What 10.8 Million Customer Comments Reveal About How People Buy

Adlicio's users are ad operators. Over the last seven months they scraped 10,858,279 customer comments from 137,375 threads, product pages, and review feeds while researching the niches they sell into: 6,389 subreddits, plus YouTube videos, Amazon listings, Trustpilot pages, and more. That corpus is a strange and useful thing: not a random sample of the internet, but a map of the places where buying conversations actually happen, selected by people whose job is finding them.

We ran the numbers on it. Six findings, then the methodology.

1. 1% of comments get 57% of the attention

Across a 211,456-comment sample of Reddit threads, upvotes are radically concentrated:

46.5% of comments earn one upvote or fewer. The practical read for anyone mining comments for ad angles: the swipe file is small. A thread with 400 comments usually contains four that the audience has already voted into headline status. Sorting by score is not a shortcut; it is the method.

2. Reddit asks, reviews answer

The share of comments containing a question splits the platforms into two funnel stages:

Nearly one in five Reddit comments is a question: what worked for you, is it worth it, does it help with X. That is pre-purchase language, the objections and doubts an ad has to answer. Reviews sit on the other side of the wallet: people who already bought, reporting back. Mine Reddit for the questions your ads should answer, and reviews for the answers your ads should quote.

3. The long testimonials live in reviews, not threads

Share of comments over 500 characters, the length where a comment becomes a story:

Nearly a quarter of Amazon reviews are full narratives: the problem, the doubt, the purchase, the result. Amazon reviewers also name a time-to-result ("within days", "after a month") about 6 times more often than Reddit commenters (1.71% vs 0.30%). If you need before-and-after material with a timeline, reviews are where it already exists.

4. One in nineteen Reddit comments hands you its own targeting

5.19% of Reddit comments contain an "as a ..." self-label: as a nurse, as a new dad, as someone with ADHD. The commenters are doing audience segmentation on themselves, unprompted. Collect these labels across a niche's threads and you get the call-out lines ("Night shift nurses:") that out-target any interest stack, in the audience's own words.

5. Every platform has an emotional signature

Rates of language types by platform, from a 481,028-comment sample:

Trustpilot is the risk conversation: scam-doubt runs 16x Reddit's rate and refund language 28x. It is the single best source for the trust objections your funnel must answer. Amazon is the endorsement conversation: nearly 12% of reviews contain recommendation language, and failed-alternative stories run 4x Reddit's rate. Reddit is earlier and rawer than both, which is exactly why its phrasing makes hooks.

6. Half of Reddit is replies, and that's where objections get resolved

49.3% of scraped Reddit comments are replies rather than top-level comments. The thread is not a list of opinions; it is a negotiation. Someone raises a doubt, someone else answers it with their experience. Those answer-shaped replies are objection-handling copy written by the most credible author possible: a customer with no stake in the sale.

What this means if you write ads

Each surface has a job. Use Reddit threads for hooks, questions, and identity call-outs. Use Amazon reviews for proof, timelines, and recommendation language. Use Trustpilot for the trust objections and the refund fears. Use YouTube comments for fast reactions to competitor content. This division of labor is the entire premise behind voice of customer research, and the reason Adlicio scrapes all of these surfaces rather than one.

Methodology

Corpus: 10,858,279 comments from 137,375 scrapes collected by Adlicio users between January 15 and August 19, 2026, spanning 6,389 subreddits plus YouTube, Amazon, Trustpilot, Steam, Facebook, and other public comment surfaces. Platform-level analyses ran on a 5% random sample (481,028 comments); the upvote-concentration analysis ran on a separate 2% random sample of Reddit threads (211,456 comments). Language markers are case-insensitive phrase matches, which undercounts each category (people express skepticism in more ways than we pattern-matched), so the true rates are floors, not ceilings. One bias to keep in mind: these threads were hand-picked by ad operators researching buying conversations, so the corpus over-represents commercial niches relative to the internet at large. For studying how buyers talk, that is the point. No individual comments are quoted; all figures are aggregates.

Every figure from this study, plus a few more cuts, lives on the voice of customer statistics page in citable one-line form; all of it is free to reference with a link. Questions about the data, or a cut you'd like to see? Ask us; the corpus grows daily.