How do linguists tell real language relatives from lookalikes?

They ignore resemblance and look for regularity. A single pair of similar words proves nothing: English day and Latin dies look and mean alike, yet linguists say the match has no value as evidence. What counts is a sound in one language lining up with a sound in the other again and again, across many words that weren't borrowed. When those correspondences are numerous, regular and systematic enough to rule out chance, the languages must descend from a common parent. This is the comparative method.

Lookalikes are traps. English initial d- doesn't regularly match Latin d-, so day and dies are a coincidence. Devil and Latin diabolus do match, but only because both were borrowed from Greek. Even a whole vocabulary can mislead: Modern Persian has more words from Arabic than from its own ancestor, yet it isn't related to Arabic. Meanwhile, real relatives can look nothing alike. Latin habēre and German haben both mean 'to have', but they're false friends. German haben and English have are actually related to Latin capere, 'to seize', while habēre is a cousin of English give.

People noticed resemblances long before they had a method. Some Romans explained the likeness between Latin and Greek by imagining that Rome was a Greek colony speaking a debased dialect. In 1786 Sir William Jones, an English scholar in India, saw that Sanskrit, Greek and Latin were too alike to be accidental and must share a source that perhaps no longer exists. A century later, scholars declared that sound laws have no exceptions, and that regularity became the method's foundation.