Author: Shahbaz Alam

  • Reddit Citations in ChatGPT Are Collapsing – But ChatGPT Isn’t Searching Reddit Less

    Reddit Citations in ChatGPT Are Collapsing – But ChatGPT Isn’t Searching Reddit Less

    Analysis of more than 2,400 prompts reveals a striking divergence between how ChatGPT searches Reddit and how often it ultimately cites it.

    Reddit has historically been one of the most visible sources in AI-generated answers, particularly for questions involving recommendations, reviews, opinions and first-hand experiences.

    But recent data from a B2B SaaS company suggests that ChatGPT’s relationship with Reddit may be changing.

    We analysed more than 2,400 tracked prompts across three consecutive weekly periods in August 2026, looking at two different layers of ChatGPT’s search behaviour:

    1. The fanout queries generated as ChatGPT researches a prompt.
    2. The citations that ultimately appear in the final answer.

    The two signals tell a surprisingly different story.

    Reddit citations collapsed.

    But Reddit-related searching in fanout queries did not.


    Reddit citations fell from roughly 26% of prompts to 3%

    Across the three-week period, the proportion of tracked prompts that received at least one reddit.com citation changed approximately as follows:

    WeekPrompts receiving a Reddit citation
    Week 126.1%
    Week 211.9%
    Week 33.0%

    That’s an approximately 89% decline in the share of prompts citing Reddit between the first and third weeks.

    And this wasn’t simply a one-week anomaly.

    The decline happened consistently across the three-week period.

    More importantly, the change was broad-based rather than concentrated in one particular topic.

    Prompts covering SEO platforms, reporting, content, rankings, enterprise SEO, AI search visibility and other areas all showed the same general pattern.


    But ChatGPT didn’t simply stop searching Reddit

    This is where the analysis gets interesting.

    One obvious explanation for the citation decline would be that ChatGPT had simply reduced the number of searches it sends to Reddit.

    Our data doesn’t support that explanation.

    When we analysed the underlying fanout queries across the full tracked population, Reddit targeting remained broadly stable:

    Reddit fanout signalWeek 1Week 2Week 3
    site:reddit.com0.17%0.75%0.62%
    Any Reddit mention~1.2%~1.4%~1.2%

    While Reddit citations fell from 26.1% to around 3% of prompts, Reddit-related fanout activity stayed at roughly the same level.

    That creates an important distinction:

    ChatGPT appears not to be citing Reddit less simply because it stopped looking for it.

    Instead, something may be changing between retrieval and final source selection.


    A closer look at Reddit-citing prompts

    We also analysed a more specific subset: prompts whose ChatGPT responses cite reddit.com.

    Within this subset, the fanout behaviour changed dramatically between early and late August.

    Metric1–8 Aug16–22 Aug
    Seed Prompts62866
    Avg. Fanout queries per Prompt1.033.2
    Maximum fanout depth36
    site:reddit.com queries0.6%12.3%
    Any Reddit mention4.5%22.3%

    The number of seed prompts fell sharply, but ChatGPT explored each remaining topic much more deeply.

    The average number of Fanout queries per Prompt increased by more than 200%, while maximum fanout depth doubled.

    At the same time, explicit site:reddit.com searches increased by roughly 20× within this particular prompt subset.

    So we have a potentially important divergence:

    For prompts associated with Reddit citations, ChatGPT became more aggressive in targeting Reddit during its research process, while Reddit’s presence in the final citation layer declined sharply.


    ChatGPT’s search behaviour also became much more structured

    The Reddit change occurred alongside a much broader shift in fanout behaviour.

    Search operators were rarely used during the earlier period but became common later:

    Search behaviour1–8 Aug16–22 Aug
    site:0.9%53.1%
    Quoted phrases0%50.2%
    OR0%4.3%
    Exclusions (-)0%3.8%
    Any operator0.9%66.4%

    The fanout also shifted towards more focused competitive and review research.

    Fanout Queries mentioning named Brand rose substantially, while searches focused on sentiment and third-party review sources also increased.

    This indicates that ChatGPT’s research process was becoming more targeted, domain-specific and investigative, rather than simply issuing broad natural-language searches.


    Reddit isn’t disappearing completely

    The data doesn’t suggest that ChatGPT has completely stopped using Reddit.

    In fact, the prompts that continued to receive Reddit citations shared an interesting characteristic.

    They were disproportionately prompts that explicitly asked for community or review content, such as requests involving Reddit, G2, Trustpilot or other peer-review sources.

    In other words, Reddit appears much more resilient when the user specifically asks ChatGPT to investigate forums or communities.

    The decline is therefore more pronounced in general-purpose prompts, where Reddit previously appeared as one of several incidental sources.

    That distinction is important.

    It suggests Reddit may be losing its position as a default supporting source, rather than being completely excluded from ChatGPT’s research process.


    Where did the Reddit citations go?

    We then looked at the most obvious alternative sources.

    G2’s share increased modestly, but other major review and forum domains did not show anything close to the increase necessary to explain Reddit’s decline.

    For example:

    DomainWeek 1Week 2Week 3
    G28.44%10.22%10.85%
    Trustpilot0.91%1.62%1.45%
    Capterra2.49%4.24%2.91%
    Quora~0%~0%~0%
    Stack Exchange0.33%0.17%0.50%

    So this does not look like a simple case of Reddit being replaced by one particular review platform.

    There is another factor.


    The overall citation pool became much larger

    Average citations per prompt increased substantially over the same period.

    The prompts that lost Reddit citations were not necessarily receiving fewer citations overall.

    Instead, ChatGPT was often citing more sources per answer.

    Reddit’s share of that growing citation pool therefore fell significantly.

    This distinction matters:

    Reddit’s declining citation share does not necessarily mean ChatGPT is producing fewer sourced answers.

