OL OpenLens research · August 2026

ChatGPT stopped citing Reddit in five days. The links went to government, standards, and docs.

As of August 18, about 1 in 1,000 ChatGPT citations points at Reddit. On August 6 it was 1 in 35. ChatGPT still returns 10 to 11 links per answer. It cites different things now: government and law pages went from 8% to 21% of all citations in the same window. No other AI engine moved.

Data through 2026-08-18 7.25M citations · 11 engines · 909,005 ChatGPT 24 client projects · 529 prompts with data in both windows Re-check promised 2026-09-15
Abstract The whole story in twelve sentences
  1. Since August 6, 2026, Reddit's share of ChatGPT citations fell from 2.9% to 0.1% (as of August 18).
  2. ChatGPT still gives 10 to 11 links per answer. It cites different sites, not fewer.
  3. The same prompts run in both windows show the same fall, so the questions did not change.
  4. 219 of 227 prompts that cited Reddit before now cite it zero times, across all 24 client projects.
  5. It happened in two dated steps: a dwindle from August 8, the day Promptwatch documented a change in how ChatGPT Search splits queries, and a cliff on August 14. Promptwatch saw the same cliff in its own data.
  6. This is ChatGPT only. Google AI, Perplexity and Gemini held flat on Reddit. Claude has never cited Reddit.
  7. Reddit is one casualty. Mainstream news (WSJ, Reuters, FT, El País) fell 88% to 1 in 500 citations. arXiv fell 93%. Forums and social fell 79%.
  8. Government and law sites took almost all of it: 8% to 21%. Groups that set rules and certifications (ISO, IEC, Fairtrade) went 4x. Official product docs went 2.7x. PubMed rose 65%.
  9. Cochrane shows the pattern. Its review database went from 0 to 46 citations per 10,000 while the website about it fell 77%. ChatGPT wants the record, not the page about the record.
  10. News, preprints and forums fell together while no other engine moved. That is the signature of a phased update to ChatGPT's search system. We cannot tell whether it is OpenAI's ranking or its search backend, and we do not know if it will revert. We re-check on September 15.
  11. These numbers come from brand-monitoring prompts, skewed commercial and Spanish-language. Company pages are still 30% of citations and still get cited when they read as official documentation.
  12. In ChatGPT, a regulatory filing, a certification listing, a docs portal, or a PubMed paper now beats any Reddit thread or press placement. Reddit is still influential. For citations it is now a Google AI and Perplexity play.
Figure 1 Reddit fell in two steps, then hit zero
Reddit share of all ChatGPT citations 0% 2% 4% 6% 5.48%3.14%2.74%1.22%0.12% JuneJulyAug 1 to 7Aug 8 to 13Aug 14 to 18 435,994 cites300,16752,67673,91546,253 step 1: Aug 8 dwindle step 2: Aug 14 cliff
What to seeTwo drops, both dated. Raw shares over all ChatGPT citations in each period (totals under each bar). The peak was 8.85% on June 17. On August 16 it was 2 citations out of 9,526. Total citations per period shrink because fewer prompts were run, not because ChatGPT cites less: citations per answer stayed at 10.2 to 11.5 the whole time.
Figure 2 Government and law took almost all of the lost share
Change in share of ChatGPT citations, percentage points (same prompts, before vs after) Government & law8.2% → 21.2% (+13 pp) Standards & certification0.5% → 2.0% (4x) Vendor docs & official product0.8% → 2.2% (2.7x) Associations & nonprofits (.org)2.3% → 3.4% Wikipedia & wikis0.09% → 0.04% Review & comparison sites0.9% → 0.6% Tech & trade press1.4% → 0.6% Academic & journals5.7% → 4.9% (arXiv -93%, PubMed +65%) Mainstream news & finance1.7% → 0.2% (-88%) Health publishers3.4% → 1.7% Forums & social (UGC)3.2% → 0.7% (Reddit 2.9% → 0.1%) Company, SMB & marketing sites39.5% → 30.5% (still the largest) 0+10 pp-10 pp Unclassified long tail (32%, mostly small company sites) not shown.
What to seeOne green bar dwarfs the rest. Government and law gained 13 points. Every other gainer combined got about 4. Computed on the 79,297 before and 31,678 after citations from prompts that ran in both windows.
Figure 3 Only ChatGPT moved
Reddit share of citations, by engine (Jul 18 to Aug 7 vs Aug 14 to 18) 0%1%2%3% 2.950.12 1.722.15 0.680.83 0.800.79 0.00 / 0.00 ChatGPTGoogle AIPerplexityGeminiClaude -96%flat, drifting down since Juneflatflatnever cites Reddit beforeafter
What to seeOne engine fell off a cliff; the others did not notice. Google AI cites forums and social at 14 to 16% of its citations (mostly YouTube), about 21x ChatGPT's post-break 0.68%. Claude has 0 Reddit citations in 49,218. Grok is omitted: too few citations and measured through the API, so not comparable.
The argument Each heading is a claim. Read the headings alone and you have the piece.

