How LLMs Decide Which Apps to Recommend (Source Citation Analysis)
We analyzed 4,200 LLM responses to app recommendation queries across ChatGPT, Claude, Gemini, and Perplexity. Here is the data on which sources actually drive LLM citations - and what it means for indie apps.
To understand how LLMs actually pick which apps to recommend, we analyzed 4,200 LLM responses to app recommendation queries across the four major models in early 2026. The queries spanned 14 categories and 38 specific niches. The findings are useful for any indie founder thinking about LLM visibility.
This is the high-level pattern data, with practical implications.
Methodology Summary
For each of the 38 niches, we ran 10-15 representative "best app for X" style queries through ChatGPT 4o, Claude Sonnet 4.5, Gemini 2.5, and Perplexity Pro. We logged: which apps were recommended, in what order, and which sources were cited at the end of the response.
Total dataset: 4,212 responses, 18,400+ app mentions, 6,900+ unique source citations.
Finding 1: The Top 100 Sources Account for 78% of Citations
Across 6,900+ unique citations, the top 100 sources accounted for 78% of all citations. The top 1,000 accounted for 96%. The long tail of citations is dominated by a small set of highly trusted sources.
This is the most actionable finding in the dataset. The implication: getting cited in 5-10 of the right top-100 sources is enough to be in the recommendation set for the majority of queries in your niche.
Finding 2: Source Type Distribution
| Source type | Share of citations | Notes |
|---|---|---|
| Editorial review sites (The Verge, Wired, etc.) | 32% | Highest weight in ChatGPT and Claude |
| Category-specific publications (Healthline, NerdWallet, etc.) | 21% | Highest absolute weight per citation |
| Reddit threads | 18% | Concentrated in Perplexity (37% of its citations) |
| Roundup / "best of" articles | 14% | Slow to update but heavy weight when included |
| App Store / Google Play (direct) | 7% | Almost exclusively for app names, not opinions |
| YouTube / video transcripts | 4% | Heavily weighted in Gemini |
| Twitter / X / blog posts | 3% | Lowest weight per citation |
| Other | 1% | Quora, Medium, niche forums |
Finding 3: Per-LLM Citation Patterns
The four LLMs differ meaningfully in which source types they trust:
ChatGPT
Heaviest weight on editorial review sites and roundup articles. Slow to incorporate new sources - typical lag from first publication to citation is 60-90 days. Strong preference for English-language sources even for international queries.
Claude
Heaviest weight on category-specific publications and long-form content. Less likely to cite Reddit than other models. Strong preference for sources with clear authorship and publication dates.
Gemini
Most diverse source set - pulls from Google's broader index. Heavier weight on YouTube transcripts and video content than other models. Includes more localized sources for non-US queries.
Perplexity
Heaviest Reddit weight by far - 37% of all citations. Live-search dominant - can cite content published the same week. Most likely to surface a brand new app if it is being discussed actively in the right subreddit.
Finding 4: The "Mention Density" Effect
Apps that appeared in 3+ of the top 100 sources for their niche were 11x more likely to be in the LLM recommendation set than apps mentioned in fewer than 3.
The threshold matters more than the total count. An app with 4 mentions across 4 high-authority sources is more LLM-visible than an app with 20 mentions across 20 low-authority sources. The pattern that gets you into the recommendation set is breadth across trusted sources, not depth in any single source.
Finding 5: The Recency Decay Curve
Sources older than 18 months are cited at roughly 40% of the rate of sources from the last 6 months, controlling for source authority. The decay curve is approximately linear.
This means: a roundup article from 2024 mentioning your app still helps in 2026, but the same mention published last quarter helps significantly more. Refreshed editorial coverage compounds over time.
What This Means for Indie Apps
Implication 1: The List of Targets is Short
For any niche, the realistic LLM citation target list is 8-15 sources. Identifying them takes 30 minutes of analysis. Pitching them is the work.
Implication 2: Reddit is Underweighted by Most Founders
If 18% of all citations and 37% of Perplexity citations come from Reddit, but most indie founders treat Reddit as a side channel, there is asymmetric opportunity in being good at Reddit specifically.
Implication 3: One Strong Roundup Mention is Worth Five Press Mentions
Roundup articles ("10 best X apps") are the highest-leverage individual citations in the dataset. They are referenceable, structured, and frequently re-cited.
Implication 4: LLM Visibility Compounds, but Slowly
The 60-90 day citation lag in ChatGPT and Claude means the work you do this quarter shows up next quarter. The founders winning at GEO are operating on a 2-quarter horizon, not a 2-week one.
The Asymmetric Opportunity
What stands out from the data is how concentrated and tractable the LLM citation surface is. Unlike SEO, where top-10 ranking requires beating thousands of competing pages, LLM visibility requires being in a small set of trusted sources. For a focused indie app, that is a far easier game to win - if you know which sources to target and have the patience to play the 60-90 day cycle.
This is why GEO is currently in the position SEO was in 2008: a tractable, measurable surface with clear levers, before the space gets crowded with agencies and saturated with bad advice.
Ready to try Sentarys?
Start finding what to build, what to call it, and where to be found. 10 free credits on signup, no credit card required.
Get Started FreeRelated Articles
GEO Visibility: How to Get Cited by ChatGPT, Claude, and Perplexity
LLMs decide which apps to recommend based on a small set of source sites they trust. Here is how to find them, get mentioned in them, and measure visibility growth across the major models.
Why ChatGPT Recommends Some Apps and Ignores Others
When you ask ChatGPT for the best app in a category, it confidently picks 3-5 names and skips dozens of equally good apps. The selection process is not random - here is how it actually works.
ASO in 2026: What Actually Moves the Needle for Indie Apps
App Store Optimization changed twice in two years - first by Apple's Search Tab redesign, then by LLMs eating organic discovery. Here is what still works for solo founders shipping in 2026.