The honest answer is that you largely cannot cleanly isolate three simultaneous changes after the fact, so the real lesson is not to make them simultaneously in the first place. When you ship a title rewrite, an internal-linking change, and a content update in the same window, the data afterward shows only the combined effect, and no amount of analysis can reliably split a single blended outcome back into three clean contributions. Pretending a tidy attribution method exists for confounded changes would be a lie; what exists is loose triangulation with real limits, and a discipline that prevents the problem next time.
What you can do after the fact is triangulate, with each tool’s weakness named. Timing helps a little: if rankings moved noticeably right after one change was crawled but before another took effect, that sequence hints at which one mattered, though crawl timing is fuzzy and changes rarely register on neat separate days. Affected-pages analysis helps more: if only the pages that received internal links moved while pages that only got the content update stayed flat, that points toward the linking change, because the footprint of each edit differs across your site. Segmenting by query or page type can sharpen this further. But every one of these is suggestive, not conclusive, because the changes overlap on the same URLs and the same time window, and confounding does not dissolve just because you stared at it harder.
So treat past confounded changes as loosely diagnosable at best. You may end with a reasonable hunch about which change drove the result, and a hunch is worth having, but you should hold it lightly and not build future decisions on it as if it were proven. The cost of the simultaneous rollout is exactly this permanent uncertainty, and the only real fix is procedural, applied going forward.
The action, then, is two-part. Going forward, change one variable at a time, wait the re-crawl-and-settling window, judge the result, and only then make the next change, so each effect is cleanly attributable. For the mess you already have, triangulate by timing, by which pages were affected, and by segment, draw your best provisional conclusion, and label it provisional. Buy clean attribution with sequencing in the future rather than chasing it in data that can no longer give it to you.