Insights and live numbers

Updated 2026-11-08 09:17 (UTC+7).

Live numbers

Focus plan, highest return only

1. 1. Judge the 07:36 releases at about 15:00. Re-rank the 282 phrases and the 41 new-listing phrases 7 hours after the release. Keep the names that climbed. Do not release the same listing again before then.
2. 2. Owner steps, in this order. The Atlassian Partner tabs and security contact, a new Forge token before 12 November, the publisher domain zovo.one (one DNS TXT record), the Apify ID check, a PayPal.me link, the Obsidian submission, a new Claude sign-in before 13 November, the Stripe Identity check on ECOHELP-182590 before 6 December, and a new Modrinth token.
3. 3. Ratings on the listings that earn installs. Count ratings each day on DOCX, 3D, NetCDF, tensor, table, PPTX and PDF. Rating ask v2 goes to the other listings in their next release.
4. 4. SVG viewer, then hex viewer (GO). The 07:30 re-measure said GO. Build and gate them, then publish after the new-extension cap frees at about 12:17.
5. 5. DOCX Viewer for the web. Ship 0.3.1 with VS Code for the Web support at 14:36 or later, after a fresh real-app check. Only 4 of 36 DOCX rivals work on the web.
6. Not doing now. No new stores and no paid promotion. Each one adds cost before the first dollar arrives.

