Mirakle Research · June 14, 2026

What happens when someone changes their mind?

A review on the Shopify App Store isn't set in stone — people can go back and change the words, or the stars. We've been watching for it. And at first the numbers looked like a disaster: 74,146 reviews changed their rating, and 99.7% of those changes were downgrades — almost all of them dropping straight from five stars. A mass exodus of goodwill. Except it never happened.

Daily review changes. The red line is rating changes; the teal line is real text edits. Notice anything about May 5?

On a single day — May 5, 202673,018 reviews "changed their rating," 98% of every rating change we've ever recorded, all at once. Reviewers did not have a collectively terrible morning. We had a bug.

Showing our work

Every "downgrade" started at five stars

For a while, our scraper read the star rating off the wrong piece of markup and recorded almost every review as a perfect five. When we fixed it and re-scanned the store, tens of thousands of reviews snapped to their true rating — and our change detector dutifully logged each one as a "5 → 1" or "5 → 4." Lay the transitions out in a grid and the artifact is impossible to miss: a single blazing row where the old rating is five, and almost nothing anywhere else.

Rating changes, old rating (rows) versus new (columns). The entire story is one row — the fingerprint of a measurement fix, not 74,000 sad customers.

This is the boring, important part of working with data nobody else has: the most dramatic number in your dataset is often the one your own tools invented. So we threw all 74,146 of them out and kept only the edits we can actually trust — the ones where a human went back and rewrote their own words. The bug never touched text. There are 1,861 of those, and they tell the opposite story.

The real finding

When people truly change a review, they forgive

Take the edits where the reviewer rewrote their text and changed their rating from something other than the suspect five — 245 genuine changes of heart. 89% of them are upgrades. The single most common move on the whole board is the most dramatic one possible: one star to five, which happened 80 times. People don't mostly come back to a review to twist the knife. They come back to take it back.

Genuine rating edits (text also rewritten, original rating not the buggy five).
The most common real transitions. Green is an upgrade, red a downgrade.

The redemption arc

"Lost so much money" → "App is perfect. Thank you."

These are real, anonymised before-and-afters: the same review, furious one week and grateful the next. Read them as a genre. Almost every one has the same hidden second character — a support team that showed up.

Two days of downtime, then four words of total absolution.
"Do not buy" became "highly recommend" once someone replied.
You can almost watch the blood pressure drop between the two boxes.

Who saved the review

It's almost always a person with a name

Read enough of these and a pattern jumps out: people don't come back to thank a company, they come back to thank a human. The most common kind of non-rating edit is a positive review that returns weeks later to bolt on a postscript — an "Update:" naming the support rep who fixed things. The review was already five stars. They edited it anyway, just to give someone credit.

The "Update:" that exists purely to name a helpful human.
Came back specifically to credit the person who helped.

The shape of a second draft

Edited reviews grow

When someone reopens a review, they almost always add to it rather than cut it back: 1,344 edits made the review longer and only 482 made it shorter, a median of 22 characters added. A review is rarely deleted in anger; it's amended, like a diary entry with a calmer note added underneath. And the smallest edits are their own quiet comedy — a "coll" fixed to "cool," a second exclamation mark that the first one clearly needed.

What an edit does to a review's length.
"100% recomend" → "recommend." The irony was, eventually, noted.
One exclamation mark was not, on reflection, enough.

…and occasionally

The rare genuine downgrade

They do exist — the real ones, where a happy review curdles. Strip out the bug and only 26 genuine downgrades remain in the whole window, usually a calm review that came back to append a specific grievance: a price hike, a feature taken away, a support thread that went cold. They read less like rage and more like disappointment, which is somehow worse.

"Works fine" — until it didn't, and they came back to say so.

So: is the most negative-looking metric on the App Store actually a catastrophe? No. Underneath a measurement bug that faked 74,000 downgrades, the real story is small and stubbornly hopeful. A few thousand people went back to reviews they'd already written — and mostly, they returned to be kinder. The App Store's reviews aren't carved in stone. They're carved in pencil, and the eraser usually points up.

How this was measured

  • We detect a review edit by scraping the same review more than once: 1,861 text edits and 74,146 raw rating changes across 923,667 tracked reviews.
  • The rating-change count is mostly a fixed bug. Our scraper once parsed nearly every review as five stars; the corrective rescrape on May 5 made ~73,018 reviews appear to "drop" from five to their true rating in a single day. All real findings here use the bug-immune subset: edits where the text changed and, for direction, where the original rating was not the suspect five.
  • Every example is hand-sanitised. All before/after pairs are stripped of store names, reviewer and person names, emails and links, and shown anonymously with neutral placeholders. We are here for the pattern, not to identify anyone.
  • The detection window is short and the true counts are lower bounds — there are surely more edits we never caught between two scrapes.

Independent research by Mirakle. Not affiliated with or endorsed by Shopify or any developer. The bug has, of course, since been fixed.