Wait long enough for Google to re-crawl the changed pages and for enough data to accumulate to be meaningful, which in practice usually lands somewhere in the range of a few weeks to a couple of months rather than a single day. There is no honest fixed number, so treat that range as a working guide, not a rule. Checking tomorrow tells you almost nothing, because Google may not have even seen the change yet, and even if it has, a day of traffic is far too little to separate a real effect from ordinary noise.
Two clocks run before a judgment is fair, and you wait for the slower of them. The first is the crawl clock: your edit does nothing to rankings until Google re-crawls and re-indexes the page, and how fast that happens depends on how often the site is crawled. A large, frequently updated, well-linked site may be re-crawled within days, while a small or new site can take considerably longer, so a low-crawl site needs a longer window simply to register the change at all. The second is the data clock: even after re-crawl, you need enough clicks and impressions for the before-and-after comparison to mean something. A page with heavy traffic reaches a readable sample quickly, while a low-traffic page may need many weeks to gather enough data to trust.
This is why a precise day count makes a poor rule. The right window for a high-traffic, fast-crawled page is genuinely shorter than for a sleepy, rarely crawled one, and forcing both into “wait 14 days” will have you judging one prematurely and the other too late. The patience threshold is set by your situation, not by a calendar figure someone repeated online. Worth confirming against your own crawl stats and traffic levels rather than adopting a number wholesale.
So before you make the change, set the window: estimate how quickly these pages get re-crawled and how much traffic they pull, and pick a judging date accordingly, leaning longer when crawl is slow or traffic is thin. Then hold to it, resisting the urge to read the early wobble, and evaluate only once both clocks have run. Decide the window up front, by crawl frequency and data volume, so you are judging a settled result instead of reacting to noise.