Analysis · 6 August 2026
ChatGPT loves to cite from press sites, the App Store and owned pages
Almost all cited websites fall into three buckets: press, app store and owned pages. Why an app appears in the answer might be a separate question from why a URL gets cited.
The number 1 mentioned mobile app in the Fitness category is Nike Training Club. It’s consistently at the top for the last three months. In the Apple App Store Health and Fitness category*, Nike Training Club is on rank 104 (US, iPhone).
The number 1 in the App Store category is Strava. In my August index Strava fell 5 places to rank 11 with a score of 11, compared to 38 for Nike Training Club.
What makes ChatGPT recommend Nike Training Club above all? What drives mentions and recommendations from AI assistants in general?
To be transparent: My set of 35 prompts doesn’t cover the breadth of “health and fitness” and is rather biased towards workout apps. The development of the following hypotheses is general but the hook is the health and fitness category because the discrepancy is so clear.
The methodology
To understand what drives recommendations by AI assistants I looked at the citations in the responses to all 35 prompts. I focused on ChatGPT. The data comes from the three existing runs in June, July and August and from one extra run I did not using the API but a logged-out ChatGPT browser session.
I argued previously that grounding (web-search enabled when prompting the AI assistant) matters a lot, because otherwise it answers only from training data and what’s the purpose in asking the same questions regularly when literally nothing happens between the runs (link)?
Interesting finding: Using ChatGPT as a logged-out user on web meant using ChatGPT without web-search in 16 of 35 answers (46%). The model changed silently from gpt-5-5 to 5-5-mini throughout the run. gpt-5-5 used web-search on 1 of 10 answers, the mini model web-searched on 18 of 25 answers.
Almost all cited websites fall into 3 buckets: press, app store and owned sites
This is true for the citations from responses via the API at least. There’s variation in month over month and in this category the owned pages increase while the app store sites decreased as cited sources. Across the three months together the split is press 36%, App Store 34%, owned pages 29%.
The picture looks different to the responses I got using ChatGPT on web:
Reddit and Youtube take 18% of the citations and vendor listicles (comparison articles) are 10%. That is not an artifact of my Most-aware questions: 8 of the 11 Reddit and Youtube citations sit on discovery questions. The bucket that really is Most-aware driven here is owned pages, at 8 of 15. Interestingly, app store citations are much less on web. I’ll leave that here and will pick this up again with a greater sample size next time.
The same sites get cited over and over again
The most cited domain is apps.apple.com, which points to the app store listings. More on that below.
Other than Apple, Tom’s Guide appears to be ChatGPT’s favorite website, followed by healthline.com, followed by owned pages and other press sites (notably techradar.com).
170 different pages were cited, 114 of them just once, and only about 15% of the pages appear in multiple months. One page is the exception: Tom’s Guide’s “best workout apps” appeared 30 times. The second most cited page appears 11 times.
Hypothesis: The content on owned websites is a favorable source for citations
Writing answers to questions users ask their AI assistants on owned websites seems to matter for citations:
- Example 1: question “best app for beginners”, cited page from Nike “nike.com/help/a/ntc-info”, part of the response of the AI is “Nike’s own help page specifically says it’s ‘a great way to learn new movements’ for beginners”, almost word-for-word what the page says.
- Example 2: question “best lifting tracker”, cited pages from Hevy (hevyapp.com/features), Strong (help.strongapp.io/what-is-strong) and Fitbod (fitbod.zendesk.com/Feature-Overview) where all the key words of the answers appear.
| question | cited owned page | what the answer says |
|---|---|---|
| best app for beginners | nike.com/help/a/ntc-info | ”Nike’s own help page specifically says it’s ‘a great way to learn new movements’ for beginners” |
| best lifting tracker | hevyapp.com/features (3x) | “easy workout entry, solid stats, custom routines” |
| best lifting tracker | help.strongapp.io/what-is-strong | ”designed to be intuitive and straightforward for tracking” |
| best lifting tracker | fitbod.zendesk.com/Feature-Overview | ”focuses on personalized workout recommendations” |
Out of the total of 101 owned citations 84 (83%) are out of the “Most-aware” prompt group which already mentions different apps specifically (like “What’s a good free alternative to Peloton?”; “Peloton” in this case would not count as an app mention).
