The Reddit collapse: Why AI visibility can’t be gamed

On the morning of Monday 18 August 2026, working marketers tracking their AI visibility noticed something odd. Reddit, which had been the single most cited source in ChatGPT for the best part of two years, had almost vanished from the citation pool. Klaas Foppen at Promptwatch posted the chart on X, Ira Bodnar shared a category-level breakdown days later, and LinkedIn caught fire well before any newsroom.

Reddit’s share of ChatGPT Search citations had held steady at 3.83% through late July and early August, and by 17 August it was averaging 0.52%. That is an 86% drop in six days, with no public statement from OpenAI or Reddit to explain it.

The Reddit story is really a case study in what AI visibility actually is, why it moves the way it does, and why the marketers working hardest to hack it are the ones most exposed when it moves.

What actually happened to Reddit in ChatGPT

Reddit’s share of ChatGPT Search citations averaged 3.83% between 18 July and 7 August 2026, then collapsed to 0.52% between 14 and 17 August. That is an 86.4% relative drop in a single week. Google AI Overviews and AI Mode showed only mild, gradual declines over the same window. The drop was noticed first by practitioners tracking their own data, not by any official channel.

AI citation share is the proportion of a given AI system’s cited sources that come from a specific domain, measured across a sample of prompts. If 3.83% of the pages ChatGPT cites when answering a prompt come from reddit.com, Reddit’s citation share in ChatGPT is 3.83%.

Reddit’s citation share moved very differently across the three engines over the same window, per Promptwatch’s own tracking data:

• ChatGPT Search: 3.83% down to 0.52% (an 86.4% relative drop, sharp cliff on 14 August)
• Google AI Overviews: 2.37% down to 2.10% (an 11.3% relative decline, gradual)
• Google AI Mode: 2.22% down to 1.54% (a 30.5% relative decline, gradual)

If Reddit had lost credibility across the AI ecosystem generally, all three engines would have moved together, and only ChatGPT dropped off a cliff. Search Engine Journal has pointed out that the cause is not formally confirmed yet, but the divergence between engines makes the general direction of the story hard to dispute.

Why this matters far beyond Reddit

AI visibility is how often a brand or website is surfaced as a source inside answers generated by AI systems like ChatGPT, Perplexity, Claude, Gemini, and Google’s AI Overviews and AI Mode. The Reddit collapse matters because it shows, in public, that AI visibility on any single platform is not a stable asset. A citation share held for weeks can be revoked in days.

Generative engine optimisation, or GEO, is the practice of shaping content and authority signals so AI systems select your pages when they generate answers. Answer engine optimisation, or AEO, is the same discipline focused specifically on answer engines like ChatGPT and Perplexity.

Reddit had been the single largest citation source in ChatGPT for two years, and plenty of brands had built AI visibility strategies on the back of that fact. When the citation share evaporated, so did every strategy that leaned on it. That is what platform risk looks like in practice.

AI Overviews work in a similar way, pulling from a shifting pool of sources that Google does not publish. The mechanism differs, the volatility is the same. Any AI surface can change what it cites without notice, and no marketer is inside the room when it decides to.

The tactics that just stopped working

A whole set of AI SEO tactics that worked reliably in early August stopped working in mid-August. The most public was seeding Reddit threads with brand mentions in the hope that ChatGPT would cite them. The tactic worked in early August, and by mid-August it was almost entirely ineffective in ChatGPT, while the marketers running it were still being paid to do so.

Black hat AI SEO is the AI-search equivalent of black hat SEO. It covers any tactic that tries to manipulate AI systems into citing a brand through means the platforms would not endorse if they were disclosed. Prompt stuffing is a related idea, meaning inserting brand names or promotional language into content for AI retrieval to pick up rather than for a human reader to benefit from.

None of these tactics were being run in secret. They were discussed at conferences, in newsletters, on LinkedIn. Working marketers were paying agencies and freelancers to run them at scale, and some were paying enterprise tool subscriptions built on top of them.

The tactics that came under threat in August, roughly in order of how public they were:

• Seeding Reddit threads with brand mentions and hoping ChatGPT would cite them
• Comment spam and upvote manipulation on discussion platforms
• Review farming on G2, Capterra, and Trustpilot, which collapsed alongside Reddit in the Bodnar model-version comparison
• Prompt-stuffed content on cheap third-party blogs
• Fake persona posting across forums and comment sections

The return curve on every one of these has been shortening every year, from months down to weeks, and the Reddit collapse compressed the curve down to days.

Why “game the system” strategies fail predictably in AI search

AI search platforms iterate faster than traditional search engines and rarely publish what they change. Retrieval logic moves weekly to monthly. Model versions get swapped in and out with limited notice, and publishers can revoke crawler access at any time. The window in which any single manipulation tactic works has shrunk from the years it lasted in classic SEO to a matter of weeks or, as with Reddit, days.

Query fanout is the process by which an AI system decomposes a single user prompt into several more specific background searches, retrieves results for each, and synthesises an answer. Changes to how fanout works change which sites get cited, even when no ranking algorithm has changed.

