Run notes

What each run turned up

Every month the index runs the same questions and produces a new snapshot. These are the readouts of those runs. Claude writes them from the data, I read them before they go up, and the numbers in them come from the index rather than from the model. They are here because the record should be complete, including the calls that later turned out to be wrong.

What I make of the data is separate, and it is in Analysis.

• 3 August 2026

  Two apps moved in three months. The rest just wobbled.

  I track about a hundred apps across five categories. Three snapshots in, two of them have actually moved. Everything else went up, came back down, and landed where it started, including the changes I called real last month. Eight of the thirteen I was confident about reversed within a month.

  READ →

• 13 July 2026

  Hevy and Strong gained on the leader. Fitbod's gain didn't hold up.

  Fitness had no favorite in the first snapshot. The second shows the fragmentation has a direction: Hevy and Strong both gained on Nike Training Club across the largest-sample engines, not just the small AI Overview capture. Fitbod's score moved by the same amount, but the four big engines were flat for it. The gain traces to one extra AI Overview mention.

  READ →

• 13 July 2026

  How much of a month's movement is real?

  Every category moved between June and July. I checked which swings held up across the larger-sample engines and which rode on a six-question capture, including a case where I got the read wrong the first time and corrected it. A working rule for telling a real shift from a noisy one, not a proof that the rule is right.

  READ →

• 3 June 2026 · updated 1 July 2026

  I asked the assistants for the best AI app. They mostly didn't pick themselves.

  A self-preference test, checked twice a month apart. In June, Google's AI Overview led its own 'best AI app' answer with Gemini and left ChatGPT out entirely. In July the same query leads with ChatGPT and Claude, and Gemini is named in only half. The one bias I found in the first snapshot didn't survive a second month.

  READ →

• 3 June 2026 · updated 1 July 2026

  On the busiest shelf, the model and the surface agreed within a month

  Photo and video editing names more apps than any category I track. In June the chat models led with CapCut while Google's AI Overview crowned Splice. In July I ran it again: the Overview now leads with CapCut too, and Splice's mentions fell to zero. A surface-versus-model split that looked stable closed in a single cycle.

  READ →

• 3 June 2026

  Duolingo owns language learning. The AI tutor isn't Duolingo.

  Duolingo is named in more than four in ten grounded answers and leads language learning by a wide margin. Then I change one word. Ask for the best AI app to learn a language and Duolingo vanishes: the assistants name Langua and Speak, newcomers most people haven't heard of. And asked which AI tutor is best, the assistant refuses to name itself.

  READ →

• 3 June 2026

  The demand the app store can't see

  I tried to weight these questions by how often people search them, and the app store's own data said almost nobody does. Everything floored except 'best VPN'. It was looking in the wrong place. The 'best X' questions where AI makes its recommendations are invisible to app-store analytics, and that gap is the reason this index exists.

  READ →

• 3 June 2026

  When the grounded answer just echoes the affiliate listicle

  'Best VPN' is the most affiliate-gamed query on the web, so grounding every assistant is a stress test. Mostly they repeat the paid consensus, Nord, Proton, Surfshark, the same names users distrust enough to append 'reddit' to their searches. But ask for the most private VPN and Mullvad, which pays no commissions, moves to the front.

  READ →

• 3 June 2026

  The first question in AI visibility is whether the model can see the web

  Whether an engine retrieves live or answers from training data moves the recommendation more than the model or the prompt does. I ran one model both ways: without search it over-listed, recommended a dead app 13 times in 15, and favored its own product; with search on, all of that went away. If you optimise for AI answers, check this variable first.

  READ →

• 3 June 2026

  Asking more didn't change the answer

  AI answers vary between runs, so the index asks each question once and lets the count of questions do the averaging. I ran fitness three times to check that was safe. The combined leaderboard held: the same five apps led on every run, with only the order at the top and the tail below them moving.

  READ →

• 3 June 2026

  AI has no favorite workout app

  The leader is named in only about a third of grounded answers, and the rest of the field sits close behind. What a fragmented category looks like to an assistant.

  READ →