We asked AI assistants which GLP‑1 app to use, 1,410 times.

• Engines. OpenAI GPT-5.5 (gpt-5.5-2026-04-23) through the Responses API with web search; Anthropic Claude Opus 5.5 through the Messages API with its web search tool; Google Gemini 3.8 Flash through the Gemini API with Google Search grounding. These are the developer APIs, not the consumer apps.
• Questions. 110 questions, 55 written for the UK and 55 for the US. Per market: 21 asking for an app, 13 about a problem (such as muscle loss or a stalled scale), 10 comparisons, 3 follow-ups and 8 asking about our app by name. UK questions say “jab” and lead with Mounjaro and Wegovy; US questions say “shot” and lead with Zepbound, Ozempic and Wegovy.
• Markets. OpenAI and Anthropic searches were set to the UK or US. Gemini’s API has no location setting, so the market is in the wording only.
• Runs. Each question five times on each engine: 1,650 answers. The 240 answers to questions that name our app are left out of the shares, which leaves 1,410.
• Date. All answers were collected on 19 October 2026, between 20:25 and 21:09 UTC.
• Scoring. App names were matched in each answer’s text against a list of 37 GLP-1, weight and body-tracking apps (scorer version s1, question set v1 (draft)). When an answer used one of our app’s names or linked to it, and for every question asking about our app by name, a judge model decided whether our app was mentioned, recommended and described correctly. All 241 saved judgements (240 by-name answers and 1 other) use judge prompt version j2.
• Judge check. Before use, both j2 judges, Claude Opus 5.5 and GPT-5.5, were checked against 98 hand-labelled answers: 90 real answers from an earlier probe and 8 constructed controls written for cases the probe didn’t contain, such as an answer that recommends our app. On whether our app was mentioned, both judges matched the hand labels on 98 of 98. On whether it was recommended, Claude matched on 98 of 98 and GPT-5.5 on 97 of 98. The one disagreement was a constructed control that mentions our app without recommending it, which GPT-5.5 scored as a recommendation.

Some of the questions, word for word:

• “What’s the best app to track body changes on Mounjaro?”
• “Which app tracks my GLP-1 shots and progress photos together?”
• “How can I tell if I’m losing muscle on tirzepatide?”
• “Alternatives to Shotsy for GLP-1 tracking”
• “Which GLP-1 apps show you body shape changes, not just weight?”

The same list is in ai-app-recommendations-questions-2026-10-19.csv, with each question’s ID and the scoring versions.

• It isn’t a ranking of which app is best. It counts what assistants said.
• The developer APIs can answer differently from the ChatGPT, Claude and Gemini apps you use on your phone.
• It is one evening. Answers change from run to run and week to week, which is why we asked everything five times and report shares, not positions.
• Only the 37 apps on our list were counted. Others were sometimes named and are missing here.
• We wrote the questions, and we make one of the apps. The full question list is in the method above, and the counts are in the CSV.