What changed
Product and data changes, newest first. Corrections to the data get listed here too — if a scoring change moves what you saw last week, you should be able to find out why.
Build memos, watchlists, trend lines, app pages, a yearly plan
Pro can have a build memo written for any opportunity: verdict, wedge, MVP spec, pricing and acquisition, every claim cited to a review. Any account can watch an opportunity and get its changes in the Monday email. Detail pages show a 90-day trend line, similar opportunities and the solution type, which is also a filter. New per-app pages. Categories collapsed from 544 spellings to 92 real markets. Pro is also $199 a year, and a single opportunity can be bought in full for $49.
Named incumbents, chosen quotes, and the top three free in full
Every opportunity now names the apps its reviews were written about, with store links. Quotes are picked for how typical and how intense they are, one per app, instead of by an engagement count that was zero for most reviews. The three highest-scoring opportunities open in full without an account. Pro gained search. New category pages and a blog of data pieces.
Three new sources, and a nine-day gap closed
Product communities on Discourse, Firefox add-on reviews and WordPress.org plugin reviews now feed the corpus. Collection had silently stopped on 23 August when the nightly clustering job outgrew its time budget and blocked everything else; it now runs on its own thread and a health check reports the pipeline’s state every half hour.
Badges rewritten so they mean something
“Broken incumbent” was appearing on 96% of clusters and “void” on none — an app review always names the app and complains about it, so the old tests were true of almost everything. Both now key off how dominant the top named product is, and a new Fragmented badge marks markets where several products are named and none leads.
Weekly digest, and pagination
Subscribers get the top opportunities by email on Monday mornings, and free accounts get the top three. The opportunity list is paginated rather than one very long page.
Rebuilt the interface
New typography and colour throughout, and a home page that shows the corpus resolving from scattered complaints into clusters — which is literally what the pipeline does to them. The method, the source mix and the scoring are now published rather than implied.
Scores no longer reward single complaints
Three of the four scoring inputs are read from the members of a cluster, so a cluster with one member was scoring on one opinion of one post. Scores are now scaled down when the evidence is thin: the top single-mention score fell from 72.6 to 46.8, and the top twenty no longer contains any.
Clustering fixed
The similarity threshold was set so tight that 96% of clusters had exactly one member — the grouping was barely happening. Measured across every pair in the corpus and reset to 0.73, which is the point where members are genuinely the same complaint. Opportunities with repeat evidence went from 108 to several hundred.
Extraction stopped losing half its work
Roughly half of all extraction batches were being discarded because the language model wrapped valid output in a code fence. Fixed; throughput went to 100% of attempted batches.
First version
Nine collectors, LLM extraction, embedding, clustering and scoring, running continuously.