Three sources. That is the working threshold a Wikipedia essay called WP:THREE sets for anyone about to start an article — not three sources that exist somewhere on the internet, but three that would, on their own, carry a subject through a deletion discussion if one were ever opened.
The test sits earlier in the process than most people expect, and it has nothing to do with how the finished article reads or how confident the drafter feels about the subject's importance. It asks a narrower question, before a single sentence gets written: can three pieces of coverage be found, each independent of the subject, each from a source Wikipedia's editors would call reliable, each substantial enough to draft a paragraph from? Passing mentions don't clear that bar, however many of them turn up.
Most drafts run the sequence backwards. Someone who already knows the subject writes what they know, then goes looking for footnotes once a full draft exists. The citations get attached afterward, chosen because they were easy to find rather than because they satisfy anything specific. Such a draft can read as complete and still fail the only test that decides whether it survives, because completeness was never the question being asked.
A paragraph can always be written around a fact. What can't be manufactured after the fact is a second newspaper that happened to think the same fact was worth reporting on its own.
Plenty of material fails the test without looking like it should. A profile on a business database, populated from a questionnaire the subject filled out, carries the subject's own account of itself, dressed up as a listing. An interview is mostly the subject's own words, arranged by a journalist who supplied the questions — a useful quote, but not one of the three. A press release rewritten by four outlets in the same afternoon counts as one source wearing four bylines. A listicle entry — '40 under 40', 'Companies to watch' — usually earns a sentence, and a sentence rarely holds the substance the bar requires.
The bar moves depending on what's being written about. WP:NCORP, the notability guideline for organizations, sets routine business coverage — funding rounds, new hires, product launches, quarterly results — aside from consideration unless the reporting goes beyond the announcement itself, with independent analysis or investigation the company didn't supply. Three sources comfortable enough for a person's biography can fall short for a company's, because the corporate guideline was written with press-release churn specifically in mind.
Editors who run this search well tend to work the same way. They read past the first page of results. They look at what a comparable, already-published article cites — a way to learn which trade publications and beat reporters cover that kind of subject seriously, since one article's sourcing says nothing about another's standing. Then they set aside anything the subject's own team wrote, distributed, or was quoted in, and count what's left.
When the count comes up short, that itself is useful information. It's cheaper to learn it before a draft exists than after one has been nominated for deletion, and cheaper still than after a client has been told the page is coming. Wikipedia's community still makes the final call on anything that reaches a discussion — no search, however careful, decides that in advance. What the search decides is whether there's anything worth bringing to them at all.