What we learned

1. Rank wins, not product count. On 8 November at 07:30, 29 live VS Code listings had 102 installs and the factory had $0 revenue (F, public gallery). The 7 listings added at 22:57 brought 1 install. The 22 older listings brought 46 overnight with no release. DOCX Viewer has 28 of the 102 installs from its top-3 rank on "word viewer" and "docx preview". Make each kept product the best on its platform for its main phrase before building new ones.
2. Wave11: installs grew 85 % overnight with no release. Installs went from 55 at 19:01 on 7 November to 102 at 07:30 on 8 November (F). No listing had a release in that time, because of the 7-hour rule. Top-3 phrases held at 103 of 282 (102 at 18:16). So the gain came from the ranks that the 7 November renames won, and from the US day. Batch-1 listings made 5.9 installs a day each, batch 2 made 1.7 and batch 3 made 2.6 (F, re-measure).
3. Wave11: more listings in small niches did not bring installs. The 7 extensions that went live at 22:57 (G-code, EPUB, ZIP, Font, HAR, Audio and the viewer pack) are top 3 on 4 of 41 phrases and have 1 install after 8 hours (F, RANK9). Their ranks did not change from 1 hour to 7 hours after release. A phrase ranks well when each of its words is in the display name, in any order: median rank 6 when all words are in the name, median rank 39 when one word is missing.
4. Wave11: the install-earning products got their biggest updates. DOCX Viewer 0.3.0 (equations, charts, a 47 MB report in about 1.5 s), PPTX 0.4.0 (a 197 MB deck in under 5 s, SmartArt, slide to PNG), NetCDF 0.4.0 (0.25-degree grids and a 472 MB file now plot, coastlines), Video 0.2.5 (sound without ffmpeg), Tensor 0.4.0 (.pt and .pth files with a safe reader), plus fixes in SQLite, Table and PDF. 9 names gained a missing search word, for example "PPT Viewer", "MKV" and "SQLite DB Viewer". B-REL published 29 of 29 staged updates to the VS Code Marketplace from 07:36 to 08:17, one per listing, with 0 failed runs. The gallery showed all 29 new versions at 08:18 (F). Open VSX had the same version for 21 of 29 at 08:18, and a follower script publishes the rest.
5. Wave11: a second check in real VS Code found 8 defects after all tests passed. Two red-team lanes ran 9 staged builds in real VS Code against the gallery builds, on more than 113 real files each. They found and fixed a PPTX zip bomb (webview memory 10 GB down to 146 MB), a tensor pickle loop, 9 unsafe pickle names, a text .pth regression, a SQLite view that never ended, a rating ask after an empty file, Windows-1252 CSV text, and 2 walkthroughs with text that failed the style check. 74 malicious pickle files ran 0 payloads.
6. Wave11: Open VSX opened, and the money gates did not. Eclipse closed issue 15379 as granted at 03:54, and the zovo namespace is verified with 29 extensions (F). Cursor, Windsurf and VSCodium users can now find them. Revenue is still $0. Atlassian added an owner step at 03:36: a Stripe Identity check on ECOHELP-182590 by 6 December. JetBrains (4 plugins), Apify, Modrinth and Obsidian did not change.
7. Wave11: ratings are still 0 on 29 listings. Crowded phrases (pdf viewer, sqlite viewer, excel viewer, claude code usage) belong to listings with 100,000 or more installs and many ratings. Rating ask v2 (3 good opens on 2 days, at most 2 asks, never after an error) is in the 8 updated builds. The #1 SQLite and Excel viewers get 1-star reviews about paywalls, so our listings lead with free, no ads and no account (COMP1).
8. Wave10: a name that starts with the search phrase is the fastest lever. From 15:00 to 18:16 on 7 November, top-3 phrases went from 70 to 109 of 290 tracked phrases (55 at 07:35). Top-10 phrases went from 139 to 176. 13 releases after 15:00 (12 of them name changes) gave 29 new top-3 phrases in 1 hour or less, with 1 loss. Example: the JSON viewer went from 0 to 5 top-3 phrases, and "json tree viewer" went from #37 to #3.
9. Wave10: installs followed the ranks. Installs went from 41 at 15:02 to 54 at 18:17 and 55 at 19:01 (F). That is 4.0 an hour, against 0.47 an hour from 10:45 to 15:02. 11 of the 13 new installs went to listings with a release that afternoon, for example the first 2 installs ever on the video player after "mov" reached the top 3. The time of day is a confounder, because the US morning starts then. The 8 November 07:30 re-measure is the test.
10. Wave10: wait 7 hours before you judge a change. Some changes reach search rank in minutes. Others take 4 to 7 hours. 11 new top-3 phrases came on 4 listings with no release after 13:21 local time, for example "claude code history" from #56 to #1 on the session viewer and "dxf viewer" from #6 to #1. The 15:00 conclusion that keyword-only changes do not move rank was wrong. Do not release the same listing again within 7 hours.
11. Wave10: crowded phrases do not move with a name. The SQLite viewer got "Database Viewer" in its name and went from #13 to #4 on "database viewer", but "sqlite viewer" stayed at #10 and "db file" fell from #3 to #5. The PPTX viewer got no new top-3 phrase. "pdf viewer" went from #19 to #12, not top 3. "claude code usage" stays at #79, because #1 has 77,519 installs. Here, installs and ratings decide, and we have 0 ratings on all 22 listings.
