---
title: "Reddit's ChatGPT Citations Fell 86%. Last Year They Fell 95% And Came Back."
description: "A circulating six point post says ChatGPT rebuilt its sources. One study is real. The other five numbers came from elsewhere and broke in transit."
published: 2026-08-19
author: "Ben Tabas, Founder"
publisher: "Zion Labs"
canonical: https://zionlabs.io/research/chatgpt-citation-shift
source: Zion Labs research
---

# Reddit's ChatGPT Citations Fell 86%. Last Year They Fell 95% And Came Back.

## 1. The claims

A six point post is circulating about ChatGPT's sources. We could not trace the original author across five distinct searches, so the claims are quoted here and attributed to nobody. In summary form they read:

- Reddit citations went from about 15% to zero.
- G2 and Capterra went from about 7% to zero.
- Documentation and help centres now take 32% of citations.
- Small company pages fell from 66% to 32%.
- App marketplaces take 17%.
- Overall, ChatGPT has shifted toward established brands.

One of those six describes a measurable event. The other five are real figures from unrelated studies, with their denominators swapped, their dates moved, or their units changed. This piece gives the provenance for each, then covers what actually changed and what a practitioner should do about it.

## 2. Provenance

| Claim | The study it came from | The error |
|---|---|---|
| Reddit 15% to zero | rocketblue.ai: 14.29% to 0.21%, 3M+ citations, 8 models, August 5 to October 29 2025 | Rounded up, then zeroed, then dated to the wrong year |
| G2 and Capterra 7% to zero | Derivatex: 40 B2B SaaS categories, 10 runs each, 233 recommendations, 169 cited URLs, published May 2026 | The 7% baseline exists in no study we could find |
| 32% docs and help centres | Ahrefs: 32.3% of the top 1,000 pages ChatGPT cited in September 2025 were influenceable by outreach | A page type census read as a source category share |
| Small pages 66% to 32% | Position Digital: 66% of brand recommendations carry no citation to that brand's own site. 278 responses, 95 brands, July 2026 | Not a measurement of company size, and not a change over time |
| App marketplaces 17% | No traceable study. No traceable study | No traceable source |
| Shift to established brands | No traceable origin | Contradicted by the largest citation dataset available (below) |

## 3. Reddit did this before, and came back

The one real event is measured by Promptwatch, over July 7 to August 17 2026. From July 18 to August 7, Reddit held a steady 3.83% average share of ChatGPT citations. It fell below 1% on August 14. The August 14 to 17 average was 0.52%, an 86.4% relative decline. The first movement came earlier: on August 8, the same day ChatGPT changed its query fan out behaviour, Reddit's share slid from the high 3s to the mid 2s.

Google's surfaces did not show the same cliff over that window. AI Overviews declined 11.3% relative and AI Mode 30.5%, both gradual, with no single sharp break.

Two things have to travel with that number. First, the publisher's own caveat, verbatim: "A shift in ChatGPT's source selection is the obvious candidate, but a data-collection issue cannot be ruled out, so treat the size of the drop as provisional while we keep monitoring." Second, no sample size is disclosed for this specific analysis. The platform reports 26 billion data points overall, and that figure does not attach to this chart. There is also no independent replication. The event was first flagged by a Promptwatch cofounder, which means the vendor flagged its own chart, and no second measurement of it exists.

Now the part almost nobody circulating the post has, and the reason the "Reddit is finished" reading fails.

Reddit already collapsed once, harder, eleven months earlier. rocketblue.ai measured 3M+ citations across 8 major AI models from August 5 to October 29 2025. Reddit peaked at 14.29% of cited sources in ChatGPT in early August 2025 and fell to 0.21% by mid September, a 95% decline in roughly a month, staying below 1% through October. That is a bigger collapse than 2026's 86.4%. By July 2026 Reddit was back at 3.83%. Reddit round trips out of these.

The same page shows 1.50% over the past seven days and 3.15% over thirty, against 0.52% in the collapse window. Neither is evidence of recovery. The seven day window still contains two pre collapse days and the thirty day window contains the entire pre drop baseline, so both averages are consistent with a flat floor plus arithmetic. Promptwatch itself claims no recovery. The honest position is that we do not yet know, and the 2025 precedent is the only reason to expect one.

### The 2025 mechanism, stated honestly

The popular explanation for 2025 runs through Google's removal of the num=100 search parameter in mid September 2025. Search Engine Land and Locomotive place the change on September 12 to 14, with secondary sources saying around September 10, so use the range and not a single day. The measured downstream effects were large: 87.7% of sites lost impressions and 77.6% lost unique ranking terms.

