by link discovery.

Ranking the web backwards

Mechanically, Marginalia is a conventional keyword information-retrieval engine. What makes it unlike anything else at this scale is its value function.

Löfgren has described the mechanism directly: the engine can look at a document's HTML and, if it carries too many advertisements, downrank the site, or expose that judgement as a user filter. The documentation describes heuristics that prefer text-heavy websites and cap heavy modern web design inventions. Filters for affiliate links, tracking and JavaScript are exposed to the user.

Layered on top is an explicit Lindy-effect hypothesis: content that has survived a long time is presumed more likely to be worth reading. Age is a positive signal, the exact inversion of the staleness penalty every commercial engine applies. That single reversal explains most of what feels strange about Marginalia's results, and it is why the engine hands you a 1998 personal homepage no other engine will show you.

This is where most criticism of Marginalia goes wrong. Its results are not worse; they are differently valued. When it returns something odd, that is the product working as designed (crawler documentation). One real scope limit: the project states an English-language focus, so non-English coverage is limited by design rather than oversight.

A small index, on purpose

The FAQ gives roughly 300 million documents as of 2024, occupying about a terabyte, and that page flags itself as outdated; the current about-site says only "hundreds of millions of documents." No dated 2025 or 2026 figure exists, so none is asserted here. Against Google's trillions this is four to five orders of magnitude smaller, and the project treats that as a design position rather than a shortfall. The FAQ's own formulation is the clearest sentence about the whole project: an index with a million documents that are all of high quality is better than an index with a billion documents where only a fraction are interesting.

Under the hood it is custom Java with MariaDB for metadata, crawl data with a shelf life of two to three months, calendar-versioned releases, and an AGPL-3.0 licence with some MIT-co-licensed parts. It is self-hostable in a barebones mode and offered as a white-label engine. It began on PC hardware over domestic broadband and moved to a proper server in October 2023 — meaning an independent web index of a few hundred million documents ran for two years on consumer hardware and a home connection.

One person, and roughly two hundred dollars a month

Marginalia is principally developed and operated by Viktor Löfgren, a Swedish software engineer, working essentially alone. No company, no board, no investor. Wikipedia records maintenance at around an hour a week.

The about page gives operating costs of approximately $200 a month and describes the project as running on donations, grants and commercial API deals, with no investor pressure — designed, in its own words, to scrape by. There is no advertising and none planned. The commercial leg is a keyed API: SearXNG's Marginalia module, for instance, requires a key issued under Marginalia's own licence terms.

Grant funding came from NLnet's NGI0 Entrust Fund, backed by the European Commission's Next Generation Internet programme, from December 2022 to November 2024, with stated goals of maturing the project, improving search quality and range, reducing manual operations and producing portable data for adjacent efforts. The amount is not published, and whether funding continued past November 2024 is not documented.

The $200 figure reframes the argument that independent search is impossible. It does not make Marginalia a Google competitor. It means the cost floor for an independent index of a few hundred million documents is roughly a phone bill. Löfgren's own account is that computers have grown enormously powerful, so ordinary people can now dabble in a space that once required millions in funding (NLnet interview, October 2023).

Privacy that falls out of the architecture, and no AI at all

Marginalia states that it complies with the GDPR by not collecting any personal information, and publishes a policy on access-log retention. No accounts, no advertising, no ad-tech, no tracking, no profiling.

What makes that credible is that it is structural rather than promissory. Most privacy engines ask you to trust a policy. Marginalia has no advertising business and so no commercial reason to profile anyone; it is open source under AGPL, so the code can be read; and decisively, because it runs its own index, your query is never forwarded to Google, Bing or any third party. Nearly every other engine in this category has to send the query somewhere. That is categorically stronger than any proxy arrangement, and it follows from the index rather than from a promise.

The honest caveat: no third-party privacy or security audit was located, and a single-operator service means privacy rests on one person's practices and server security, with no organisational controls behind it.

There are also no AI features, deliberately — no AI answers, no chat, no LLM summarisation. The project positions against the trend, holding that traditional information-retrieval approaches still offer capabilities that have largely become lost. It is a keyword engine by conviction, for users who want to steer a query rather than ask a question. The inverse is quietly true: because it offers a clean keyed API over an index of non-commercial text, it has become attractive to people building AI and research tooling. Marginalia does not do AI; AI systems increasingly want Marginalia.

History, and the launch-date confusion

26 February 2021 — first commit: per the creator's one-year retrospective, a sketch for a website crawler and some data models.
2021 — the engine goes public at search.marginalia.nu, built solo on PC hardware over domestic broadband.
2022 — press discovery. The New Yorker covers it, ABC News and The Verge follow, and it becomes the canonical example of small-web search.
Dec 2022 – Nov 2024 — NLnet NGI0 Entrust grant, funded through the European Commission's NGI programme.
October 2023 — migration to a proper server.
January 2025 — rebrand and move from search.marginalia.nu to marginalia-search.com, announced by the creator; the old interface stays online at old-search.marginalia.nu.
2025–2026 — still operating and actively developed. No lawsuits, breaches or acquisitions found.

