Reddit, YouTube and the Shifting UGC Share of AI Citations
Short answer: The user-generated-content share of AI citations moves faster than any other part of the citation mix. Goodie AI measured YouTube’s share of social citations roughly doubling while Reddit’s halved over late 2025, and Promptwatch recorded Reddit’s ChatGPT Search share falling about 86% in mid-August 2026 — a drop its own publisher calls provisional. Treat UGC as a channel you influence honestly, never as a foundation you depend on.
Answer-engine advice treats citation sources as a fixed hierarchy: Reddit, then YouTube, review sites, news, Wikipedia, your own domain below. It is not fixed. In the past twelve months the UGC slice has re-ordered itself twice, and one of those moves is still unexplained by the people who measured it.
What actually moved, and who measured it
YouTube up, Reddit down — as a share of social citations
Goodie AI’s social citations study analysed 6.1 million citations across 10 AI platforms between August and 15 December 2025. Over that window YouTube’s share of social citations rose from 18.9% to 39.2%, while Reddit’s fell from 44.2% to 20.3%. The two almost exactly traded places.
The same study reports something most coverage dropped: Reddit’s absolute citation volume increased in November and December. Reddit did not shrink — the pool grew, and YouTube grew faster inside it. A falling share with rising volume is not a decline, and the two get conflated constantly in internal reporting.
August 2026: Reddit’s ChatGPT Search share falls off a cliff
Promptwatch’s tracking found reddit.com held a steady 3.8% share of ChatGPT Search citations from 18 July through 7 August 2026, then collapsed to under 1% on 14 August, averaging 0.5% across 14–17 August — an 86% relative decline.
The effect was not universal. Across Promptwatch’s wider 7 July to 17 August 2026 tracking window, it measured Reddit at 2.37% of Google AI Overview citations in the first seven days and 2.10% in the final seven (an 11.3% relative decline), and 2.22% of AI Mode citations falling to 1.54% (a 30.5% relative decline). Near-total collapse in one product, mild drift in two others — same domain, same summer.
Three reasons to hold this loosely
The measurer flagged it as provisional. Promptwatch states a data-collection issue cannot be ruled out and asks readers to treat the size of the drop as provisional. That caveat should travel with the number.
The proposed mechanism does not line up. Search Engine Journal’s Matt G. Southern, on 19 August 2026 in Why Reddit’s ChatGPT Citation Drop Isn’t Fully Explained, notes Promptwatch links the fall to a change in ChatGPT Search’s fanout queries: on 8 August, fanout queries using the site: operator jumped from 0.37% to 16.8% of all fanout queries in a day. But Reddit’s share did not collapse until six days later, and the piece is explicit that Promptwatch does not explain 14 August.
The popular explanation is not supported by the evidence we could check. A widely-shared theory held that Reddit had just added a blanket Disallow to its robots.txt in August 2026. Fetched live on 26 August 2026, https://www.reddit.com/robots.txt does contain User-agent: * followed by Disallow: / — and the Internet Archive’s capture of 1 July 2024 shows those same two directives, two years earlier. On that evidence the block is not new.
That is as far as the evidence goes, and it is worth saying why. The archived record of this particular file is not internally consistent. Captures of www.reddit.com/robots.txt one day apart in June 2025 return two entirely different files — 1 June 2025 the blanket Disallow: /, 2 June 2025 Reddit’s older permissive rule set. August 2026 captures alternate too, between the blanket block and a variant headed # Allowed for archiving that Reddit appears to serve to the archive crawler specifically. A robots.txt capture is evidence of what one crawler was served on one date, not proof of what the file said on every other date. So the defensible claim is the narrow one: nothing we could verify supports a new August 2026 block, which leaves the robots.txt theory unsupported — not that the file is proven never to have changed.
The direction is real and multiply observed; the magnitude is provisional and the cause unknown. That combination is the normal state of AI citation data — see AI answer volatility.
Why the UGC share moves faster than everything else
Access is commercial, not editorial. Large UGC corpora reach engines through negotiated arrangements and API access — renegotiated, repriced and switched off on business timelines unrelated to content quality. A source class whose availability is a contract term will always be more volatile than one gated only by a crawl.
Retrieval plumbing changes silently. The fanout behaviour Promptwatch observed is the example: a change in how an engine decomposes a prompt into background searches re-weights entire domain classes overnight, with no announcement and no way to detect it but measuring outcomes.
Format substitution is live. YouTube’s rise did not come from Reddit losing quality. It came from engines getting better at reading transcripts and video metadata as text. Any improvement in an engine’s ability to read a format shifts share toward whoever already publishes in it — which makes format coverage a distribution question, as in multi-format AEO.