The argument then goes: OpenAI does not crawl Google directly, it buys search data from providers who relied on num=100. With the parameter gone, those providers see only the top 10 to 20 results, and 57.8% of Reddit's ranking keywords sit outside the top 20 per Ahrefs. So most Reddit pages fell out of the feed. Note the correction: the figure circulating as "roughly 60%" is 57.8%.

That chain is plausible and nowhere measured. No source demonstrates it. rocketblue itself does not mention num=100 at all. It offers three candidate causes, declines to pick one, and observes only that "the synchronized nature of the decline across multiple models points to a broader systemic change rather than individual algorithm updates."

## 4. Why these numbers cannot be compared

Three states get collapsed constantly in this literature, and keeping them apart is most of the work.

- **Visible citation**: the domain appears as a source link in the final answer.
- **Retrieved**: the page was fetched during fan out and may never surface.
- **Trained on**: present in the parametric weights, and unmeasurable from outside.

The gap between the first two is enormous. Across 57 conversations, 3,554 retrieved pages produced only 110 citations in final answers, a retrieved to cited rate of 3.1%. Pages that were merely fetched were cited 2.1% of the time. Brands named in the fan out query itself were cited 68.9% of the time. Being retrieved is close to worthless. Being the target of the query is close to decisive.

The gap between citation and mention runs the other way, and it runs differently per engine. In the Semrush AI Visibility Toolkit dataset of 3,981 domain appearances across 115 prompts, 14 countries and 4 engines, 61.7% of citations are ghost citations, where a source is cited without the brand being named. ChatGPT cites 87.0% of the time but names brands only 20.7% of the time. Gemini is the mirror image, naming brands 83.7% of the time and citing 21.4%.

Position Digital's July 2026 dataset shows the inverse failure on ChatGPT specifically: 278 responses across 95 brands produced 552 prose mentions of a brand against only 375 citations to that brand's own domain, and the two overlapped just 185 times. Being recommended and being cited are close to independent events.

This is why the Reddit numbers in circulation cannot be reconciled. Conductor puts Reddit at 1.01% in January 2026 across 238,212 prompts. Promptwatch puts it at 3.83% in July 2026. rocketblue puts it at 0.21% in September 2025. These are not three readings of one quantity. They differ by engine, surface, denominator and prompt mix. Our own standing rule is that any Reddit claim must name its surface and its window, and it exists because three studies inside eight weeks pointed in different directions, with half of one decline reversing inside three weeks.

The same discipline destroys the G2 and Capterra claim. Derivatex ran 40 B2B SaaS categories with one buyer style question each, 10 runs per question, producing 233 software recommendations across 219 tools and 169 analysed cited URLs on one engine. Review aggregators came to 0.9% of citations, and G2 and Capterra each received zero. But 0.9% of 169 URLs is one to two URLs. That zero is a real result from a fragile denominator, and it is contradicted by a larger one. Omniscient Digital, February 2026, measured 25,755 citations including 9,466 bottom of funnel citations, across 200 prompts and 5 engines: G2 sits at 2.09% of BOFU citations, ranked fourth behind Reddit, YouTube and LinkedIn, with Capterra at 0.94%. Publishing the zero means publishing the weaker of two conflicting measurements.

## 5. What actually changed

The largest source family dataset available is TryAnalyze's: 22,295 AI answers, 115,843 citation events, 460 distinct B2B prompts, 37 tracked organisations, across ChatGPT, Perplexity and Google AI Mode.

| Source family | ChatGPT | Google AI Mode | Perplexity |
|---|---|---|---|
| Brand websites and product pages | 68.8% | 37.9% | 35.4% |
| Lists, comparisons and reviews | 13.9% | 33.1% | 34.4% |
| Editorial and educational | 11.9% | 23.8% | 25.2% |
| Community and social | 5.3% | 4.8% | 4.7% |
| Directories and marketplaces | 0.13% | 0.46% | 0.26% |

Directories and marketplaces are 0.13% on ChatGPT. That ends the 17% claim without further argument. And 68.8% of ChatGPT citations going to brand websites and product pages, is the opposite of a flight to established brands.

For page types rather than families, Ahrefs' census of the top 1,000 pages ChatGPT cited in September 2025 found 32.3% influenceable by outreach, composed of educational 19.4%, reviews 5.8%, news and media 5.2%, blog and article 1.9%. The non influenceable remainder is Wikipedia at 29.7%, homepages at 23.8% and the App Store at 6.6%. That is a snapshot of 1,000 URLs from September 2025, not a time series, and it predates every 2026 model change discussed here.