On 8 August 2026, Promptwatch measured ChatGPT’s use of the site: operator in its fanout queries jumping from about 0.4% to nearly 17% of all fanout queries in a single day, roughly a 46-fold increase. ChatGPT had started pointing its background searches at specific named websites, in addition to broader open-web searches. The fuller mechanics are covered in our piece on how ChatGPT actually searches the web. Discussion platforms are rarely the target of a site: query, whereas documentation, official brand sites, and institutional domains are.

Multiple GEO practitioners reported that Reddit had added a blanket disallow to its robots.txt around 14 August, cutting off crawler access to the entire domain, in what looks like a publisher decision as consequential as any retrieval change. Reddit has not confirmed this publicly and the timeline is contested. What is not in dispute is that publishers are actively reconsidering their crawler arrangements. The Wall Street Journal reported in July that Reddit was already weighing whether to cut off Google’s AI access, alongside other publishers doing the same.

A comparison shared by GEO practitioner Ira Bodnar showed that model versions themselves change what gets cited. Under GPT-5.1, Reddit accounted for roughly 15% of ChatGPT citations. Under GPT-5.6 Sol, that fell to near zero. Help centres and product documentation rose from 2% to 32%, and established company websites rose from 4% to 18%. None of that appears in a changelog, and marketers only notice by watching their own data.

A black-hat backlink strategy in classic SEO might have worked for eighteen months before Google caught up. AI search does not offer that timeline, and the gap between a tactic working and failing publicly is now measured in weeks.

What durable AI visibility actually looks like

Durable AI visibility is built on first-party, well-structured content published on a domain you control, tracked separately per engine, and diversified across surfaces so no single platform change can wipe it out. Treat it as a discipline that combines editorial judgement, technical hygiene, and continuous measurement. The brands that hold citation share through platform shifts are already doing this before the shift arrives.

Per-engine tracking means measuring AI visibility separately on each AI platform, rather than combining them into a single blended score. Blended scores hide platform-specific shifts, and the Reddit collapse would have been invisible in a metric that averaged ChatGPT and Google AI Overviews together. Durable AI visibility comes from five practical choices, taken together:

1. First-party, single-purpose pages. Structure your site so each priority question your audience asks has a dedicated page that answers it directly. Forum threads and third-party review pages both collapsed in the category shift Bodnar measured. Docs and established company sites gained.
2. Explicit definitions and clean structure. AI retrieval systems extract answers most reliably from content that states what a term means, in one sentence, on first use. Vague writing loses citation share to writing that answers the question outright.
3. Per-engine tracking. Track ChatGPT, Perplexity, Claude, Gemini, Google AI Overviews, and Google AI Mode as separate systems. A blended AI visibility score hides the shifts that matter most.
4. Diversification. Do not build a strategy on citation share from a single platform. The Reddit precedent is enough on its own to make that point.
5. Continuous human oversight. AI search moves faster than any dashboard can react to alone. The professionals watching this every day are the ones who can reroute strategy when the ground shifts.

Semrush’s own analysis confirmed the shape of the Promptwatch numbers days later, and OpenAI has said on record that it has not changed how it selects sources. The mechanism will keep being contested by the platforms and the trackers, while the volatility keeps affecting brands regardless of the explanation.

What to check in your own AI visibility right now

Any brand running AI SEO work over the past twelve months has four things worth checking this week. Whether your own domain gets cited in ChatGPT before and after 14 August. Whether your Reddit-sourced citations specifically have dropped. Whether your GPTBot crawl activity has changed. And whether your pages actually answer questions directly, or bury the answer in a thread of replies.

The practical audit runs in four steps.

1. Compare your own cited URLs before and after 14 August. A broad loss across many pages points to the platform-wide change. A narrow loss concentrated in a few pages or topics suggests something more specific worth investigating separately. Run this in whatever citation-tracking tool you already use, split by engine.
2. Check your GPTBot activity. If ChatGPT is now running domain-scoped searches, your own domain is a more important retrieval surface than it was in early August. That should show up as changed crawl activity in your server logs. If GPTBot is visiting more often, indexed pages matter more. If it is visiting less, that is a different problem to solve.
3. Audit your priority content for single-purpose structure. Look at the highest-value questions your audience asks. Ask whether each has a dedicated page that answers it directly, with the answer near the top, or whether the answer is buried mid-way through a thread of related content. Discussion-style structures are losing citation share. Direct, one-question-per-page structures are gaining.
4. Move any budget currently going into hack-the-model tactics. Payments made to agencies, freelancers or tools for Reddit seeding, comment spam services, review farming, or anything marketed as guaranteeing AI citations should stop. The return on those tactics is now negative more often than positive. Move that budget toward first-party content, per-engine tracking, and expert oversight.

These checks are not exhaustive, but any brand serious about its AI visibility should be able to answer them within a working week.

Why humans are still the point

The Reddit collapse was noticed by working marketers before it was reported anywhere else. Klaas Foppen at Promptwatch surfaced the drop on X. Ira Bodnar published a category-level breakdown days later. LinkedIn discussion moved faster than the press coverage did. AI visibility has become a domain where attention is the scarce input, and attention is a human activity that no dashboard replicates on its own.