ChatGPT cites the same number of links. It chose different sites.

11.0 citations per answer in early June, 10.2 to 11.5 through August. If our collection had broken, that number would have fallen. It did not. The total citation counts in Figure 1 shrink because we ran fewer prompts in later periods, not because ChatGPT cited less.

The same prompts show the same drop, so the questions did not change.

All headline numbers use only prompts that ran in both windows: 79,297 citations before, 31,678 after. On those, Reddit went 2.87% to 0.11%. Raw, uncontrolled numbers say 2.95% to 0.12%. Of 227 prompts that cited Reddit before the break, 226 fell and 219 went to exactly zero. All 24 client projects fell; 15 went to zero.

Why the control mattersRaw data showed a cluster of Indian hospital sites (Apollo, Fortis, Practo) "surging" after August 14. Under the control they vanish: they came from newly added prompts, not from ChatGPT. Anyone re-running this on their own tracker should apply the same filter.

The drop has two dates, August 8 and August 14.

Share went 2.74% (Aug 1 to 7) to 1.22% (Aug 8 to 13) to 0.12% (Aug 14 to 18). August 8 matches a change Promptwatch documented in how ChatGPT Search expands one question into several sub-searches. OpenAI's release notes say nothing about August 14. Promptwatch saw the same August 14 cliff in its own data, 3.83% to 0.52%, and published first.

News, preprints and forums fell with Reddit. Government, standards and docs rose.

Mainstream news: 1.66% to 0.20%. WSJ 57 to 1 raw. El País 377 to 1. The Verge, ITPro, Diario AS, Heise to 0. arXiv 770 to 22. Meanwhile osha.gov rose 33x, webstore.iec.ch 34x, fda.gov 18x, cisa.gov 12x, apa.org 11x, developers.google.com 9x, eur-lex 6x, nist.gov 5x. All rates per 10,000 citations, same-prompt basis. Full tables in the appendix.

Cochrane shows what "authority" means to ChatGPT now.

cochranelibrary.com, the database of systematic reviews, went from 0 to 46 per 10,000. cochrane.org, the organisation's website, fell 77%. Same institution, opposite fates. ChatGPT now prefers the primary record over the page that describes it. PubMed is up 65% and is now the most-cited academic source; arXiv preprints are down 93%.

This is a ChatGPT search update, and we cannot say which part changed.

Four facts point the same way: the change is sharp and dated, it hit forums, news and preprints together, it replicates across every project, and no other engine moved. That is the signature of a phased update to ChatGPT's consumer search system. What we cannot tell from citation data: whether the change is in OpenAI's source ranking or in the search backend it uses, whether it is fully rolled out, and whether it will revert. There is no public evidence of a Reddit licensing dispute, and a licensing change alone would not explain news and arXiv falling too.

Company pages are still the biggest category, and the prompt panel shapes every percentage here.

Corporate and SMB pages are 30% of citations after the break (40% before), the largest single bucket by far. The panel behind every number in this post is brand-monitoring prompts from 24 client projects, skewed commercial and Spanish-language (Chile especially). Absolute shares will differ on your panel. The direction of the same-prompt changes should not.

Monday Three things to do this week

B2B and tech: check whether your docs portal is crawlable.

ChatGPT now cites vendor documentation 2.7x more and trade press 0.4x as much. Crawl your own docs, support center and developer portal. Fix noindex, gated PDFs and JS-only rendering. One engineer-day, zero cash, measurable next week.

Publishers: re-weight across engines instead of retreating.

The ChatGPT drop is ChatGPT only. The New York Times is at 0 per 10,000 on ChatGPT but 5.1 on Google AI and 5.4 on Perplexity. Perplexity's trade-press share is at a three-month high. Claude's news share hit a dataset high last week (small volume, direction only). Specialist and reference content survived better than news-cycle coverage: trade press fell 57%, general news 84% on the full-July baseline.

Everyone: get your facts into a primary record.

A certification registry, a standards listing, a regulatory docket, a professional association's guidelines, or a PubMed-indexed paper now outranks a Reddit thread or a press hit in ChatGPT. Slow to build, but it is where the citations went. Keep Reddit for Google AI and Perplexity; Google alone cites forums and social at 21x ChatGPT's rate.

