Reddit lost 86% of its ChatGPT citations. Here’s who actually took them.
Not the publishers. While everyone was writing Reddit’s obituary, the vacated share was already being collected by YouTube, Wikipedia, and, in Google’s AI answers, Facebook. On ChatGPT itself, the biggest winner may be the one surface you control: your own documentation.
The second cliff in eleven months
Let’s get the event out of the way fast, because it is already well covered. PromptWatch data shows reddit.com holding a steady 3.83% average share of ChatGPT Search citations from July 18 to August 7, 2026. On August 14 the share fell below 1%, and the August 14 to 17 average of 0.52% is an 86.4% relative collapse. RDDT dipped in premarket trading today partly on search-traffic risk. This is the second time in under a year that Reddit has lost more than 80% of its ChatGPT citation share in days: in September 2025 the share fell from 29.2% to 5.3% after Google killed the num=100 scraping parameter, and Reddit’s stock dropped 14.4% over five trading days ($240.11 to $205.50, per G2’s breakdown).
Crash #1 · Sept 2025
Trigger: Google removed the num=100 results parameter that OpenAI’s data suppliers scraped. RDDT fell 14.4% in five sessions.
Crash #2 · Aug 2026
Trigger: ChatGPT Search changed query fanout on Aug 8, then Reddit fell off the cliff on Aug 14. RDDT dipped premarket Aug 19.
The questions everyone is asking this week: is Reddit dead in AI search, and who picks up the traffic? The first answer is no. The second answer is where it gets interesting, because the heirs are not who the trade press assumed.
First correction: Reddit did not drop everywhere
Most coverage treats “Reddit dropped in AI” as one event. PromptWatch’s tracker measured three different curves over the same July 7 to August 17 window, and they tell three different stories:
| Platform | Before | After | Relative change | Shape |
|---|---|---|---|---|
| ChatGPT Search | 3.83% | 0.52% | -86.4% | Single-day cliff, Aug 14 |
| Google AI Overviews | 2.37% | 2.10% | -11.3% | Gradual drift, no break |
| Google AI Mode | 2.22% | 1.54% | -30.5% | Gradual, step in late July |
So any practitioner playbook has to be per-platform. Reddit’s Google visibility is fading slowly and is still worth defending. Its ChatGPT visibility was switched off in a single afternoon, and that is where the inheritance question matters.
Who took the share on ChatGPT: structured first-party pages
The most direct answer arrived today, one day after PromptWatch published the cliff. GEO practitioner Ira Bodnar posted a source-category comparison of what ChatGPT cites under GPT-5.1 versus GPT-5.6 Sol (OpenAI’s current flagship, generally available since July 9, 2026), and it was written up by explainx.ai with the right caveat attached: this is a model-version comparison from a single practitioner, not independently re-verified. Treat it as a second data point pointing the same direction as PromptWatch, not as a replica of the 86% number.
That destination matches the mechanism we can see on the retrieval side. On August 8, ChatGPT Search started running site:-scoped queries at scale: the share of fanout queries containing a site: operator jumped from about 0.4% to roughly 17% overnight, and average fanouts per response climbed from 1.08 to 1.83. ChatGPT stopped only searching the open web and seeing what comes back; it now deliberately queries specific domains, and longer, qualifier-heavy sub-queries retrieve authoritative single-purpose pages instead of broad forum threads. Reddit’s indexed content is mostly the second shape. Your documentation is the first.
The other steady ChatGPT heir is Wikipedia. PromptWatch’s domain tracker has it as ChatGPT’s most-cited knowledge base, and when the first crash hit in September 2025, Profound data analyzed by G2 found Wikipedia’s visibility rose as Reddit’s fell: canonical, structured content that consistently ranks in the top results is what retrieval systems reach for when the forum goes dark.
Who took the share on Google: YouTube, Wikipedia, and the Facebook surprise
Google’s AI surfaces tell a different inheritance story, and this is where the counterintuitive part lives. Surfer SEO’s analysis of 46 million AI Overview citations, synthesized by 5W in May 2026, puts the top three sources of Google AI Overviews at: YouTube, 23.3% of all citations; Wikipedia, 18.4%; Google.com itself, 16.4%. Every legacy news outlet sits below that top three.
YouTube’s overtake of Reddit is not even news from this week. ADWEEK reported in January 2026 that Bluefish data showed YouTube appearing as a cited source in 16% of LLM answers over the prior six months versus 10% for Reddit, making YouTube the most-cited social platform in AI answers. The mechanism is unglamorous: transcripts, chapters, descriptions, and metadata make video content machine-readable text, which is exactly what a retrieval layer can ingest and attribute.
