Unbiased Search Engines: Does One Exist?

By Bill Hartzer · Updated 19 August 2026

There is no unbiased search engine, and that is not a dodge

A search engine's job is to take millions of pages that all match your words and put one of them first. Doing that requires a value function — a decision about what makes a page better than another page. Fresher? More linked-to? More authoritative? Faster to load? Written by someone with an audience? Every one of those is a judgement, and there is no neutral place to stand while making it.

The best proof of this is not an argument, it is an engine. Marginalia Search ranks the web deliberately backwards from Google: heavy JavaScript, dense advertising, affiliate links and tracking scripts push a page down, and age is treated as a positive signal rather than staleness, on the assumption that content which has survived a long time is more likely to be worth reading. Search the same query on Google and on Marginalia and you get two almost disjoint sets of pages. Neither is neutral. One prefers commercial signals and one penalises them, and both are internally coherent. What distinguishes Marginalia is not that it has no values but that it publishes them, along with its source code.

So “which search engine is unbiased?” has no answer. But it dissolves into five narrower questions that do have answers, and that are worth far more than the original.

The five things people actually mean

Personalisation. Am I being shown different results from everyone else, based on what I have clicked before? This is the filter-bubble complaint, and it is fully separable — some engines simply do not do it.
Commercial self-preferencing. Is the engine putting its own products, or the products of whoever paid, above the results that would otherwise rank?
Editorial intervention. Has the operator demoted or removed particular sources by hand, for reasons it may or may not have disclosed?
State censorship. Is the law of the engine's home country removing results, and are you told when it has?
Access bias. Whose content can the engine see at all? This is the newest and least discussed form, and it is now decisive.

These are different problems with different remedies, and lumping them together as “bias” is why the discussion usually goes nowhere.

Personalisation: the one that is genuinely solved

If your worry is that you and your neighbour are seeing different webs, several engines have already answered it — by not personalising at all.

Mojeek is the clearest case. Its ranking is keyword and relevance based, with no personalisation and no behavioural signals, and every user issuing the same query sees the same page. That is not just an ethical stance; it makes Mojeek uniquely useful for anything requiring reproducibility — comparative research, checking whether a page is indexed anywhere outside the duopoly, or seeing what the web looks like without engagement optimisation. Results you can reproduce are results you can cite.

DuckDuckGo makes the same commitment for the same reason: no user profiling and no personalised reordering, so everyone searching the same term in the same region sees essentially the same page. It calls this “no filter bubble,” and it is the real product. The limitation is that the links underneath are Bing's, so you are getting an unpersonalised view of Microsoft's judgements rather than an independent one.

Marginalia and self-hosted SearXNG also do not personalise, in Marginalia's case because there is no advertising business to profile for, and in SearXNG's because you configure it yourself.

Google, by contrast, is built the other way round. Signed-in search feeds an advertising profile and the controls are opt-out rather than opt-in. That is not a scandal — it is the design of an advertising-funded product, and Google does not claim otherwise — but it means Google results are not reproducible in the way a research citation needs.

Commercial self-preferencing, and what regulators have found

This is the form of bias with the most documented findings behind it, because competition regulators have spent a decade on it.

The European Commission fined Google €2.42bn in June 2017 over the treatment of its own comparison-shopping service in results, €4.34bn in July 2018 over Android, and €1.49bn in March 2019 over AdSense. In the United States, Judge Amit Mehta ruled on 5 August 2024 that Google “is a monopolist, and it has acted as one to maintain its monopoly,” specifically over exclusive default-placement contracts; the remedies issued on 2 September 2025 banned exclusivity, ordered Google to share search index and click-and-query data with qualified competitors, and required it to license results and search text ads — with both sides appealing and no decision as of August 2026. Separately, Google Search is a designated core platform service under the EU's Digital Markets Act, which imposes obligations on self-preferencing and data use. Microsoft Bing is not designated, so it carries none of them.

The everyday version of this is simpler than the litigation: on commercial queries, the organic links are pushed well down a page dense with ads, AI summaries and the engine's own modules. That is a ranking decision in the plainest sense.

