Does AI resume rewriting actually improve your match rate?

Less than the tools claim, and here is the measurement: we took 52 rewrites that our own pipeline had marked as successfully addressing a job requirement, then re-scored the merged résumé with an independent grader that hadn't seen the rewriting session. Only 5 of the 52 — about 9% — were confirmed as genuinely meeting the requirement. A tool that grades its own homework will happily report the other 47 as wins.

The self-grading trap

The inflation mechanism is simple: after you answer a clarifying question, the tool marks that requirement as addressed and recomputes the score locally. In our logs this local re-score ran roughly 20 points above what an independent re-run of the real engine produced. Worst case: users answered "no, I don't have that experience" 33 times, and in 21 of those the displayed rating didn't move at all — the screen still counted the requirement as met. Any before/after number where the "after" is computed by the same code that did the rewriting should be treated as marketing.

What actually moves the number

Real but unwritten experience. When we measured end-to-end lift on 10 real JD-résumé pairs — same frozen rubric, same grader, before and after — the honest average was about +3.5 points. That's meaningful but modest, and it came almost entirely from experience the candidate genuinely had and simply hadn't written down: quantified outcomes, tools used, scope of ownership. Rewriting surfaces what exists. It cannot create qualifications, and the +3.5 only became measurable after we pinned the rubric, because the unpinned noise floor (±10) was three times larger than the effect.

The honest ceiling

If a hard gate fails — the years requirement, a required title, a required credential — no amount of phrasing fixes it, and a rewriting tool that raises your score anyway is scoring its own optimism. There is also a darker failure mode: pushed to improve a score, models start inventing experience. That is a separate problem with its own writeup (see: is it safe to let AI rewrite your resume), but it means an unreviewed AI rewrite can be worse than none.

How to use AI rewriting anyway

• Demand an independent after-score: the final number must come from a fresh scoring run, not from the session that produced the edits.
• Feed it raw material, not adjectives: your real numbers, team sizes, and outcomes. The 9% that genuinely improved were all cases of true facts finally written down.
• Diff every draft against your actual history before sending — you are signing it, not the model.
• Distrust any tool whose score only ever goes up while you use it.

This measurement is why OfferCoach deleted its own local re-score: when you finish a session, it reruns the real engine once, and if that run fails it shows "couldn't compute" rather than a flattering guess.