The mechanism that best fits the August 2026 timing is fan out behaviour, and it is properly sourced. Peec AI measured that as ChatGPT 5.6 became the default, single fan out prompts fell from 94.0% to 43.5%, average retrieved sources roughly doubled from about 12 to about 24, and use of the `site:` operator inside fan out queries went from about 0.3% to about 23%. Nectiv, over roughly 4,000 prompts, measured fan outs per prompt rising from 2.17 to 7.61 with `site:` present in 64% of queries. The two disagree on prevalence, 23% against 64%, a difference attributed to API versus consumer interface collection.

The direction is what matters. A surge in site scoped queries mechanically starves a domain that wins on breadth rather than on being a named target. Reddit wins on breadth. Recall that pages merely fetched are cited 2.1% of the time while brands named in the query are cited 68.9% of the time. If the retrieval layer increasingly asks for named domains, an unnamed general purpose domain loses share without anyone having demoted it.

Nobody has connected that mechanism to the Reddit measurement directly. We are not doing so either. We are pointing out that the two are dated to the same week and that no one has run the test.

### The counter argument, taken seriously

Reddit's absence may be temporary for reasons that have nothing to do with quality. Two arguments are worth stating as arguments rather than as findings, because neither is measured anywhere I could source. A retrieval mix can shift again, and a model that answers subjective questions without user generated content has a quality problem pushing back toward including it. What is measured is the 2025 precedent, and it says recovery is the base case rather than the exception. Anyone rebuilding a content strategy around Reddit's permanent removal is betting against the only prior instance of this event.

## 6. What to do

**Measure mention rate, not citation share.** Citation share is the metric that moved and the metric that lies. Position Digital's 552 mentions against 375 citations, with 185 overlapping, means a brand can be recommended constantly and score as invisible. Track how often a model names you in prose. Track citations separately, as a second series.

**Treat citation share as infrastructure dependent.** Both Reddit collapses coincided with a change in retrieval plumbing rather than a change in Reddit. A metric that can move 86% in a week without any change to the underlying content is not measuring the content. Never report a citation share number without the engine, the surface, the denominator and the window attached.

**Diversify across engines, and know that Reddit was never an engine level phenomenon.** Reddit appears 14,267 times in the Consensus Gap dataset yet is universal across engines only 0.1% of the time. Independent measurement puts Reddit at 1.8% on ChatGPT, 2.2% on AI Overviews and 6.6% on Perplexity, and across 98,217 citations in 7 verticals corporate and editorial content is 94.7% of all citations, with Reddit at 2% to 5% in most verticals. A ChatGPT only Reddit collapse is precisely what a single engine phenomenon predicts. The reverse also holds: a strategy that reads as strong on one engine is untested until it is measured on the others.

## 7. The measurement problem underneath all of this

None of these disagreements are resolvable with current instrumentation, and practitioners should price that in.

There is no impression log. We cannot see how often a page was retrieved and discarded, only the 3.1% that surfaced. There is no reliable timestamp, so a model change, an index refresh and a data collection fault all look identical in a time series. And there is often no visible citation at all, which is why 61.7% of citations are ghost citations and why mention and citation have to be tracked as separate series rather than as one number.

The six point post is what happens when those limits are ignored. Five studies with five different denominators get flattened into one narrative with a clean shape. The shape is the giveaway. Real citation data almost never produces round numbers, and it never produces six of them that agree.

## Frequently asked questions

### Did Reddit's ChatGPT citations really fall to zero?

No. Promptwatch measured Reddit at a 3.83% average share of ChatGPT citations from July 18 to August 7 2026, falling to a 0.52% average over August 14 to 17, an 86.4% relative decline. That is a collapse, not a zero. The publisher also states that a data-collection issue cannot be ruled out, discloses no sample size for that specific analysis, and the finding has no independent replication.

### Has Reddit's share collapsed in ChatGPT before?

Yes, and harder. rocketblue.ai measured 3M+ citations across 8 major AI models from August 5 to October 29 2025. Reddit peaked at 14.29% of cited sources in ChatGPT in early August 2025 and fell to 0.21% by mid September, a 95% decline in roughly a month. By July 2026 it was back at 3.83%.

### Do G2 and Capterra get zero citations from ChatGPT?

That zero comes from a fragile denominator. Derivatex analysed 169 cited URLs on one engine, where review aggregators totalled 0.9% of citations, which is one to two URLs. A larger measurement disagrees: Omniscient Digital, February 2026, covered 25,755 citations including 9,466 bottom of funnel citations across 200 prompts and 5 engines, and put G2 at 2.09% of bottom of funnel citations and Capterra at 0.94%.

### Do app marketplaces take 17% of ChatGPT citations?

No traceable study supports that figure. The largest source family dataset available, TryAnalyze's 115,843 citation events across 22,295 AI answers and 460 B2B prompts, puts directories and marketplaces at 0.13% of ChatGPT citations.