Research / Edition 2026-09
YouTube is the most cited source when AI models recommend software
Key findings
- YouTube received 2,669 of 29,119 citations (9.2%) across 33 categories and 10 models, two and a half times the next source.
- YouTube was the single most cited source in 16 of 33 categories. Reddit led 2, Firecrawl's own site led 2, and a single vendor or content site led each of the rest.
- Reddit was second overall with 1,058 citations (3.6%). G2 was cited 45 times, Product Hunt 31, TrustRadius 0, Capterra 0, GetApp 0, Software Advice 0.
- Five developer-infrastructure categories cited YouTube zero times in their top sources: authentication, inference hosting, LLM APIs, databases and transactional email.
- Citation volume differs by model an order of magnitude: Gemini 3.5 Flash returned 57 citations per answer, GPT-5.6 Sol 4.
YouTube accounted for 9.2% of every citation that ten current AI models returned when asked which software a startup, developer or creator should use. That is 2,669 of 29,119 cited URLs across 33 categories in the September 2026 edition of the Memetik Index. Reddit, the next source, drew 1,058. No other host passed 2%.
Which sources do AI models cite when recommending software?
Every host with 150 or more citations. Counts are raw URL citations returned by the model APIs, resolved to their host.
| Source | Citations | Share of all citations |
|---|---|---|
| youtube.com | 2,669 | 9.2% |
| reddit.com | 1,058 | 3.6% |
| firecrawl.dev | 582 | 2.0% |
| medium.com | 409 | 1.4% |
| zapier.com | 374 | 1.3% |
| braintrust.dev | 280 | 1.0% |
| valueaddvc.com | 242 | 0.8% |
| dev.to | 226 | 0.8% |
| mailtrap.io | 193 | 0.7% |
| guideflow.com | 190 | 0.7% |
| vellum.ai | 184 | 0.6% |
| therundown.ai | 175 | 0.6% |
| heygen.com | 161 | 0.6% |
| codeant.ai | 159 | 0.5% |
| confident-ai.com | 159 | 0.5% |
| daily.dev | 156 | 0.5% |
| sequenzy.com | 150 | 0.5% |
The tail is very long: 1,959 distinct hosts. Seven of the seventeen hosts above are vendors citing into their own category or an adjacent one (Firecrawl, Zapier, Braintrust, Mailtrap, Vellum, HeyGen, CodeAnt, Confident AI). The rest are tutorial and comparison publishers.
In which categories was YouTube the top source?
Sixteen of 33: AI code review (210 citations), online course platforms (186), AI coding agents (182), AI app builders (171), AI website builders (164), payroll (126), and AI avatars, AI image generation, AI music, AI voice, creator monetization, link-in-bio, automation, paid communities, short-form clipping and CRM.
The pattern is software with a large tutorial and review economy on YouTube. The surprise entries are developer tools: coding agents and code review are the two most YouTube-cited categories in the Index, ahead of every creator category.
Where does YouTube not appear at all?
Five categories had no YouTube citation among their top sources: authentication for developers, inference hosting, LLM APIs, databases for startups and transactional email. All five are backend infrastructure. In those categories the cited sources are documentation, engineering blogs and vendor comparison pages.
Which review sites were cited?
Almost none. Across 29,119 citations: G2 45, Product Hunt 31, TrustRadius 0, Capterra 0, GetApp 0, Software Advice 0. The review-site layer that has dominated Google results for “best CRM” style queries for a decade is close to absent from what the models cite.
Does citation behaviour differ by model?
By an order of magnitude. Across the same 165 prompts each:
| Model | Citations returned | Per answer |
|---|---|---|
| Gemini 3.5 Flash | 9,475 | 57.4 |
| Gemini 3.6 Flash | 5,665 | 34.3 |
| Sonar Pro | 3,223 | 19.5 |
| Sonar Reasoning Pro | 3,206 | 19.4 |
| Claude Fable 5 | 2,009 | 12.2 |
| Claude Sonnet 5 | 1,761 | 10.7 |
| Claude Opus 5 | 1,557 | 9.4 |
| GPT-5.6 Terra | 843 | 5.1 |
| GPT-5.6 Luna | 735 | 4.5 |
| GPT-5.6 Sol | 665 | 4.0 |
Google’s models cite most, OpenAI’s least. OpenAI’s API returns citation annotations only when its search tool fires and only for the pages it used, so its source profile is under-measured relative to Gemini, which returns grounding redirects for everything it read.
What this means for a vendor
For consumer software and developer tools alike, the reviewer’s video and the “X vs Y” comparison are what the models read first. A vendor with no YouTube presence in its category is absent from the layer the models cite most. For backend infrastructure the layer is documentation and engineering blogs. In both cases the traditional review site sits far below.
Limitations
One edition: 33 categories, 1,650 answers. Citation counts measure what the models returned as sources, not everything they read. OpenAI models are under-represented in citation counts for the reason above. Host counts treat a YouTube video and a channel page as one host. Gemini’s redirect URLs were resolved to real hosts before counting; a small number failed to resolve and were dropped.