12. Wave10: 4 new products for 2 new stores, 0 public yet. Agents built and tested 3 JetBrains plugins (AI Usage Meter 35877, DOCX & PPTX Viewer 35878, NetCDF, HDF5 & NumPy Viewer 35879) and 1 Obsidian plugin (Office Viewer). All passed a test in the real app. All 4 JetBrains plugins, Context Drift included, are in moderation with 0 downloads (F, 19:01). Obsidian now takes new plugins only through community.obsidian.md, so the owner must sign in and submit. The install estimates (40 to 530 in 30 days each) are E until the stores show numbers.
13. Wave10: the real-app check still finds defects after every unit test passes. Fresh lanes ran the new builds from the exact zips. They found 1 High and 3 Medium defects in the Obsidian Office Viewer, 8 defects (2 Medium) in the 2 JetBrains viewers, and 2 broken README links in the PDF viewer. All were fixed. The JetBrains fixes wait until moderation ends.
14. Wave10: no money gate moved. Revenue is still $0. PayPal is still empty in funding.json. Apify waits for the owner ID check. Atlassian waits for the owner answers on the Partner tabs. Open VSX issue 15379 has had 0 comments since 10:47 on 7 November. 5 free Atlassian apps got the prosemirror-view 1.42.9 security pin and pass 4 of 5 smoke tests.
15. 53 releases moved installs by 1 on the same day. Installs were 39 at 10:51 and 40 at 14:40, while 7 new extensions went live and 46 updates shipped (F, public gallery). The one new install is on DOCX Viewer. Rank moves first and installs follow later, so judge the releases on the 8 and 9 November re-measures, not today.
16. Search rank can move within hours of a release. The video player went from #22 to #8 on "video player" and from #4 to #1 on "mp4 viewer" within about 3 hours of its new version. The SQLite viewer went from #45 to #9 on "sqlite viewer". Some phrases need longer: claude code usage is still outside the top 100.
17. A clear name in a small niche ranks on day one. The 7 new extensions got names that start with the search phrase. On the first reading they are top 3 on 18 of 41 phrases, for example #1 on "tiff viewer", "obj viewer" and "mbox viewer". Graphviz and Avro are not, because older listings with many installs hold those phrases.
18. Small niches give top 3 on the first day. Big phrases do not. NetCDF Viewer is top 3 on all 4 of its phrases, and the table viewer reached #2 on "xlsx diff" and "excel compare" right after its release. The big phrases (csv viewer #12, excel viewer #15, claude code usage) belong to listings with thousands to millions of installs. A name alone does not pass them. Installs and ratings do.
19. A check on real data still finds bugs after every test passes. Fresh lanes ran every new version from the gallery on real data in real VS Code: the Claude Code, Codex and Gemini folders of this Mac, a 5.2 GB GGUF file and a 124 MB JSON file. They found 6 bugs that the unit tests and the release test missed, for example no weekly limit on Team plans and raw ffmpeg text on an empty video. All 6 were fixed and published the same hour. The last check found none.
20. Every product now points to the others. All 22 VS Code extensions, 9 Apify Actors and 43 site pages show GitHub Sponsors, the GitHub profile and 3 to 5 related Zovo products. Some lists still show the old names of 9 renamed extensions, and the links still work. The next rollout fixes the names. PayPal needs only a link from the owner. One command adds it everywhere.
21. Owner sign-ins are still the gate, and they move when agents prepare them. On 7 November an agent filled the Atlassian Partner Verification form from the company documents, and the owner sent it. Agents also sent the security questionnaire on all 10 paid tickets. JetBrains waited 70 minutes over 2 lanes for a sign-in. When the owner signed in, an agent uploaded Context Drift within the hour, and it is now in review. Open VSX has had no reply on issue 15379 since 10:47. The Apify company certificate is approved, and only the owner ID check is left.
22. Code review is not enough. Run the app. A browser test on 2026-11-06 found that 7 of 19 Atlassian apps crashed at start, although all 7 passed review with scores of 92 to 100. We fixed and redeployed them the same morning. Now no app can be marked ready without a runtime pass on its deployed version. 17 of 19 apps pass. Bulk Page Ops is fixed (version 1.1.1 reads labels through the v2 API, because Confluence removed the old one). Portal SLA and StatusBeacon need a Jira Service Management project to test.
23. Store search is where Apify users come from, and we are not in it. For 10 of our 11 main keywords, none of our Actors is in the top 100 Store results. The top 10 for "youtube transcript" have 15,238 users in 30 days. Free channels such as Discord and Reddit bring about 7 paying users in total.
24. The Atlassian apps hold most of the expected money. The base case is about $92,000 in 12 months for all 19 apps. The first 5 to list are AskSpace ($23,000), SignOff ($13,600), Markdown + Mermaid ($8,100), MathBlock ($7,900) and EpicTotals ($5,700). These are estimates from rival install rates, and the method is on the Money page.
