YouTube search runs on three factors, and none of them is the tag box: relevance (does your video match the query), performance (do searchers of that query pick it and stay), and personalization (who is asking). Everything a creator controls routes through the first two.
Here is the honest model, including the detail most explanations skip: performance is measured per query, which changes how you should think about everything.
The three-factor model
Relevance: metadata plus the words you say
Relevance reads your title, your description's context, and, underrated, your speech: auto-captions make the words said on camera searchable text. Saying your topic naturally in the first minute is metadata work done out loud.
The tag box contributes its minor typo-catching share, per this site's standing calibration, and nothing more. Relevance gets you into the audition; it never wins the part.
Performance: the audition, scored per query
Here is the load-bearing detail: performance is evaluated for each query separately. The system watches what searchers of "sourdough proofing" do with your video, and separately what searchers of "why is my bread dense" do with it, and ranks it differently for each based on each.
Two practical consequences. First, promise-match decides everything: a video satisfying exactly the searchers its title targets builds a winning record on that query, which is the retention side of the clickbait-debt loop. Second, the topic-research habit of matching format to intent is literally rank engineering: format mismatch shows up as bounce, per query, forever.
Personalization: why your own checks lie
The third factor means the results page is per-viewer: your history tilts your results, so searching for yourself measures your bubble. Judge search performance where it is actually recorded, Studio's traffic-source and query reports, not in an incognito ritual that still is not neutral.
Search is the compounding surface
Suggested feeds spike and fade with sessions; search demand recurs as long as people have the problem. A video with a strong per-query record collects views for years, which makes search-targeted videos the library and suggested-feed hits the fireworks.
Both have a place. But the ranking model above is why the library builds: relevance you write once, a track record that compounds, and a queue of searchers arriving daily to extend it.
The one-line takeaway: relevance auditions you, per-query performance ranks you, personalization shuffles the final order per viewer. Write clear metadata, say your topic out loud, match the format to the intent, and let satisfied searchers build the track record that compounds.