Analysis
What the numbers show
The index produces the data. These are the findings worth reading, written up.
-
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
Give Mistral web search and it stops looking like the odd one out
The French engine looked insular and stale next to the US models. Then I noticed I'd run it without web search. With search on it converges: same top apps, no more dead products, and it stops naming its own Le Chat. The home bias was a memory artifact.
READ → -
3 June 2026
I asked the assistants for the best AI app. They mostly didn't pick themselves.
A self-preference test on a recursive category: four of the assistants are also apps on the board. The big chat models cross-recommend each other; the bias hides at the margins, in Le Chat and in Google’s AI Overviews.
READ → -
3 June 2026
AI has no favorite workout app
In budgeting the top app is named in 60% of answers. In fitness the leader barely clears 40%. What a fragmented category looks like to an assistant.
READ → -
3 June 2026
A budgeting app that shut down in 2024 is still recommended in 2026
Mint closed two years ago. The chat assistants still name it. Google AI Overviews are the exception. On grounding, staleness, and how much the assistants disagree.
READ →