    It may mean that ChatGPT is selecting a broader range of other sources while giving Reddit a smaller role in the final citation mix.


    Reddit citations became concentrated

    Another interesting signal emerged when looking at the total number of Reddit citation instances.

    Although the number of prompts containing a Reddit citation fell by almost 90%, the total number of individual Reddit citation instances fell much less.

    That means Reddit citations became increasingly concentrated.

    Earlier in the period, Reddit could appear as an incidental citation across a wide range of prompts.

    Later, Reddit was much more likely to appear when a prompt specifically called for community content and when it did appear, multiple Reddit pages could be cited within the same answer.

    In simple terms:

    Reddit went from being a background source across many answers to a concentrated source within a much smaller number of answers.


    What does this mean for AEO and GEO?

    The most important lesson from this research may have little to do with Reddit specifically.

    It highlights a fundamental distinction between retrieval and citation.

    When an AI search engine researches a question, a source can potentially move through several stages:

    Query → Retrieval → Evaluation → Synthesis → Citation

    A website can therefore be discovered without ultimately being cited.

    That means measuring only final citations can hide important changes in the underlying search process.

    The data suggests that this is particularly important for Reddit.

    Across the full prompt population:

    Reddit querying remained broadly stable.

    Yet:

    Reddit citations collapsed.

    Within the Reddit-citing subset:

    Reddit targeting actually increased sharply.

    That makes a simple “ChatGPT stopped searching Reddit” explanation difficult to reconcile with the data.


    What changed inside ChatGPT?

    We don’t yet have enough evidence to determine the exact mechanism.

    Possible explanations include changes to:

    • source ranking
    • trust or quality assessment
    • relevance filtering
    • duplicate-content handling
    • citation selection
    • how retrieved user-generated content is incorporated into final answers
    • the broader search and citation system

    At this point, these are hypotheses rather than confirmed explanations.

    The data tells us that the relationship between retrieval and citation changed.

    It does not tell us precisely which internal mechanism caused it.


    The bigger lesson for AI visibility

    For SEOs and GEO practitioners, this is a useful reminder that AI visibility is not one metric.

    It may be necessary to monitor at least two separate signals:

    1. Retrieval visibility

    Is the AI search system actively looking for and retrieving your content?

    2. Citation visibility

    Does your content ultimately survive the model’s selection process and appear as a citation in the final answer?

    Reddit’s recent behaviour demonstrates why the distinction matters.

    A source can remain relevant to the AI’s research process while becoming significantly less visible in the final answer.


    The next question

    The next stage of this research is to determine whether the same pattern exists beyond Reddit.

    Are other forums, communities and user-generated-content platforms experiencing the same divergence between retrieval and citation?

    And does the pattern appear across other AI search engines, or is it specific to ChatGPT?

    Those answers could tell us whether we’re looking at a Reddit-specific change or a broader shift in how AI search systems evaluate and cite user-generated content.

    For now, the clearest conclusion from the data is this:

    ChatGPT doesn’t appear to be simply searching Reddit less. It appears to be becoming much more selective about when Reddit survives into the final answer as a cited source.

    For anyone measuring AI visibility, that’s an important distinction.

    Being found is one thing. Being cited is another.

  • Serving Markdown to Search & AI Bots: A Practical Strategy for the Agentic Web

    Serving Markdown to Search & AI Bots: A Practical Strategy for the Agentic Web

    As AI-powered search, assistants, and autonomous agents become more common, website owners are looking for ways to make their content easier for machines to consume.

    One question comes up frequently:

    Should websites provide content in Markdown for AI systems?

    The short answer is yes, but how you do it matters.

    Why Markdown Matters

    AI systems can read HTML, but HTML often contains a lot of extra information such as navigation, styling, scripts, and layout markup.

    Markdown provides a cleaner, more structured version of the same content, making it easier for AI systems to process while reducing unnecessary overhead.

    Common Approaches

    There are three main ways websites can provide Markdown:

    1. Bot Detection

    Some websites detect AI crawlers and serve Markdown only to those bots.

    While possible, this approach requires ongoing maintenance because user agents change frequently and new AI bots appear regularly.

    2. Separate Markdown Pages

    Another option is to create dedicated Markdown resources, such as:

    • llms.txt files
    • Separate .md pages
    • Parallel Markdown URL structures

    The downside is that you now have two versions of the same content to maintain, increasing complexity and the risk of content becoming inconsistent.

    3. Content Negotiation (Recommended)

    A cleaner approach is to use HTTP content negotiation.

    With this method:

    • Browsers continue to receive HTML.
    • AI agents can request Markdown.
    • The same canonical URL is used for both formats.
    • Only one source of truth needs to be maintained.

    For example, an AI agent can request:

    Accept: text/markdown

    And the server responds with:

    Content-Type: text/markdown

    This follows existing web standards and avoids duplicate content issues.

    Why It’s SEO-Friendly

    Content negotiation keeps the canonical URL unchanged and does not create separate indexable pages.

    Search engines continue to receive standard HTML, while AI systems can access Markdown when they explicitly request it.

    This makes it possible to support both humans and machines without affecting existing SEO foundations.

    Growing Adoption

    Support for Markdown delivery is growing across the web.

    Companies such as Cloudflare and Vercel have introduced solutions that use content negotiation to provide AI-friendly content, and some AI tools are already requesting Markdown through HTTP headers.

    Final Thoughts

    If you’re exploring ways to make your website more AI-friendly, content negotiation is currently one of the simplest and most scalable approaches.

    Rather than maintaining separate AI content layers, you can serve Markdown only when requested while keeping a single canonical URL and source of truth.

    It’s a standards-based solution that helps prepare your website for the growing role of AI in content discovery and consumption.