Sources disagree about the launch date. English Wikipedia gives 2024, with development beginning in 2021. The creator's own dated blog post puts the first commit at 26 February 2021, and the engine was publicly searchable and press-covered well before 2024. The later date reflects the rebrand, not the launch; this page uses 2021 and treats the rebrand separately.

No usage figures exist. Marginalia does not appear in StatCounter's breakdowns at any level and has never published query or user numbers. It has never claimed scale, and its funding model does not require it.

Who it suits, who it does not, and what people get wrong

It is genuinely good at finding things that cannot be found any other way: a hobbyist's page on a narrow topic, a university department's 2003 reference notes, an unmonetised how-to, an essay by someone with no audience. Google structurally will not surface these, because its ranking reads the absence of commercial signals as an absence of quality. The inversion is not a gimmick but a working retrieval system with a different value function, and the only one of its kind at this scale. It is also reproducible: the code and method are public, so its claims can be checked rather than believed.

It is bad at almost everything a general-purpose engine exists for. It cannot answer a factual question, cannot do news on an eight-to-ten-week refresh, and does not do shopping, maps, images, local results or anything transactional. Non-English coverage is limited by design, and a few hundred million documents against Google's trillions means many queries have no good answer in it. There is also an unhedged bus-factor risk: one person, one server, about an hour of maintenance a week, funded by donations and a grant that expired in November 2024. Nothing suggests trouble; nothing structurally protects it either.

It does not suit anyone looking for a Google replacement, anyone chasing current events, products or local information, or anyone needing non-English results.

What people get wrong

"Small engines can't crawl the web, so it must be metasearch." It crawls the open web itself, with a documented, blockable, IP-published crawler. The belief that independent crawling is infeasible is what this project falsifies.
"Its results are bad." They are differently valued, on purpose.
"A small index means it's a toy." Roughly 300 million documents in 2024 is a genuine web index.
"There's an AI mode." There is not, deliberately.
"Marginalia Science is the same project." Unrelated publication, similar name (Wikipedia).

Frequently asked questions

Does Marginalia have its own index?

Yes, and it is one of the best-evidenced independent indexes anywhere. Marginalia runs its own crawler, identified as search.marginalia.nu in both its user-agent and robots.txt token, from a published static IP range, on an eight-to-ten-week refresh cycle. No Google or Bing feed is involved at any point. Cloudflare's bot directory independently lists the crawler and names Marginalia as its operator.

Who owns and runs Marginalia?

Viktor Löfgren, a Swedish software engineer, working essentially alone. There is no company, no board and no investor behind it. The project runs on donations, grants and commercial API deals against operating costs the about page gives as roughly $200 a month. Grant funding came from NLnet's NGI0 Entrust Fund, backed by the European Commission's Next Generation Internet programme, from December 2022 to November 2024.

Why are Marginalia's results so different from Google's?

Because it values pages by opposite criteria. Heavy advertising, tracking scripts, affiliate links and heavy modern web design push a page down; text-heavy pages rise. Marginalia also treats age as a positive signal under a Lindy-effect assumption, where commercial engines apply a staleness penalty. That single inversion is why it surfaces personal homepages and university pages from decades ago that nothing else will show you.

Is Marginalia a replacement for Google?

No, and it does not claim to be. The project describes itself as complementing Google and Bing rather than replacing them, offering a minority report that keeps them honest. It cannot do news, shopping, maps, images, local results or anything transactional, and its index holds a few hundred million documents against Google's trillions. Used as a daily driver it will disappoint immediately.

When did Marginalia launch?

The first commit was on 26 February 2021, per the creator's own dated one-year retrospective, and the engine was publicly searchable during 2021 and covered by The New Yorker in 2022. English Wikipedia gives a launch date of 2024, which reflects the rebrand and the move from search.marginalia.nu to marginalia-search.com in January 2025 rather than the launch itself.

How do I stop Marginalia crawling my site?

One standard robots.txt directive, using the token search.marginalia.nu, which is the same string the crawler sends in its user-agent. The project also publishes its crawler IP range, 193.183.0.162 to 193.183.0.174, as a machine-readable text file so it can be allowed or denied programmatically, and gives a human contact address for complaints. Few engines of any size document themselves this thoroughly.

Is Marginalia private?

Its privacy properties are structural rather than promised. It has no advertising business and so no commercial reason to profile anyone, the source is public under AGPL, and because it runs its own index your query is never forwarded to Google, Bing or any third party. The caveat is that no independent audit exists, and a solo project means everything rests on one person's practices.

How big is Marginalia's index?

The project FAQ gives roughly 300 million documents as of 2024, occupying about a terabyte, and that page describes itself as outdated. The current about-site says only hundreds of millions of documents. No dated 2025 or 2026 figure has been published, so any precise current number quoted elsewhere is unsourced. The project regards a small, high-quality index as the point rather than a limitation.