Reddit, G2, YouTube: the off-site sources AI trusts most

When ChatGPT recommends a product, it isn't consulting your website. It's reading what the rest of the internet says about you — and “the rest of the internet,” for AI answers, is a surprisingly short list. Across hundreds of millions of AI citations analyzed in 2025 and 2026, the same handful of platforms keeps deciding who gets recommended: Reddit, Wikipedia, YouTube, LinkedIn, and a tight circle of review sites and editorial publications.

This guide maps where each AI engine actually pulls its answers from, why those sources won, and what showing up on each one really takes — with the folklore stripped out. The short version: the winners aren't who you'd guess, upvotes and view counts matter far less than everyone assumes, and the compliant, honest play is also the one that works.

01The citation map: who AI actually reads

Profound analyzed 680 million AI citations from August 2024 through June 2025. On ChatGPT, the single most-cited domain was Wikipedia at 7.8% of all citations — nearly half of ChatGPT's top-ten source mix on its own. Perplexity leans hardest on Reddit, which makes up 46.7% of its top-ten mix. Google's AI results skew to its own ecosystem plus community content: Reddit, YouTube, Quora, and LinkedIn lead.¹

Ahrefs' June 2026 snapshot of Google AI Overviews, drawn from more than 3 million US queries, puts the current hierarchy in sharp relief:²

One caveat belongs at the top of any citation chart: these shares are volatile. Semrush tracked 230,000 prompts over three months in 2025 and watched ChatGPT's Reddit citation rate collapse from roughly 60% of responses in early August to about 10% by mid-September, while Wikipedia fell from ~55% to under 20% in the same window.³ The specific percentages move; the set of platforms that matters has stayed remarkably stable.

02Why these sources won

The citation map isn't an accident of taste. It's largely a story about access.

Reddit signed a reported ~$60M-per-year licensing deal with Google in February 2024, giving Google real-time API access to Reddit content,⁴ and announced a partnership with OpenAI that May, wiring Reddit content into ChatGPT.⁵ The results are visible in both directions: Reddit went from the #68 most-visible domain in US Google search in mid-2023 to #5 a year later — the single biggest SEO winner of 2024⁶ — and its data-licensing revenue line went from effectively zero to $114.7M for full-year 2024.⁷

Access works in reverse, too. Review platforms that block AI crawlers — Yelp and Trustpilot block them entirely — are measurably weaker in AI answers, while platforms that allow search-time access (G2, Capterra, TripAdvisor, Software Advice) or unrestricted crawling (Clutch, SourceForge, TrustRadius) soak up citations in their categories.⁸ When a source can't be read, it can't be cited — and someone else defines your category.

03Which platform matters depends on what you sell

Averages hide the useful part. Hall analyzed 456,570 citations across ChatGPT, Perplexity, Google AI Overviews, and Copilot in mid-2025 and found each engine has different “house favorites” per category.⁸ For B2B software queries, ChatGPT routed 47.65% of its review-platform citations through GetApp — while G2 took just 8.25%. Copilot preferred SourceForge and Software Advice. For travel and local queries, TripAdvisor took 73–94% of review citations on every engine. For agencies and digital services, Clutch took 66–84.5% everywhere.

SE Ranking's December 2025 snapshot of 30,000 commercial keywords adds a concentration warning: in Google AI Overviews for software queries, 88% of review-platform citations flowed to just five sites — Gartner Peer Insights, G2, Capterra, Software Advice, and TrustRadius. Smaller directories got zero.⁹

That's the paradox to internalize: TrustRadius traffic fell 92% while it stayed inside the five sites that define software categories for Google's AI.⁹ The influence didn't disappear — it moved upstream, into the answer itself. Your G2 profile's job is no longer to be visited. It's to be quoted.

04Reddit: forget karma, answer questions

The folklore says you need viral threads. The data says otherwise. Semrush analyzed 248,000 Reddit posts cited by Google AI Mode, Perplexity, and ChatGPT Search and found that most cited posts have fewer than 20 upvotes and fewer than 20 comments.¹⁰ Engagement isn't the ranking signal — topical fit and clarity are.

• Format wins: over half of cited Reddit content is Q&A-style threads, followed by comparisons and structured discussions.¹⁰
• Age compounds: the average cited post is ~2.5 years old. A thread answered well today is an asset that pays out for years — and nothing seeded this quarter will rescue this quarter.¹⁰
• Astroturfing gets caught: in 2025 an agency seeded ~100 fake “organic” comments promoting a game studio's launch, got exposed by users, banned, and turned into a public case study in what not to do.¹¹ Reddit flags coordinated behavior by account age, karma, and timing patterns.

05Review platforms: recency and depth, not blitzes

G2's own scoring documentation says the quiet part: review weight decays over the first 90 days and holds meaningful weight for about 18 months, and longer, detailed reviews are weighted more heavily than thin ones.¹² A one-time review blitz produces a spike that quietly expires. A steady drip — a few detailed reviews every month, sourced from real customers at natural moments — compounds into the profile AI keeps citing.

Two rules keep this clean. The FTC's rule on consumer reviews (effective October 21, 2024) bans fake and AI-generated reviews, paying for positive or negative sentiment, and undisclosed insider reviews — with civil penalties up to $51,744 per violation.¹³ And Google separately prohibits “review gating”: selectively steering only happy customers toward the review form can get all of your reviews removed.¹⁴ Ask every customer, disclose incentives, and let the honest distribution stand.

06YouTube: long-form, chaptered, views irrelevant

OtterlyAI studied more than 100 million AI citation instances across six engines and isolated what gets a YouTube video cited. The findings invert most channel strategy:¹⁵

• 94% of cited videos are long-form; Shorts take just 5.7%. The most-cited duration cluster is 10–20 minutes.
• Popularity is statistically irrelevant — view count correlates with citation at r ≈ −0.03, and 40.8% of cited videos had under 1,000 views. A 200-subscriber channel can be the source AI quotes.
• Chapters multiply citations: 78% of timestamped videos get cited repeatedly across 2–5 chapters — chapters work like subheadings that give the model more places to point. Yet only 31% of cited videos use them.

If you make one YouTube change after reading this, add descriptive chapters and a real description to every video you already have. It's the highest-leverage, lowest-effort move in the study.

07What this means for your next quarter

Pick your column from the category map, then work the short list: answer the standing questions on Reddit as yourself, keep a steady drip of detailed reviews flowing to the two platforms that own your category, and publish long-form, chaptered video for the questions buyers actually ask. None of it is fast — the average cited Reddit thread is two and a half years old — which is exactly why the brands that start now end up being the answer later.

And measure it. Citation shares moved 6× within a single six-week window last year; the only way to know which sources decide your category's answers this month is to track the answers themselves.