Source classes: weight, influence and risk
The useful question is not “which source is biggest this month” but “how much can I move, and what happens if it moves without me”. Weight below reflects the broad direction of published citation studies, not one figure.
| Source class | Typical AI-citation weight | Vendor influence | Risk |
|---|---|---|---|
| Your own site | Low per page; reliable, cumulative | Total | Low — unless too thin or promotional to quote |
| Reddit and forums | Historically very high; now engine-dependent, volatile | Indirect only — reputation, never placement | High — manipulation breaches platform rules and leaves public evidence |
| YouTube | Rising fast as a share of social citations | Moderate — you publish; retrieval is not yours | Medium — production cost; transcripts must exist |
| Review sites (G2, Capterra) | Moderate; strong for comparison prompts | Moderate — solicit honestly, never fabricate | Medium — incentivised reviews breach most terms |
| News and trade press | Moderate to high; durable once indexed | Low to moderate — earned coverage only | Low, but slow and unpredictable |
| Wikipedia | Moderate; heavily trusted for entity facts | Very low — editing your own entry breaches policy | High if influenced directly; low if facts are verifiable elsewhere |
What a vendor should actually do
Be the thing people cite. Reddit threads get retrieved because they carry specific, first-hand, comparative detail with the marketing removed. Publish that on your own domain: real trade-offs, figures with owners attached, the cases where you are the wrong choice.
Participate genuinely. Named people, disclosed affiliation, answering questions in their real area of expertise, over months. It is the only forum strategy that survives moderation — practitioner detail in our post on Reddit citations for B2B SaaS.
Cover the formats, not just the topics. If YouTube’s share of social citations can double in five months, having no transcript-bearing video is a distribution cost. Publish the transcripts on pages you own, so the content survives whichever platform falls out of favour.
Own the pages comparison threads link to. When someone answers “which vendor should I use”, they link somewhere. A current comparison or methodology page gets the link, and the link outlives the thread’s citation window.
Measure share and volume separately, per engine. Goodie AI’s split (share down, volume up) and Promptwatch’s engine-by-engine split are the template; single-engine trackers mislead in both directions.
What not to do
Do not astroturf. The published Reddit Rules are unambiguous. Rule 2 requires users to “Participate authentically in communities where you have a personal interest, and do not spam or engage in disruptive behaviors (including content manipulation) that interfere with Reddit communities”. Rule 5 states: “Be authentic. You don’t have to use your real name, but do not intentionally mislead others or impersonate an individual or entity in a deceptive manner.” Sock-puppet accounts recommending your product breach both.
Do not buy votes or reviews. Beyond the rule breach, a vote is not a sentence. Engines quote prose, and a purchased score contributes none of it — the enforcement exposure is real and the citation upside is zero.
Do not outsource “community engagement” to anyone who will not name themselves. If the deliverable is anonymous positive mentions, that is astroturfing with an invoice.
Do not rebuild your strategy around a provisional number. The 86% figure carries the measurer’s own uncertainty caveat and no confirmed cause. Direction, yes; a year’s budget, no.
Do not confuse a share shift with a demand shift. Buyers did not stop reading forums in August 2026; an engine changed how it retrieves.
Where Zian fits
Zian AI builds autonomous AI sales agents for teams that want pipeline worked whilst marketing works on visibility: SmartReach AI™ orchestrates message, channel and timing by country, industry and profile, and PrecisionPitch AI™ split-tests approaches against real outcomes. Zian is not an AEO tool and will not win you citations. It is in waitlist and partnership beta.
Frequently asked questions
Has YouTube overtaken Reddit as an AI citation source?
Within social citations, Goodie AI’s study of 6.1 million citations across 10 AI platforms found YouTube’s share rose from 18.9% in August 2025 to 39.2% by 15 December 2025, while Reddit’s fell from 44.2% to 20.3%. That is a crossover in one dataset over one window, not a ranking — Goodie AI notes Reddit’s absolute volume rose in November and December.
Did Reddit really lose 86% of its ChatGPT citations?
Promptwatch measured reddit.com’s share of ChatGPT Search citations falling from a 3.8% average (18 July to 7 August 2026) to a 0.5% average (14–17 August 2026), an 86% relative decline. Promptwatch states a data-collection issue cannot be ruled out and asks readers to treat the size of the drop as provisional. Search Engine Journal’s 19 August 2026 coverage is titled “Why Reddit’s ChatGPT Citation Drop Isn’t Fully Explained” for good reason.
Was it caused by Reddit blocking crawlers in robots.txt?
No available evidence supports that. Reddit’s live robots.txt on 26 August 2026 does contain User-agent: * and Disallow: /, but the Internet Archive’s capture of 1 July 2024 shows the same two directives, two years before the August 2026 drop, so the block is not new. Draw no more than that from archives: captures of the file one day apart in June 2025 return two entirely different versions, so the archive is not a reliable record of what the file said on any given day. The accurate statement is that the robots.txt theory is unsupported by the evidence we could check, not that the file is proven unchanged. As at 26 August 2026 no confirmed cause has been published.
How often should we re-measure our AI citation mix?
Monthly at minimum, per engine, tracking both share and absolute volume. The August 2026 episode moved a major domain’s share by an order of magnitude inside four days on one product while barely touching two others. Quarterly single-engine reporting would have missed it entirely.
Every figure above was checked against its publishing page on 26 August 2026. If you want autonomous AI sales agents working your pipeline whilst you work on visibility, Apply For Partnership.