A meta-search engine holds no index. It has never crawled a page. When you press enter it sends your query out to other engines, waits for their answers, merges and de-duplicates them, and shows you one combined list. Every result you see originated somewhere else, and was retrieved in the second or two you spent waiting.
Why that was a genuinely good idea in 1995
In the middle of the 1990s no search engine had indexed most of the web, and the ones that existed had indexed different parts of it. Crawlers started from different seed lists, followed different link paths, refreshed at different rates and capped their indexes at different sizes. Two engines could return almost entirely different results for the same query and both be working correctly. Under those conditions searching one engine meant knowingly missing most of the answer, and searching five by hand was a chore. Merging them automatically was an obvious and valuable service.
MetaCrawler, built at the University of Washington and described in a 1995 paper by Selberg and Etzioni, is the clean statement of the idea: post the query to several services in parallel, collate the returned references, and — a step later systems mostly dropped — actually fetch each returned URL to confirm it still existed and to score it. That verification step was doing real work at a time when link rot was rampant and engines routinely returned pages that had been gone for months. SavvySearch came out of Colorado State University in 1995, ProFusion out of the University of Kansas, and Inference Find, Highway 61, Mamma and Dogpile followed. Several were academic projects before they were companies, and the field was well-populated enough that "which metasearch engine" was a real question with a real answer.
What changed
Three things happened, and together they removed most of the value.
- Index coverage converged. Google's index grew until missing pages stopped being the ordinary experience of search, and ranking rather than coverage became what distinguished engines. Merging two ranked lists that were drawn from largely the same corpus mostly reorders results without adding any.
- The suppliers consolidated into two. This is the decisive one. In 1996 a metasearch engine querying half a dozen services was reaching genuinely different indexes. Today an engine querying Google, Yahoo and Bing is reaching two, because Yahoo has served Bing's results since the migration completed in the early 2010s. Several of the properties in this category still advertise supplier lists assembled when those names meant different things. MetaCrawler's own about page, retrieved in August 2026, still named a video service that shut down years ago and contained an unreplaced internal template string — a fair indication of how much maintenance a legacy metasearch property receives.
- Being queried stopped being free. Search engines have a direct commercial interest in serving their own results pages with their own advertising, not in feeding a competitor's merged list. Access moved behind paid APIs, then behind fewer paid APIs: Microsoft raised Bing API pricing sharply in 2023 and retired the public Bing Search APIs entirely on 11 August 2025. The free-scraping alternative is actively defended against. SearXNG's public issue tracker documents Google blocking fresh instances after a handful of consecutive queries, with the number of public instances that had a working Google connection collapsing over a matter of days during one 2023 episode; the project ships a rate-limiting component precisely because this pressure never stops, and instance directories publish per-engine failure rates because engines break constantly.
That last point is structural rather than a bug awaiting a fix. A metasearch engine's results depend on the continued willingness of companies that have a commercial reason to refuse. No amount of engineering resolves that.
Where metasearch still earns its place
The idea is not dead, but its purpose has shifted from coverage to interposition. SearXNG — free software, no company, no central service, a community fork of the older searx project begun in 2021 — is metasearch used as a privacy mechanism. Your IP address, cookies and browser fingerprint reach the instance you use and go no further; the upstream engines see a server, not you. Because anyone can run an instance, and because which upstreams are enabled, how they are weighted and whether anything is logged are all per-instance settings, any sentence of the form "SearXNG gets its results from X" is meaningless unless it names an instance. Two instances can differ in both results and privacy. Self-hosting resolves the trust question and inherits all of the blocking problems.
The other durable uses are narrow and real: aggregating specialist sources that genuinely do not overlap, and comparing engines side by side to see how differently they rank the same query — a research use rather than a daily driver.
What the surviving commercial properties actually are
Dogpile, WebCrawler and MetaCrawler are all now owned by the same company, InfoSpace Holdings, a System1 subsidiary, and are operated as advertising properties over bought-in feeds. None publishes its current supplier contracts. WebCrawler is the one worth pausing on: launched in 1994, it was the first engine to index the full text of pages rather than just titles and headings, and it switched to serving another engine's database in 2001. The site named after the first full-text web crawler has not crawled anything for a quarter of a century.
Presearch is the modern variant of the argument. It distributes the query-fetching work across a network of operator-run nodes and pays them in a token, and it has announced that it now runs an independent index of its own. On published evidence as of August 2026 the documented mechanism is still federated metasearch: there is no bot information page, no user-agent string, no robots.txt token, no published crawl IP ranges and no index-size figure — the artefacts every real crawler generates, and which a one-person project like Marginalia publishes in full. Distributing retrieval across volunteer nodes also distributes labour, not authority; the operator still runs the front end, the accounts, the token and the supplier relationships.