Hypothesis (the data says no): the app store description is not what gets you cited
The average app description in my set covers 5 of my 11 specific questions (“for beginners”, “no equipment”, “free”). That is the same whether ChatGPT cited the listing or never touched it.
The description gives the AI the sentence once the app has been picked. Nine answers quote a listing, and each quote is one clean claim about one attribute. Example: FitOn’s “Fitness is always 100% free” sentence was pulled into the response twice. And it sits deep into the app description, so depth on the page did not stop it from being read.
Hypothesis: Apple’s pages get cited, Google Play’s don’t
Apple is far more cited than Google: 103 citations (29%) vs 17 citations (5%) on the API column. Google Play has Editors’ Choice and curated collections, but no editorial, chart or Editors’ Choice URL was cited in the three months, while Apple’s were cited 14 times.
Why an app appears in the AI response might be a separate question from why a URL gets pasted after the app’s name. Does ChatGPT answer “Hevy” and then search for the best source about Hevy? Does it pick the app store page because it’s Hevy’s page, not because of what it says?
I’m not sure what’s cause and effect. It might be that I am overstating the value and impact of web-search and that the AI assistants are mainly or fully answering from knowledge. Web search might just be used to confirm their response (or simply because the user told the tool to use it).
- In the logged-out run 16 of the 35 answers used no web search, and 13 of the 14 apps named across the whole run appear in at least one of those unsearched answers. Caliber is the only app that shows up exclusively in a searched answer, once.
- Across 105 grounded API answers, 12% of the mentions belong to apps that never appear without a citation. Two of the three apps behind that number, Centr and Sweat, turn up in the logged-out run in answers that ran no search at all, so they are in the model’s memory as well. That leaves Ladder, 2 of 130 mentions, 1.5%.
Hypothesis: A citation supports an app mention that would have happened anyway
If that is the case, I put too much emphasis (and money) on grounding and high-frequency (i.e. daily) monitoring of app mentions without model changes is even more pointless.
An exception I see in the data:
Unknown and lesser-known apps might make the answer because they carry the answer in their title and subtitle
The question “Best app for home workouts with no equipment” brought 13 store citations, six unknown apps, and all of them have the answer in their title or subtitle field:
| listing title | subtitle | how the answer introduces it |
|---|---|---|
| Home Workout - No Equipments | Bodyweight Fitness & Training | ”Best overall: Home Workout - No Equipments” |
| 7 Minute Workout | HIIT Bodyweight Home Workouts | ”best if you want very short, fast sessions” |
| Home Workout No Equipment. | Bodyweight Training for Men | ”good if you want more personalization” |
| No Equipment Exercise - Home Workouts | Get Fit, Exercise anywhere | ”simple, beginner-friendly, and available on iPhone” |
| HomeFit - 7 Min Workout Plan | Stretch Exercise No Equipment | ”good for quick structured plans” |
| Seven: 7 Minute Workout | Daily HIIT Bodyweight Exercise | listed among the alternatives |
Similar patterns show the results for “Best AI personal trainer app” (“Saga: AI Personal Trainer”, “Iridium - AI Workout & Macros”) or “Best app to log my gym workouts” (StrengthLog / “Workout Log & Programs”, OverLog / “Progressive Overload & Gym Log”, Steady: Gym Log & Planner / “Workout Tracker for Progress”).
Those keyword-stuffed app titles are also why an app ranks in App Store search. I can’t tell whether the wording got the app picked or whether ChatGPT just took what ranked well.
*Source: https://apps.apple.com/us/iphone/charts/6013, data from 2026-08-03.