Lovable Rewrote a Changelog Entry to Add AI Training

Last week I wrote that PostHog raised its entry AI price 5.8x in a single day, and closed with a note that a changelog looks like an append-only log and is not one. Five days later the same thing happened again, at a different company, with more at stake than two missing bullet points.

Same method as always: every quote below is a diff between two daily snapshots of a public page, captured at 06:00 UTC. Nothing here was announced.

What did Lovable change about AI model training?

On the morning of August 6, Lovable's changelog carried a new entry dated July 31, 2026. It read as an unambiguous privacy win. Title: Opt your personal data out of AI model training. First line, verbatim:

"Every user, on any plan, can now exclude their own personal data from AI model training."

The next morning, the same entry — same position, same July 31 date — had been rewritten. Title: Opt out of AI model training. First line, verbatim:

"Every user, on any plan, can now set their training-data choice in advance: starting September 9, 2026, Lovable may use customer data from Free and Pro plans for AI model training unless you opt out."

Here is what actually changed between the two captures:

• The title. "Opt your personal data out of AI model training" became "Opt out of AI model training". The words that scoped it to your own data are gone.
• What it announces. On August 6, a control you can use. On August 7, a default you are subject to.
• The default for Free and Pro. Not mentioned at all on August 6. On August 7, training is on from September 9 unless you opt out.
• Business and Enterprise. Excluded by default in both versions — unchanged.
• The date shown on the entry. Jul 31, 2026 in both.

The Business and Enterprise carve-out is identical in both versions. The new default lands on Free and Pro only — the two plans whose users are least likely to have a procurement process that re-reads a vendor's changelog.

So what: the two versions of this entry are not the same announcement with better wording. The first tells you about a control you can use. The second tells you about a default you are subject to, and it arrived inside a post dated a week before it was written. If you read Lovable's changelog on August 6 and considered yourself informed, you were — about a different policy than the one now in force.

Did the entry's date change when its meaning did?

No. As of publication the entry still shows July 31, 2026, still sits in the July 31 slot, and carries no correction note, no "edited" marker and no follow-up entry. The only record that the earlier version existed is a snapshot someone happened to keep.

This is the second changelog entry I have caught being rewritten after publication in two weeks. The first was Cursor's, which announced Google Docs and Google Sheets support and then quietly removed both from the entry.

So what: the useful lesson is not that either company acted in bad faith — a legal review landing after a changelog goes out is an ordinary way for this to happen. The lesson is structural. Changelogs are rendered from a mutable CMS and presented with the visual grammar of an immutable log: dated, ordered, append-only. If you cite one in a battlecard, a security review or a board deck, cite a copy you control, because the page will silently agree with whatever the company believes today.

Did Cursor put Google Docs and Sheets back?

I said I would check. No.

Cursor's changelog still carries the Google Workspace entry dated August 3, 2026, and it now names three integrations: Google Drive, Gmail and Google Calendar. Docs and Sheets do not appear in the entry, and they do not appear anywhere else on the page. The bullets that described them — "open and read documents, write and edit text, create new docs" and "read ranges and find values, update cells, create and edit spreadsheets" — are gone with no replacement.

Eight days after they were announced as shipped, two integrations have left the public record without ever being described as delayed, pulled or cancelled.

Why did Stripe's pricing page keep flipping?

Because it was almost certainly being tested on, and this is what an A/B test looks like when you sample it once a day.

Stripe's pricing page alternated between two designs in the same week — a guided discovery block ("Not sure where to start? Tell us about your business to get personalized Stripe product recommendations"), and a row of direct sign-up buttons, one per product:

• August 3. Discovery block removed, 20 "Sign up with…" buttons added.
• August 4. Discovery block back.
• August 6. Buttons back — 21 of them now, with Link added — plus an explicit "2.9% + $0.30".
• August 7. Buttons and the rate removed, discovery block back.
• August 8. Discovery block removed, buttons back.

Underneath the flapping, one change stuck: on August 6 "Access to 100+" became "Access to 125+" in both places it appears, and it stayed.

So what: this is the difference between a change detector and a competitive briefing. Five diffs fired here, and four of them are the same experiment breathing in and out. A tool that emails you on every diff would have sent five alerts about Stripe this week and buried the one fact worth keeping — the product count moved 25% — under four false alarms. This is exactly why the rule here is that a page that alternates between two states gets rate-limited rather than reported, and why the briefing is weekly. Some signals only resolve when you stop looking daily.

What else moved

beehiiv did not turn last week's rename into a price change — at least not yet. After "Paid Recommendations" became "Recommendations" across the comparison matrix on August 5, availability stayed exactly where it was: no on Launch, yes on Scale, Max and Enterprise. What did change is the matrix around it. On August 7, beehiiv added three feature rows and moved one gate:

• Click-triggered automations — Max and Enterprise only.
• Group subscriptions — Max and Enterprise only.
• Random cohort sends — Enterprise only.
• Send API — moved from no to yes on Max; still no on Launch and Scale.

Send API moving onto Max is the one to watch: an API that was previously an Enterprise conversation is now a self-serve tier feature.

The pattern worth naming

Three of the four companies above changed what their own published pages said about work they had already announced. Lovable rewrote the terms inside a dated entry. Cursor deleted two integrations from one. Stripe served two different pages on alternating days.

None of that is visible to anyone reading the page today. It is only visible if you kept yesterday's copy — which is the entire argument for treating competitor pages as a time series rather than a bookmark. The published record of a software company is a live document that presents itself as a historical one.

How this was made

Every quote and number above is a diff between two daily snapshots of a public page, captured at 06:00 UTC and compared to the previous morning's copy. No press releases, no tips, no interviews. Lovable did not announce that it had edited the July 31 entry. Cursor did not announce that Docs and Sheets came out. Stripe did not announce that it was testing two pricing pages.

One build-in-public note. The Cursor deletion that led last week's edition was scored noise by the automated relevance step — the human review caught it. Silent removals are the blind spot: a classifier trained to notice what was added is weakest exactly where the story is what disappeared. That is the next thing getting fixed here.

If checking competitor pages by hand is not happening at your company — and it usually is not — here are the practical ways to track competitor pricing changes, including the automated one.

ScrapRival watches your competitors' pages and sends one briefing every Monday. You can read a real sample digest without signing up, or start a 14-day free trial pointed at your own competitive set.

Next Monday: whether Lovable's September 9 date survives contact with its enterprise customers, and whether beehiiv's Send API move onto Max is followed by a price.