Then there is the number nobody expected. BrightEdge’s study of more than 300 million US monthly searches, covered by Search Engine Journal this week, found Google’s AI Overviews cited Facebook as a source 19.5 million times, Instagram about 877,000 times, and TikTok around 78,000.
| Platform | AI Overviews citing it | Share of ~300M searches | Role in the funnel |
|---|---|---|---|
| 19,500,000 | ~6.5% | After-sale: ~23% of bottom-funnel citations are troubleshooting, returns, how-to | |
| 877,000 | ~0.29% | Purchase: ~90% of bottom-funnel citations are buy queries (where, price, sale) | |
| TikTok | 78,000 | ~0.026% | Trends and how-to clips |
Facebook, the platform SEOs have been writing eulogies for since 2018, is now a first-class citation source in Google’s AI answers. The posts that get cited are not the big accounts. The BrightEdge examples include a small account’s Instagram post answering “mobile payment app” (an estimated 18.5M monthly searches) and a baseball team’s Facebook post answering “where to watch Brewers versus Reds.” AI cites whatever answers the exact question, in public, in text the crawler can read.
Who did NOT inherit it: the publishers
Here is the part that breaks the comfortable narrative. The assumption after every Reddit wobble has been that AI engines would hand credibility back to traditional media. 5W’s Citation Source Audit for Q1 2026, synthesizing Similarweb’s roughly 600,000 US ChatGPT citation events from January and February 2026, says otherwise:
| Domain | Share of citations | Note |
|---|---|---|
| Wikipedia | 13.15% | #1 |
| 11.97% | #2 (pre-crash window) | |
| Reuters | 2.27% | #7, top news wire |
| Forbes | 1.38% | #18, only US business publication in the top 20 |
| WSJ | n/a | Not in top 20 |
| New York Times | n/a | Not in top 20 |
| Bloomberg | n/a | Not in top 20 |
| Financial Times | n/a | Not in top 20 |
The fast movers worth watching
- LinkedIn jumped from #11 to #5 on ChatGPT in three months, and is now cited in 14.3% of ChatGPT Search responses (SEMrush, 325,000 prompts, via 5W). The largest rank shift Profound observed all year.
- Fandom leads Google AI Mode at 7.16% of citations, ahead of Wikipedia. Structure plus depth beats brand authority there.
- Review platforms split by engine. PromptWatch’s AIO tracker has G2 as the review winner in Google AI Overviews (roughly 3 to 4x Capterra and Trustpilot), while Trustpilot trends upward on ChatGPT. 5W adds that brands listed across G2, Capterra, Trustpilot, and Yelp see about a 3x citation multiplier versus brands without those profiles. Note: in Bodnar’s August breakdown, review sites as a category collapsed alongside Reddit on ChatGPT, so treat the engine split as volatile.
- YouTube’s correlation with AI visibility is 0.737 in Ahrefs’ 75,000-brand analysis, the strongest single predictor in any 2025-2026 study 5W reviewed.
What Reddit still owns (do not over-correct)
Before anyone pulls every Reddit campaign: Conductor’s four-month study (October 2025 to January 2026) shows Reddit’s overall citation share halving from 2.02% to 1.01%, but its share of responses where Reddit is the only cited source rose 31% in the same window. When Reddit wins, it increasingly wins alone, and the intent mix of those sole citations is specific: 36.5% transactional (“is X actually worth it?”), 33.5% commercial (“best X for Y”), 25.6% informational troubleshooting, 4.4% navigational. And the categories where it never wins are equally specific: medical authority, financial definitions, reference facts, and structured how-tos with an official manual.
That is a targeting brief, not a funeral notice. Reddit remains the go-to for lived-experience and trade-off prompts on Google’s surfaces, where its decline is gradual. It is ChatGPT where the lights went out, and ChatGPT where your expectations need resetting.
The six-move playbook for this week
Caveats and the watchlist
OpenAI declined to comment for Gizmodo’s coverage. Reddit’s spokesperson said the platform “remains one of the most cited domains” per other reports and that Reddit does not rely on LLMs for traffic, which is true for its business and irrelevant to yours. What to watch next: whether the ChatGPT cliff holds through this week once PromptWatch checks collection volume around August 14, which specific domains gained as Reddit fell (PromptWatch explicitly recommends that comparison and has not published it yet), and whether Google’s gradual fades accelerate.
The pattern across both crashes is the lesson. When Reddit stumbled in September 2025, Wikipedia absorbed it. When Reddit stumbled this August, structured first-party pages, YouTube, and Facebook absorbed it. The share never flows back to general publishers, and it never flows to the least machine-readable surface available. It flows to whatever the model can read cleanly: transcripts, canonical wiki text, single-purpose docs, and public social posts with the answer in the text. The uncomfortable corollary for practitioners is also the opportunity: the biggest heir on ChatGPT is the content you already own, if it is built to be retrieved.
