Run notes · 3 August 2026
Two apps moved in three months. The rest just wobbled.
I track about a hundred apps across five categories. Three monthly snapshots in, two of them have actually moved. Everything else went up, came back down, and landed roughly where it started. That includes the changes I told you were real last month.
Run note Written by Claude from that month's data, and read by me before it went up. The numbers come from the index, not from the model. My own reading of the data is in Analysis.
Last month I wrote a rule for telling a real change from a random one. If two or more of the four big engines shifted the same way, I called it real and wrote about it. If the change came from one engine, or mostly from Google's small six-question sample, I flagged it and moved on.
I couldn't test that rule at the time. Two months of data gives you one change, and one change can't tell you whether changes stick. August gave me a second one, and the second one grades the first.
The rule doesn't work
Thirty-six apps moved at least three points between June and July. By August, 22 of them had moved back the other way.
Sorted by what I'd called them:
Changes I called real: 8 of 13 came back. Changes I called noise: 10 of 16 came back. Same rate. The rule sorted nothing.
Two engines agreeing told me nothing that one engine alone didn't. Here is every change I called real in July, and what August did to it.
| App | Category | Jun → Jul → Aug | Jul move | Aug move |
|---|---|---|---|---|
| Picsart | Photo & video | 14 → 4 → 13 | -10 | +9 |
| Strong | Fitness | 20 → 27 → 21 | +7 | -6 |
| italki | Language | 19 → 13 → 14 | -6 | +1 |
| Microsoft Copilot | AI assistants | 19 → 14 → 19 | -5 | +5 |
| Hevy | Fitness | 26 → 31 → 26 | +5 | -5 |
| Photoshop Express | Photo & video | 4 → 9 → 9 | +5 | 0 |
| Premiere Pro | Photo & video | 10 → 15 → 11 | +5 | -4 |
| Perplexity | AI assistants | 33 → 29 → 32 | -4 | +3 |
| Caliber | Fitness | 15 → 19 → 23 | +4 | +4 |
| Canva | Photo & video | 14 → 18 → 12 | +4 | -6 |
| NordVPN | Security | 26 → 22 → 21 | -4 | -1 |
| Adobe Lightroom | Photo & video | 20 → 23 → 23 | +3 | 0 |
| 1Password | Security | 16 → 13 → 16 | -3 | +3 |
Look at the middle column. Microsoft Copilot went 19, 14, 19. 1Password went 16, 13, 16. Strong went 20, 27, 21. These apps did not rise and then fall. They sat still while the measurement bounced.
How much wobble is normal
Across all three months and all five categories, the typical month-to-month change is 1 point. Nine out of ten changes are 5 points or less.
So a 3-point move, which is where I set the bar for "worth looking at," is close to ordinary. I ask each engine about 35 questions. One extra mention shifts its score by roughly 3 points on its own. Two engines each happening to mention an app one more time was enough to trip my rule. That is not a trend. That is two coin flips.
The one I got wrong in public
In July I published Hevy and Strong gained on the leader. Both had moved up on two engines at once, which my rule said was the trustworthy kind of movement. In August both dropped back to 6 percent on ChatGPT, where they started, while Nike Training Club climbed from 17 to 26 and pulled further ahead.
The careful part of that piece was wrong too. I refused to credit Fitbod's gain because it came from Google's small sample. That caution was right by accident, not because the rule worked.
I've left the article up with a note on top instead of fixing the numbers quietly. The bad call is the useful part.
What actually moved
Two apps, out of roughly a hundred, went the same direction two months running:
- Caliber, fitness: 15 → 19 → 23. Claude and Mistral both about doubled how often they name it between June and July, and have kept naming it since. ChatGPT has never mentioned it once.
- Adobe Express, photo and video: 21 → 18 → 11. Google's AI Overview named it in three of six captures in June, two in July, one in August. Gemini dropped it from 32 percent to 15 in the last month.
That's it. Everything else is either steady or noise. The top app in every category is the same one it was in June: ChatGPT, Duolingo, CapCut, ProtonVPN, Nike Training Club. None of them has been overtaken.
What I'll do differently
I'm going to wait. A change counts when it goes the same way two months in a row. One month of movement, however many engines agree, is not enough.
This means I can't report anything the month it happens. I'd rather be a month late than keep describing coin flips as news.
It also makes the case for tracking this monthly at all. If I'd asked these questions once, in July, I'd have walked away certain that Hevy and Strong were on the rise. Three readings is what showed me they weren't. A single snapshot can't tell you it's wrong. This month, mine did.
Where this comes from
The June 3, July 1 and August 3 snapshots across all five indexes, using the stored per-engine counts behind every published score. No new data was collected for this piece. The rule being graded is the one in how much of a month's movement is real, applied to June-to-July exactly as it was at the time, then checked against August.