Reddit’s share of ChatGPT Search citations was sitting at a steady 3.83% average between July 18 and August 7, 2026. By the August 14-17 window, it had fallen to 0.52%, an 86.4% relative collapse in less than two weeks. Nobody tracking AI search visibility, including the firm that spotted it, is fully sure why.
What the Data Actually Shows
The numbers come from Promptwatch, a GEO analytics firm that tracks which domains ChatGPT cites across sampled queries over time. The decline happened in two distinct steps rather than one smooth slide: a smaller dip on August 8, taking Reddit’s share from the high 3s into the mid 2s, followed by a much sharper collapse on August 14 that pushed it below 1%. A gradual decline would suggest a slow behavioral drift. A two-step drop with a six-day gap between the steps suggests something changed twice, or something changed once and took days to fully propagate through ChatGPT’s citation behavior. Neither explanation is confirmed.
The Leading Theory, and Why It Doesn’t Fully Fit
Promptwatch’s working theory centers on a change in how ChatGPT constructs its background “fanout” queries, the sub-searches it runs before assembling an answer. The firm found that fanout queries using the `site:` search operator jumped from 0.37% to 16.8% of all fanouts, a roughly 46-fold increase, starting August 8. Promptwatch co-founder Klaas Foppen described this as ChatGPT “deliberately going to specific sites to pull information from them,” rather than relying on open web search to surface Reddit threads incidentally the way it may have before.
That theory lines up cleanly with the first, smaller dip on August 8. It doesn’t explain the second, sharper one on August 14, six days later. If a single behavioral change on the 8th were the whole story, the effect should have shown up immediately and then held steady, not waited nearly a week to compound. Search Engine Journal’s own reporting on the data is blunt about this gap: the site-operator change alone “does not explain the sharper decline.”
This Isn’t the First Time Reddit’s Numbers Have Done This
A strikingly similar collapse happened once before, in September 2025. At the time, analyst Kevin Indig argued the real cause wasn’t a Reddit or OpenAI decision at all, it was Google quietly removing the `num=100` search parameter, a URL parameter that let a single search request return 100 results instead of ten. GEO tracking tools had been leaning on that parameter to gather citation data efficiently, and losing it broke how several of them sampled search results, producing what looked like a platform-level citation drop but was actually a measurement artifact on the tooling side. If the same kind of infrastructure dependency broke again this August, the story wouldn’t be about Reddit or ChatGPT changing at all, it would be about the tracking layer itself being fragile in ways that are invisible until something breaks.
Why Citation-Share Numbers Are Shakier Than They Look
Promptwatch itself is upfront that it “can’t yet rule out a data-collection issue on its own end” and is calling the finding provisional rather than settled. That’s a notably honest caveat for a GEO analytics vendor to publish about its own headline number, and it’s worth taking at face value rather than assuming the caveat is just boilerplate. Every AI-citation tracking tool works by sampling a set of queries, observing which domains get cited, and extrapolating a trend from that sample. None of them have access to OpenAI’s actual retrieval infrastructure. A sampling methodology that happens to lean heavily on one search parameter, one query pattern, or one API behavior can produce a dramatic-looking chart from a change that has nothing to do with the platform being measured. That doesn’t mean these tools are useless, it means an 86% aggregate swing is a prompt to check your own domain’s specific numbers before reacting, not a verdict to act on directly.
What This Means If You’ve Been Leaning on Reddit for AI Visibility
If part of a business’s AI-search strategy has involved seeding genuine discussion on Reddit, answering questions in relevant subreddits, building the kind of community presence that shows up in organic threads, this data isn’t a reason to abandon that approach. Reddit’s earlier climb from roughly 0.15% to 3.68% of ChatGPT’s fanout queries between January and May 2026, tracked separately by AI-visibility firm Peec AI across five million query fanouts, reflected a real shift toward AI systems pulling from human, experiential discussion rather than purely branded or editorial content. That underlying preference for authentic, specific, first-person answers over polished marketing copy hasn’t been shown to have reversed, only the visible citation-share number has moved, and the cause of that move is still unconfirmed. Community engagement built for real people, not for a citation-tracking dashboard, doesn’t stop being valuable because one metric wobbled for two weeks.
How to Check Your Own Numbers Instead of the Aggregate Chart
Search Engine Journal’s own advice on this data is the most practical line in the whole story: check your own domain’s citation trend before changing anything, because aggregate tracking data doesn’t necessarily reflect what’s happening to any single site. A small business doesn’t need an enterprise GEO platform to do a rough version of this. Pick five to ten questions a real customer would plausibly ask ChatGPT that relate to the business, the kind of question a search engine query might have answered five years ago, and ask them directly in ChatGPT once a week, noting which sources get cited in the response. It’s manual and it won’t produce a chart, but it directly answers the only question that matters for a specific business: is *this* site’s visibility changing, not some sector-wide average that may be driven entirely by one platform’s fanout-query behavior. Paid tools like Peec AI, Promptwatch, or Profound automate this at scale and add historical trending, worth it once a business has enough at stake in AI-referred traffic to justify the subscription, but the manual version costs nothing and catches the same signal for a single domain.
The Pattern Underneath Both Incidents
What connects this August’s drop to the September 2025 precedent isn’t Reddit specifically, it’s that AI search infrastructure, and the tooling built to measure it, is still young enough that neither has settled into the kind of stability classic SEO tracking has had for two decades. Google’s own search results and the tools that measure rankings against them have had twenty-plus years to shake out their edge cases. ChatGPT’s retrieval behavior, the specific fanout-query patterns it uses, and the third-party tools sampling that behavior to build citation-share charts have had roughly two years, and OpenAI has never published the kind of stable, documented API contract that search-tracking vendors could build against with confidence. A business that watched its Google rankings hold steady for months at a time is going to find AI-citation numbers considerably noisier by comparison, not necessarily because visibility is actually more volatile, but because the measurement layer itself is newer and less proven. That’s a reason for patience with any single week’s chart, not a reason to ignore the space.
A More Durable Way to Think About AI Discoverability
The Peec AI research also found that ChatGPT uses Reciprocal Rank Fusion to combine results across its fanout sub-queries, meaning content that surfaces across multiple related searches for a topic scores higher than content that only appears for one narrow phrasing. That’s a structural preference for breadth of genuine relevance over a single lucky keyword match, and it rewards the same thing good SEO always has: being a real, substantive answer to a question from more than one angle, not a page engineered to catch one specific query string. Chasing whichever platform’s citation share happens to be climbing this month, based on a single analytics vendor’s aggregate chart, is a strategy that will keep whipsawing as tracking methodologies and AI retrieval behavior both keep shifting. Building genuine expertise and community presence across more than one channel, rather than over-indexing on Reddit specifically because a chart showed it trending up in the spring, holds up regardless of which measurement tool’s numbers move next.