Everyone answers where do I find customer problems with a list of websites. I think that is the wrong shape of answer.

The thing that finally made my own research usable was switching from a destination to a property.

The property: somebody described the problem, in their own words, somewhere you can point at. Drop the pointer and you have a claim. Drop the own words and you have a summary of a claim. Drop the somebody and you have your own hypothesis wearing a citation.

The rule I now enforce on myself is that evidence never gets promoted. A search result summary is proof that a page exists and says something — it is not proof of what the page says. So anything resting on a snippet stays a snippet, however convincing the summary reads.

That sounds pedantic until you notice what it prevents. Three weeks later, an unsourced hunch and a sourced finding look identical in your notes, and the hunch is usually the one you liked.

On the run I publish in full: 41 sources consulted, 24 pages opened and read, 56 problems recorded, 32 of them carrying a link you can open. The other 24 are labelled unverified rather than deleted.

Two things about that which cut against me, and I would rather say them than have them found.

I obey robots rules, so a site that says no is skipped and whatever it knew is missing from the run. That is a genuine hole in coverage, not a rounding error.

And the best evidence is usually the least reachable. Long threads behind logins, support forums that render client-side, private communities — where people are most candid is where a fetcher cannot go. Anybody claiming complete coverage of customer complaints is describing a product that does not exist.

For doing it by hand, four habits that cost nothing: keep the URL rather than the summary, record the date, mark what you inferred, and write down what you could not reach. A search that returned nothing is a finding.

The tiers and the run: https://whittleos.com/guides/where-to-find-customer-problems