Analysis · 4 November 2026

Different models from the same AI (ChatGPT) recommend different apps.

Comparing gpt-5.4-mini to gpt-5.6-luna shows that Luna decides differently and cites different sites.

Data and method: 172 answers across the categories, plus a second run of the 35 fitness prompts. Plus the same fitness prompts with search off. All with gpt-5.4-mini which was validcube’s default ChatGPT model until October and with gpt-5.6-luna which is the new default from November onwards. Luna is also ChatGPT’s default model for free users on chatgpt.com and their app (as of October 2nd).

The big question in the analysis articles so far is about how AI Assistants search the web, how that influences the apps they mention and recommend (can it be influenced?), or if they answer from memory and only slap on citations as proof of their answer.

All five categories, 3 November 2026LunaMini
Answers that searched the web145 of 172113 of 172
Search queries per answer that searched4.82.3
Answers with more than one round of searching585
First-round search queries naming an app the question didn't name246 of 510 (48%)43 of 244 (18%)
Recommended apps that were already named in a first-round query241 of 336 (72%)75 of 250 (30%)
Queries limited to one website ("site:")355, in 94 answers21, in 9 answers
... of those, on the app maker's own website27917
Queries containing the word "official"498, in 131 answers166, in 90 answers

This table shows that the search behavior changed quite drastically between the two models from OpenAI. Luna uses search about 1.3 times as much as Mini, and searches more often per answer. In the November run, Luna had 58 answers with more than one search run, Mini had 5. One signal to find out if a model answers from memory, is to check if it uses the specific app name in its search query that wasn’t named in the prompt. Luna did this in 48% of its first search queries. Mini did this as well, but only in 18%.

This is different to Claude and Mistral who never named a brand in their first search query in this sample:

All five categories, 3 November 2026Answers that searched the webSearch queries per answer that searchedFirst-round queries naming an app the question didn't name
ChatGPT (Luna)145 of 1724.8246 of 510 (48%)
Claude (Haiku 4.5)123 of 1721.10 of 123 (0%)
Gemini (3.1 Flash-Lite)138 of 1723.3not measurable (no query order)
Mistral (Small)132 of 1721.10 of 132 (0%)

Owned pages and listicles up, press pages down

Share of cited links, all five categoriesLuna, 3 NovMini, 3 NovMini, 2 Oct
App maker’s own website56%46%52%
Press and test sites11%26%28%
Listicles (affiliate rankings, vendor “best apps” pages)13%9%4%
App stores8%15%10%
Government, research8%1%1%
Reddit, YouTube, forums0%0%0%
Other5%3%6%
Cited links (each link counted once per answer)461337338

With Luna, the share of citations from press pages drops from 26% to 11% while owned pages (brand/company websites) increased slightly from 46% to 56%. Overall, Luna retrieves 3x as many pages compared to Mini, but cites relatively less. In absolute terms it’s still 1.4x more. An interesting finding is that Luna retrieves a lot of results from Reddit, but never cites it in its answers.

Websites that gained and lost the most citations

WebsiteTypeMiniLunaLuna minus Mini
help.openai.comown site (OpenAI)1225+13
consumer.ftc.govofficial012+12
support.google.complatform help310+7
helpx.adobe.comown site (Adobe)512+7
help.fitbod.meown site (Fitbod)17+6
mullvad.netown site (Mullvad)28+6
itechguides.comlisticle27+5
apple.complatform38+5
pimsleur.comown site (Pimsleur)04+4
duolingo.comown site (Duolingo)50−5
healthline.compress50−5
techcrunch.compress50−5
androidauthority.compress60−6
tomsguide.compress2311−12
apps.apple.comapp store4728−19
techradar.compress3313−20
Cited links per website, all five categories, both models on 3 November 2026 (172 answers each, each link counted once per answer). The websites with the largest gains and losses.

In August, the analysis piece titled “ChatGPT loves to cite from press sites, the App Store and owned pages” and named the App Store, tomsguide.com and healthline.com the top 3 sites. healthline wasn’t cited once this time by Luna, tomsguide lost 12 and the App Store lost 19 citations compared to the Mini run on the same day. However, this decline can’t be attributed to the model switch in November alone, because both healthline and tomsguide already fell in September (on Mini, in fitness). On Luna, the App Store is the most cited URL in the run, but the distance is shrinking across all five categories. In the fitness category, where Fitbod is the most mentioned app, Fitbod’s own pages now beat the App Store listings (11 vs. 9).