Two non-Western engines are worth naming for the same reason. Naver does crawl the open web, but its results page is dominated by Naver's own blogs, cafes and Q&A content — self-preferencing within its own index, at a scale that means many Korean searches never leave naver.com. Baidu has historically drawn the boundary between advertising and organic results far less clearly than Google, and the consequences have not been theoretical.

State filtering, and the engines that carry it

This is where “biased” stops being a matter of taste. Three points, stated as facts rather than allegations.

Baidu filters at every stage — what is crawled, what is indexed and what is served. Politically impermissible content is absent, and the user is generally not told what is missing. The filtering also covers health, pharmaceutical, gambling and financial content. Baidu is not state-owned; it is a Nasdaq- and Hong Kong-listed private company that is legally obliged to censor, which is a different and more accurate description.

Yandex is a Russian company subject to Russian data-localisation and lawful-interception law and to Roskomnadzor blocking orders. Independent measurement has repeatedly found that results served to Russian users suppress or downrank material the Russian state has designated unlawful, including about the war in Ukraine, and that this filtering is not fully disclosed on the results page. Its news aggregator drew from a state-registered list of outlets before being sold to VK in 2022. Rambler has served Yandex's results since 2011 and inherits all of this.

Bing is the one that should give Western readers pause, because it does not stay inside its jurisdiction. Microsoft operates a filtered version in mainland China, and Citizen Lab research published in 2022 by Jeffrey Knockel found that Bing's autosuggestion system censored politically sensitive Chinese names for North American users as well as Chinese ones. Because Bing supplies Yahoo, DuckDuckGo, Swisscows and others, a filtering decision made at Microsoft propagates through a large part of the “alternative” search market automatically.

Editorial intervention by engines that promise not to

In March 2022, after the invasion of Ukraine, DuckDuckGo announced it was down-ranking sites it identified as Russian state disinformation. Some users objected that this was inconsistent with its stated position on unbiased results. Both sides of that have force, and the tension is genuine: an engine that says it does not editorialise, and then does, has taken on a responsibility it had previously disclaimed. It is recorded here not to settle it but because a page about bias that omits the awkward example is not worth reading.

Every engine intervenes somewhere — spam removal, legal takedowns, right-to-be-forgotten delistings, safe-search filtering. The meaningful distinction is between interventions that are disclosed and interventions that are not. Kiddle illustrates the bad end of that spectrum. It is a Google Programmable Search instance with SafeSearch and its own domain allow and block lists on top, marketed to children and used in schools — and after a decade, no company, editorial team, legal entity or accountable publisher has ever been identified behind it, a point a May 2026 investigation made again. A filtered index for children is a reasonable product. A filtered index for children with no identifiable operator is a different thing.

“Uncensored” engines are usually the least accountable

There is a whole category of engines marketed as the cure for search bias, and it deserves scepticism rather than sympathy.

Gibiru has described itself for years as “the only fully anonymous, uncensored search engine,” with a toggle promising pages the mainstream suppresses. It has never published a crawler, a user agent or an index size in seventeen years of operating. Its founder described a modified Google feed in 2012; independent testing by Larry Sanger in 2022 concluded that its results were routed through Google and that the “uncensored” label was, in his assessment, strictly marketing. Reviewers in 2026 still cannot identify who runs it. The general point is more important than the specific one: an undisclosed result set assembled by unnamed criteria is a worse epistemic position than a documented mainstream one, not a better one. “Uncensored” means differently filtered, by someone who has not told you how.

Presearch markets decentralisation as the answer to bias. Its documented mechanism is still metasearch — its own privacy policy logs which provider answered each search — and the own-index claim it began making in September 2025 has not been accompanied by a crawler name, user agent or index size.

Access bias: whose content the engine can even see

The newest form of bias is not about ranking at all. In 2024 Reddit blocked non-Google crawlers while Google retained access under a paid deal. Overnight, an enormous body of human discussion became visible to one engine and invisible to every small one. Mojeek, Marginalia and every other independent index lost it simultaneously, and no amount of engineering can win it back — it is a licensing barrier, not a technical one.

This is now the mechanism by which search concentration maintains itself, and it deserves more attention than personalisation does. An engine cannot be unbiased about material it is not permitted to fetch, and the set of engines permitted to fetch the most valuable material is shrinking.

What you can actually do about it

Three things are available now, and none of them requires believing an engine is neutral.