Open questions What we do not know yet

Will it bounce back? Our dataset starts February 26, 2026. We have no prior ChatGPT retrieval flip to compare against. We re-run this table on September 15 and will say so if it reverts.

Is this the reasoning model? Possibly. A fanout that rewrites "best X reddit" into "X official specification" produces exactly this pattern. We cannot test it from citations alone.

Did Reddit block the fetcher? It would not explain news and arXiv falling on the same day, but it could add to the size of the Reddit drop. Untested.

Does a citation equal traffic? Not measured here.

Method Six steps, an afternoon to reproduce
  1. Take every ChatGPT answer in the OpenLens citation database from June 1 to August 18, 2026: 81,373 answers, 909,005 citations, 24 client projects. The full database is 7,251,308 citations across 11 engines.
  2. Define before = July 18 to August 7 and after = August 14 to 18.
  3. Keep only prompts that produced answers in both windows (mix control). Report shares and per-10,000 rates, never raw counts across windows.
  4. Classify domains with a rule-based list of about 1,150 web domains. That covers 68% of citations in the two windows; the rest is a long tail of small sites, under 16 citations each.
  5. Compute category shares before and after; repeat per prompt (529 prompts with at least 20 citations in both windows) and per project (24 with at least 200).
  6. Repeat for Google AI, Perplexity, Gemini, Claude, Grok on the same windows.
Appendix The tables
A. Category shares, same prompts, before vs after
CategoryBeforeAfterChangeMultiplier
Government and law8.23%21.18%+12.95pp2.6x
Standards and certification0.50%2.01%+1.51pp4.0x
Vendor docs and official product0.82%2.23%+1.40pp2.7x
Professional associations and nonprofits2.29%3.44%+1.15pp1.5x
Wikipedia and wikis0.09%0.04%-0.04pp0.5x
Review and comparison sites0.90%0.58%-0.33pp0.6x
Tech and trade press1.39%0.60%-0.79pp0.4x
Academic and journals5.67%4.86%-0.81pp0.9x
Mainstream news and finance media1.66%0.20%-1.46pp0.12x
Health publishers and providers3.43%1.70%-1.72pp0.5x
Forums and social (UGC)3.23%0.68%-2.54pp0.2x
Corporate, SMB and marketing sites39.48%30.48%-9.01pp0.8x
B. Domains cited more, per 10,000 ChatGPT citations
DomainWhat it isBeforeAfterMultiplier
webstore.iec.chIEC standards store0.93034x
osha.govUS workplace safety1.13833x
fda.govUS FDA46818.6x
ams.usda.govUSDA Agricultural Marketing Service58918x
cisa.govUS cybersecurity agency910511.6x
apa.orgAmerican Psychological Association33611.3x
developers.google.comOfficial developer docs6569.2x
gesetze-im-internet.deGerman federal law portal272147.9x
eur-lex.europa.euEU law database13775.9x
nist.gov, csrc.nist.govUS standards institute331665.1x
nidcr.nih.govNIH dental institute361584.4x
epa.govUS EPA461092.4x
pubmed.ncbi.nlm.nih.govBiomedical literature index1161921.7x
C. News outlets, raw citations (21 days before, 5 days after; rate change per 10k). Rows with fewer than 50 before are labeled small.
OutletBeforeAfterRate change
El País3771-99.3%
TechRadar51432-84%
La Vanguardia1374-93%
Forbes12119-61%
Search Engine Land11014-68%
Tom's Guide9725-35%
La Tercera822-94%
Business Insider7915-53%
ITPro690-100%
Diario AS680-100%
TechTarget627-72%
WSJ571-96%
Reuters (small)445-72%
Financial Times (small)432-88%
Time (small)351-93%
The Verge (small)340-100%
Axios, Guardian, Washington Post, Chip.de, Heise, Independent, Emol, Infobae (small)16 to 28 each0 or 1-84% to -100%
D. Reddit share by engine, and weekly news share by engine
EngineReddit beforeReddit afterNews: June avgJuly avgWk Aug 10Wk Aug 17
ChatGPT2.95%0.12%1.49%1.76%0.49%0.28%
Google AI1.72%2.15%0.72%0.75%0.64%0.75%
Perplexity0.68%0.83%0.56%1.19%1.13%0.67%
Gemini0.80%0.79%0.89%0.69%0.35%0.44%
Claude (small volume)0.00%0.00%0.55%1.13%0.73%2.33%

Week of Aug 17 is two days of data. Google AI's Reddit "+25%" is a short-window artifact; on the June-to-August run its Reddit share drifts down 3.18% to 1.96%. Neither is a break.

Sources