How change detection works, and what it cannot tell you
Change detection compares two stored copies of the same page and reports the difference. Most alerts say a page changed. A useful one says what changed, how much it matters and how sure the detection is, which takes a few deliberate design choices.
Snapshots give you a before and an after
Each check stores the page's text with a date. A change is the difference between the new snapshot and the previous one: a price that moved, a line that was removed, a headline that was rewritten. Because both snapshots are kept, you can open the before and the after and read them yourself.
The first check only sets a baseline
Everything on a page is new the first time you look. A baseline records it without raising anything, so changes appear from the second check onward.
Importance and confidence answer different questions
Importance is how much a change matters. A 20 percent price rise on Meridian's Team plan matters more than a reworded footer. Confidence is how sure the detection is. A direct before and after on the vendor's own page is high. A forum mention of a rumoured change is lower. Keeping the two apart stops a confident detection of a trivial change from looking urgent.
| Detection | Importance | Confidence |
|---|---|---|
| Meridian Team plan: $49 to $59 per seat, same page, two snapshots | High | High |
| Meridian footer wording changed | Low | High |
| Forum post says Meridian is dropping its free tier | High | Low |
What change detection cannot tell you
- Why the competitor made the change.
- Whether it will affect your pipeline.
- Anything on a page that needs a login, or a script to render, to display its content.
- A change made and reverted between two checks.
Keep a person between a detection and an action
A detected change is a prompt to look. Triage it, promote it to a draft insight, approve that insight, and only then create an action. Each step is a place to stop and decide that the change does not matter.
How Sigblaze handles it
Sigblaze compares each check with the last stored snapshot and raises a Signal when the difference is meaningful. The comparison is deterministic, and a model may reword a signal but cannot add facts to it. Checks run daily, weekly or on demand.
This piece contains no external statistics. Fieldnote and Meridian are fictional companies, and the figures in the examples are synthetic.