25. VS Code is the only store with real pull. Context Drift has 225 downloads with no promotion (highest of 5 readings, because Marketplace servers return different counts). On 2026-11-06 we added 5 file-viewer extensions. They are Skill Shelf, Tensor Viewer, HDF5 NetCDF Explorer, DOCX Viewer and PPTX Slide Viewer. 8 extensions are now live. Each has a public repo and a landing page on this site.
26. The Apify numbers are almost all ours. Each Actor shows 2 users, which are our account and the Apify daily auto-test. All runs are our tests and fixes. Apify Store search takes days to rank new Actors, so a quiet first day is normal.
27. The 19 Atlassian apps are the largest stock of finished work. They are built, reviewed and deployed. They earn nothing until the owner adds partner payment and tax details and we submit paid Marketplace listings.
28. Owner steps are the bottleneck, not building. The sprints shipped 46 products. 4 owner steps block the money. They are the Open VSX Publisher Agreement, Apify business verification, the Atlassian Partner Agreement and verification, and Atlassian payment details. Free Atlassian listings do not need bank details.
29. Free on Atlassian is permanent. If an app moves from free to paid, existing installs lock until an admin subscribes. Apps that run in Jira and Confluence must keep licensing, so Issue Dates Calendar stays paid. 7 apps are now free builds, to collect installs and reviews first.
30. The VS Code Marketplace limits new extensions to 7 in 12 hours. Updates have no limit. The web upload form dropped uploads with no error, so for 5 hours no new extension went live. Trusted publishing through GitHub Actions fixed it on 2026-11-06. 7 new extensions went live in 15 minutes, and the next ones publish by script when a slot opens.
31. The upload path was the bottleneck, not the products. For 5 hours no new VS Code extension went live, because the web upload form dropped files with no error. A lane found trusted publishing through GitHub Actions. After that, 7 new extensions and 24 updates were published in 2 hours with 0 failures, and the account holds no token.
32. Installs come from search inside VS Code. DOCX Viewer got 11 installs from 4 web page views. The name decides the search rank, so lane A3 renamed 7 listings to start with the phrase people type, for example Claude Code History & Codex History Viewer.
33. A test in the real app finds what unit tests miss. Real VS Code runs found product bugs in DICOM, ZIP, Arrow and SMV builds, and a real Minecraft server run found that BucketBackup left out the main world data. All were fixed before release.
34. Modrinth is the one new store that needs no owner step. Agents made a token and submitted 2 mods. Async Locator 26 cut the longest server tick from 1,368 ms to 17 ms in a real server test.
35. Building is fast. Store limits set the pace. Agents build and test a new viewer in 25 to 40 minutes, including a test in a real VS Code window. The queue holds 8 finished extensions and 3 more are being built. At about 7 new extensions each 12 hours, the queue is full until 2026-11-08.
36. A test in real VS Code finds bugs that unit tests miss. The release test ran JSONL, SQLite and Agent Session viewers in a real VS Code window. It found and fixed 6 bugs. All unit tests had passed before.
37. Search rank comes from the display name. In Marketplace search, the exact search words in the display name decide the rank group. Inside a group, shorter names that start with the words win, then installs. Keywords only break ties. We renamed all 8 live extensions to match this, for example "CSV Viewer, Excel Viewer & Parquet Viewer".
38. Open VSX has real demand for the same viewers. On Open VSX (used by Cursor and Windsurf), one JSONL viewer has 24,478 downloads and one CSV extension has 361,835. We have 0 there. One script publishes all our extensions when the owner signs the Open VSX agreement.
39. Ratings help search, so each extension now asks one time. All 15 of our extensions now ask for a rating one time, after 5 good uses over at least 2 days. The prompt never shows in the first session, and the user can turn it off. The extensions send no telemetry.
40. The review loop worked. Each product went through independent reviews until it scored 92 to 100. Reviews caught an App Reviews run that could hang, a silent install failure in MathBlock, and a Secret Scanner setting that Atlassian does not allow.
41. First builds always failed their first review. The median first score was 56. One fix round added 15 to 45 points. A reviewer on a different seat finds what the builder misses.
42. Research prices were 3 to 4 times too high. We now price against the cheapest working rival, at $0.0001 per review, $0.001 per trends result and $0.0025 per transcript.
43. A real platform run is part of the gate now. All 4 first Actors failed the Apify empty-input auto-test, and no reviewer caught it. Every prefill field now needs a default.

Next steps

Numbers come from the Apify API and the VS Code Marketplace API. 48 snapshots are kept in factory/sprint/stats.json.