The tools that track AI citation share do a good job of reporting what has already changed, but they cannot yet tell anyone what to do about it in real time. The interpretation, the response, and the strategic reroute all still happen in someone’s head.

Reddit’s own on-record response makes the counter-argument better than most agencies could. Adam Collins, Reddit’s chief communications officer, told Axios that the citation drop had “no meaningful impact” on the business because LLMs account for little of its traffic anyway. His fuller line was worth writing down.

For communicators, Reddit’s value is in the real conversations that people are having about your brand today, not where it surfaces tomorrow. If you’re a brand looking at Reddit merely as a way to hack your presence on an AI platform, you risk frustrating your consumers here with no guarantees of a positive result elsewhere.

Steve Rubel of Burson, quoted in the same piece, made the wider argument bluntly. Chasing whichever source or tactic is currently winning in GEO is not the point. Building a diversified information footprint across owned and earned media, and expecting the citation mix to keep moving, is the point. Both arguments arrive at the same conclusion, and the human dimension of AI marketing is the discipline itself.

Why an expert partner is the safer path

AI visibility is not something a brand can safely leave to a general marketing team, a single tool subscription, or a quarterly consultancy engagement. It requires continuous attention, calibrated judgement, and working relationships across a fast-moving ecosystem of platforms, publishers, and data providers. That is the case for working with a specialist agency, not as a defensive purchase but as the sensible allocation of a professional function.

Fifteen years ago the same argument played out with SEO. Businesses that treated it as something anyone in marketing could pick up in a spare afternoon underperformed the ones that treated it as a discipline. The pattern is repeating with AI visibility, on a compressed timeline.

Roar has been publishing on AI search since 2022, and we track proprietary AI visibility data across UK estate agents every month. That work has taught us that the brands who hold their citation share through platform shifts are not the ones running the most tactics. They are the ones with the most consistent editorial standards, the cleanest content architecture, and someone paying close attention when the ecosystem moves.

If your current AI SEO work looks more like a set of tactics than a discipline, that is worth a conversation. Our specialist AI SEO work is built for brands that want to earn AI visibility over the long term rather than manufacture it for the quarter.

Final thoughts

Reddit’s citation collapse in ChatGPT is the case study, not the story. AI visibility is volatile because it depends on retrieval logic and publisher decisions that no marketer controls. The tactics that promise to hack it work for a while and then fail publicly, at which point the brands running them are the ones exposed.

The alternative is a discipline built on first-party structured content, per-engine tracking, diversification across AI surfaces, and the human expertise to keep all of that in tune as the ground shifts. It is slower than a hack, and it lasts.

Anyone wanting to talk through how any of this applies to their own AI visibility can get in touch.

Frequently asked questions

What is AI visibility?

AI visibility is how often, and how prominently, a brand or website is surfaced as a source inside answers generated by AI systems like ChatGPT, Perplexity, Gemini, and Google’s AI Overviews and AI Mode. It differs from traditional SEO visibility because AI systems synthesise answers rather than presenting a list of links, so being cited matters more than being ranked. AI visibility is measured per platform, not as a single blended score, because each platform selects sources differently.

Can you pay to appear in ChatGPT results?

No. Neither OpenAI nor any other AI platform currently offers a paid inclusion product for their generative answers. Sponsored placements (ChatGPT ads) and advertising products exist separately, but citations inside the generated answer are earned through retrieval, not bought. Any vendor claiming they can guarantee a paid citation in ChatGPT is either misrepresenting their offer or using a tactic that will not last.

Is Reddit still worth investing in for marketing?

Yes, but for community reasons rather than for ChatGPT citations specifically. Reddit still has more than 500 million weekly active users, still shapes how customers talk about brands, and still holds citation share in Google’s AI Overviews and AI Mode even after the ChatGPT decline. What no longer holds is any strategy that treated Reddit primarily as a route to ChatGPT visibility.

What is generative engine optimisation?

Generative engine optimisation, or GEO, is the practice of shaping content, structure, and authority signals so AI systems select your pages as sources when they generate answers. It sits alongside answer engine optimisation, or AEO, which focuses on answer engines like ChatGPT and Perplexity specifically. For a fuller view of the discipline, see our GEO service overview.

How often do AI search algorithms change?

More frequently than traditional search algorithms. Retrieval logic on the major AI platforms shifts weekly to monthly. Model versions are swapped in and out with limited public notice, and publisher access can change at any time. The Reddit citation collapse of August 2026 is the recent case study. A domain that averaged 3.83% of ChatGPT citations for weeks lost most of that share in six days, without any official announcement from OpenAI or Reddit.

What should I stop doing if I have been trying to hack AI search?

Stop paying for services or tools that promise guaranteed placement in AI answers, particularly anything that involves seeding third-party platforms, fake accounts, comment automation, or prompt-injected content. These tactics stop working faster than they generate return. Reallocate the budget toward first-party structured content on your own domain, per-engine tracking of your actual visibility, and expert oversight of how the ecosystem is shifting.