Apple App Store, app listingsApple App Store, app listings, June: 28Apple App Store, app listings, July: 34Apple App Store, app listings, August: 27Apple App Store, app listings, September: 32Apple App Store, app listings, October: 21Apple App Store, app listings, November (Luna): 9151tomsguide.comtomsguide.com, June: 20tomsguide.com, July: 12tomsguide.com, August: 17tomsguide.com, September: 3tomsguide.com, October: 4tomsguide.com, November (Luna): 359Hevy's own pagesHevy's own pages, June: 8Hevy's own pages, July: 6Hevy's own pages, August: 9Hevy's own pages, September: 9Hevy's own pages, October: 7Hevy's own pages, November (Luna): 948healthline.comhealthline.com, June: 10healthline.com, July: 13healthline.com, August: 12healthline.com, September: 5healthline.com, October: 343Peloton's own pagesPeloton's own pages, June: 5Peloton's own pages, July: 6Peloton's own pages, August: 6Peloton's own pages, September: 11Peloton's own pages, October: 8Peloton's own pages, November (Luna): 541Fitbod's own pagesFitbod's own pages, June: 2Fitbod's own pages, July: 2Fitbod's own pages, August: 7Fitbod's own pages, September: 11Fitbod's own pages, October: 5Fitbod's own pages, November (Luna): 1138Google Play, app listingsGoogle Play, app listings, June: 13Google Play, app listings, July: 4Google Play, app listings, September: 10Google Play, app listings, October: 3Google Play, app listings, November (Luna): 535techradar.comtechradar.com, June: 4techradar.com, July: 10techradar.com, August: 5techradar.com, September: 1techradar.com, October: 13techradar.com, November (Luna): 235Apple's own Fitness+ pagesApple's own Fitness+ pages, June: 1Apple's own Fitness+ pages, July: 5Apple's own Fitness+ pages, August: 4Apple's own Fitness+ pages, September: 10Apple's own Fitness+ pages, October: 6Apple's own Fitness+ pages, November (Luna): 531nike.comnike.com, June: 1nike.com, July: 1nike.com, August: 3nike.com, September: 7nike.com, October: 4nike.com, November (Luna): 824Freeletics' own pagesFreeletics' own pages, June: 1Freeletics' own pages, July: 4Freeletics' own pages, August: 3Freeletics' own pages, September: 3Freeletics' own pages, October: 5Freeletics' own pages, November (Luna): 218Apple App Store, category chartsApple App Store, category charts, June: 7Apple App Store, category charts, September: 5Apple App Store, category charts, October: 517Apple App Store, stories and roomsApple App Store, stories and rooms, June: 4Apple App Store, stories and rooms, July: 1Apple App Store, stories and rooms, August: 2Apple App Store, stories and rooms, September: 5Apple App Store, stories and rooms, October: 3Apple App Store, stories and rooms, November (Luna): 116jefit.comjefit.com, June: 2jefit.com, July: 4jefit.com, August: 1jefit.com, September: 6jefit.com, October: 1jefit.com, November (Luna): 115forbes.comforbes.com, June: 1forbes.com, July: 2forbes.com, October: 1114the other 62 sourcesthe other 62 sources, June: 6the other 62 sources, July: 13the other 62 sources, August: 25the other 62 sources, September: 31the other 62 sources, October: 27the other 62 sources, November (Luna): 54156
JuneJulyAugustSeptemberOctoberNovember (Luna)
Citations by source in the ChatGPT column of the fitness index, counted as every link in the answer text, 35 answers per month. June to October gpt-5.4-mini, November gpt-5.6-luna. Links per month: 113, 117, 121, 149, 126, 115. Apple and Google Play are broken out by the kind of page cited; a vendor's own pages are counted across all of its hosts.

Memory still dominates app mentions

Switching search off and on made no clear difference in what apps got mentioned (for Luna, in fitness). The variation is about the same when just asking the same question multiple times. 85% of the apps that Luna mentioned with search on also appeared when search was off.

If you want to dive deeper into Luna-Land, read these 20 Findings written up by AI.

The numbers here come from the index. Every category has its own leaderboard and full table, and the method is published in full.