Field note
How often help centers drift from changelogs
An early look. 42 companies, one week, a sample that is small and not random. It is a first count and an upper limit, not a rate you should quote.
12 of 42
companies had a help article we could show was contradicted by their own changelog.
- 12still contradicted when checked against the live pages
- 1contradicted, then fixed within a day
- 29no contradiction we could prove
The question is simple. A SaaS company announces a change in its public changelog. Does its public help center keep up?
At most of the companies we read, it did. In 29 of 42 we could not prove a contradiction, and some of those help centers had been rewritten within days of the change. At one more, we found articles still using a feature's old name right after a rename. When we checked again a day later the docs had caught up, so we dropped it.
What the contradictions looked like
Renames are the obvious case, but they were not the common one. The most frequent shape was an article stating a limit, such as a plan restriction or a setting you can't change, after the changelog had announced the limit was gone. Our guess at why: a launch post prompts someone to document the new thing, and nobody goes back to the sentence saying it was impossible.
- Contradictions reported
- 16
- across 12 companies
- Per company
- 9 / 2 / 1
- companies with one / two / three
- A stated limit was lifted
- 8 of 12
- the article says you can't, the changelog says now you can
- A rename
- 2 of 12
- the article uses the old name
- Other behavior change
- 2 of 12
- a step or status that no longer holds
The three kinds are our own reading of our notes, with each company counted once.
How long they stayed wrong
We have both dates for 11 of the 12 companies. From the changelog entry to the day we wrote to them, the article had been wrong for 7 to 244 days, with a median of 51. That is a floor, since we don't know when each one gets fixed. Of the 11, 4 are older than the 90-day window our paid audit covers.
Method
- Where the data comes from. Two logs we kept between October 27 and November 2, 2026, while looking for companies to write to about this service. One records a verdict per company, the other what we sent.
- Who is in the 42. Small and mid-size B2B SaaS companies where we read a dated public changelog and a public help center, and recorded a verdict.
- What counts as a contradiction. A sentence in a help article that is false today because of a dated changelog entry. Both quotes are machine-checked word for word against the live page, and a second review has to agree. Ambiguous cases are dropped.
- What the 16 covers. Only the findings we sent. Weaker ones we left out are not counted.
- Who is not in the sample. About 260 other companies in the log got no verdict: no changelog where we looked, undated or stale entries, pages we couldn't fetch, help centers too large for this pass, or we didn't get to them.
Why not to lean on this
- It isn't random. We chose categories and companies where drift seemed likelier, left some categories out, and in some categories stopped looking after a hit.
- It only sees readable changelogs. A company without a dated public changelog can't appear here at all.
- A clean verdict is not a clean help center. These passes read a sample of articles, not all of them, so we will have missed some drift.
- The log has gaps. Some companies from our first night have no recorded verdict. One of them is in the 12 because the send log shows its finding. Our notes from that night say we read roughly 35 companies and most came back clean, so 12 of 42 is too high. Read it as an upper limit.
- One borderline call. We sent one further company a finding about two of its own pages disagreeing with each other. That isn't changelog drift, so it counts as no contradiction here.
- We sell the fix. Formship sells an audit for this problem. Read the count with that in mind.
We name no company here and quote none of their text. To have yours read, see the audit.