Data

Yelp API shut down: what happens to reviews when a source dies

14 September 2026 · JustReview

The Yelp API shut down for most integrations in early 2024, and the question every business asked next was whether the reviews they had already collected went with it. In our database, they did not. The 1,952 Yelp reviews belonging to 46 businesses on the old route are still stored, still flagged as visible, and still being served today, frozen at 5 February 2024, the last day a new one arrived. That is the pattern behind every source closure we have lived through: the archive survives, the tap stops, and nothing on the page tells a visitor which of the two just happened.

This post is the inside view of three of those closures, taken from our own production database on 14 September 2026, where 1,386,276 reviews belonging to 3,741 businesses sit across 53 source identifiers. Two of the three were real. The third only looked like one, and the difference between them is the most useful thing here.

What a closure looks like from inside a database

A source closing does not produce an error. It produces silence, and silence is indistinguishable from a quiet month. The only reliable signal is a pair of dates that most dashboards never show side by side: the date of the newest review, and the date anything was last imported from that source. When both move together, the source is alive. When the import date keeps moving but the review date does not, the pipe works and the business simply is not earning new reviews. When both stop on the same day and never resume, the door closed.

The reason this matters commercially is that a frozen widget keeps working. It keeps rendering, keeps passing its own health checks, and keeps serving traffic, which is exactly why nobody notices. The failure mode is not a broken page. It is a page that quietly stops being evidence.

Case one: Yelp, 5 February 2024

We hold 94 Yelp connections. They split into two routes that are invisible to the person who set them up, because both were connected through the same screen.

Fifty connections have produced nothing in thirty-one months. Not a slow trickle, not a reduced volume: zero rows. The nineteen connections that were moved onto a different collection route kept importing normally and were still doing so a week before this was written. Yelp reviews are therefore still obtainable in 2026. They are just not obtainable through the integration most tools were originally built on, and no message anywhere told the 46 affected businesses that their list had stopped growing.

What survived is the part worth noticing. Every one of those 1,952 reviews is still in place, and eleven of the affected businesses served 16,677 widget loads in the last 90 days from that frozen list. Their social proof did not disappear the day the API closed. It stopped compounding.

Case two: Amazon product reviews, 1 November 2024

This one is larger by two orders of magnitude and is the clearest example of why holding your own copy is not a theoretical benefit.

Amazon exposes two different things that both get called reviews: feedback about the seller, and reviews about the product. We import them through separate routes. On 1 November 2024, the product route returned its last row. The seller route never stopped and was still importing on 14 September 2026.

Three hundred thirty-seven thousand reviews stopped being collectable overnight, and 337,350 of them are still visible today because they were already stored. Twenty-six of those 89 businesses served 37,856 widget loads in the last 90 days on that archive. Had those reviews lived only on Amazon’s servers behind a widget that read them live, all 26 would have shown an empty box on 2 November 2024.

The silent part: 175 live Amazon connections still have the product-review switch turned on right now. The setting is saved, the interface confirms it, and it has produced nothing for almost two years. A saved setting is not a working pipe, and this is the single most common reason a review count on a widget does not match the platform.

Case three: eBay, December 2025, the one that was not real

In November 2025 we imported 2,014 eBay reviews. In December, 30. In January 2026, eight. Every symptom of a closed door was present: a cliff between two months, no errors, no announcement, and a source that had been reliable for years.

It was not eBay. It was our own importer, which had been walking eBay’s feedback list on the assumption that it was ordered newest-first. It is not. Once a seller accumulated enough history, the importer kept re-reading pages it had already seen and starved itself of new rows while reporting success on every run.

Of the 24,824 rows imported in July 2026, 6,824 carried review dates from before the outage: 3,669 older than the gap and 3,155 inside it. Those were not new reviews. They were reviews that had existed on eBay the whole time and that we had simply failed to collect. eBay’s own record was intact, so a backfill recovered them, and the source has been healthy since, with the most recent review dated 14 September 2026 across 202,720 stored reviews.

The lesson is uncomfortable and worth stating plainly: two of these three cases were platforms closing a door, and one was us. From the outside they were identical. If you cannot tell the difference, you cannot tell whether waiting will fix it. Ours took seven months to notice because nothing was throwing an error, only producing less.

What is actually lost when a source closes

Very little, if a copy already exists. Everything, if it does not.

Reviews you already collected stay yours. In our database they keep rendering, keep their star ratings, keep their dates, and keep feeding the aggregate rating your page shows. A closure does not reach back into a stored copy. What you lose is the future: every review published after the door shuts is written on a platform you can no longer read from, and for sources that only expose a rolling window of recent feedback, those reviews eventually age off the platform too and then exist nowhere you can reach. That is the one genuinely irreversible loss, and it is why connecting a source early matters even if you do not display everything from it yet.

There is a second, quieter loss: credibility drift. A widget showing a five-star average from 2024 is not neutral in 2026. Visitors read dates, and a page whose newest review is two years old reads worse than a page with fewer, newer reviews.

How to protect the reviews you have already earned

1. Keep the reviews outside the platform. This is the whole argument. A widget that renders from a stored copy survives a closure; a widget that queries a platform live does not. Our Google reviews widget reads from our copy, which is why the Yelp and Amazon archives above are still on the internet at all.
2. Connect more than one source. On 14 September 2026, 2,818 of the 3,637 accounts with a live connection, 77.5% of them, were running exactly one. One platform decision freezes their entire widget. The fix costs nothing on most plans and takes a few minutes; see several sources in one widget.
3. Watch the import date, not the review date. Ask when anything was last imported. If that date is months old while the platform is visibly still publishing, the connection is broken regardless of how healthy the page looks.
4. Own a channel nobody can close. Reviews collected through your own request form belong to you outright. 91 of our accounts hold 2,580 reviews collected this way, and no API decision can touch them.
5. Keep an export. If you are moving from another tool or want an offline copy, importing reviews from CSV works in both directions and costs nothing to keep as a backup.

Where we stand on this

We store every review we import, permanently, and serve widgets from that copy rather than from a live call to the platform. That single design decision is why 337,350 Amazon product reviews and 1,952 Yelp reviews are still doing work for the businesses that earned them, years after the routes that collected them closed. It is also why a source outage on our side degrades into a stale list rather than an empty page.

Two caveats, because they are real. A stored copy does not make old reviews look fresh, and we will not pretend a 2024 archive is current social proof. And storage does not help with reviews you never collected: the window between a source closing and you connecting it is permanently lost. Connect early, and connect more than one.

If you want to see which sources are currently alive before deciding what to connect, we published the full measurement in which review sources still work in 2026. If you already know, you can create an account and connect your first source in a few minutes, on a plan that starts free.