Choose an engine whose values are published. Marginalia states its ranking philosophy and publishes its code under the AGPL, so its bias can be read rather than inferred. Mojeek states that it does not personalise and that results are identical for every user. That is a much stronger position than an engine that claims neutrality and documents nothing.

Change the ranking yourself. Brave Search offers Goggles — a small language for writing re-ranking filters that boost, demote or exclude whole sets of sites, shareable as a URL, with a limit of 10,000 instructions each. It is the only mainstream implementation of the idea that the user, not the engine, should choose the value function, and it exists only because Brave owns its index. No reseller can offer it.

Use two indexes, not two interfaces. This is the practical habit, and the one most readers miss. Comparing Google against Startpage, or Bing against Yahoo or DuckDuckGo, tells you nothing about bias, because you are querying the same index twice. Comparing Google against Mojeek or Marginalia tells you a great deal, because they crawled the web separately and value it differently. The gap between the two result sets is the most direct look at editorial bias any reader can get without a research budget — and it costs nothing but a second browser tab.

Frequently asked questions

Is there such a thing as an unbiased search engine?

No, and the claim should be treated as a warning sign. Ranking requires deciding what makes one page better than another, and every such decision embodies values. The useful questions are narrower: does the engine personalise results, does it favour its own products, does it filter by hand or by law, and does it disclose any of that. Those are answerable, and they differ enormously between engines that all describe themselves as neutral.

Which search engines do not personalise results?

Mojeek states that it applies no personalisation and no behavioural signals, so every user issuing the same query sees the same page — which also makes its results reproducible for research. DuckDuckGo makes the same commitment, though its links come from Bing. Marginalia does not personalise because it has no advertising business to profile for, and a self-hosted SearXNG instance does not because you configure it yourself.

Which search engine is least censored?

None is uncensored, and engines marketed as uncensored are usually the least accountable. Every engine removes spam, complies with legal takedowns and applies safe-search filtering somewhere. The meaningful comparison is disclosure: Baidu and Yandex filter in line with national law without telling users what is missing, and Bing's China-related autosuggest filtering has been documented reaching North American users. Engines that publish their ranking rules, such as Marginalia, are the honest end of the range.

Is Google biased?

In documented and specific ways, yes. The European Commission fined it €2.42bn in 2017 over the treatment of its own shopping service in results, and a US court ruled in August 2024 that it had acted unlawfully to maintain a search monopoly. It also personalises for signed-in users and places its own modules, ads and AI summaries above organic links. None of this makes its results untrustworthy on ordinary queries; it means its ranking reflects a commercial interest as well as a relevance judgement.

Do private search engines give unbiased results?

Not necessarily, and most cannot. A privacy engine that resells Google's or Bing's results inherits its supplier's ranking, omissions and filtering unchanged — the privacy layer changes who is watching, not what is shown. Startpage relays Google and Bing, DuckDuckGo's links come from Bing, Swisscows resells Bing. Only engines with their own index, such as Brave, Mojeek and Marginalia, can differ from the big two on what appears at all.

How can I tell if a search engine is filtering results?

Run the same query on two engines that crawled the web independently — Google or Bing on one side, Mojeek or Marginalia on the other — and compare. Comparing two interfaces onto the same index tells you nothing. Also check whether the engine documents its ranking approach and its removals at all: an operator that publishes a philosophy and source code, as Marginalia does, can be checked; one that publishes nothing cannot.

Are “uncensored” search engines a good alternative?

Usually not. Gibiru has marketed itself as fully anonymous and uncensored for seventeen years without ever publishing a crawler, user agent or index size, and independent testing in 2022 concluded its results were routed through Google. An undocumented result set compiled by unnamed criteria is not less filtered than a mainstream one; it is filtered differently, by an operator who has not said how, and often cannot be identified at all.

What is access bias in search?

It is bias in what an engine is allowed to crawl rather than in how it ranks. In 2024 Reddit blocked non-Google crawlers while Google retained access under a paid deal, so an enormous body of discussion became visible to one engine and invisible to every independent one. This is now a leading mechanism of search concentration, and unlike ranking bias it cannot be fixed by better engineering — only by licensing or regulation.

Unbiased search engines: what is actually possible