Consulted, not cited: how social content shapes AI answers without a link

Questions this guide answers

• Does social media affect ChatGPT answers if ChatGPT doesn't cite it?
• Should we still post on Reddit after ChatGPT cut Reddit citations?
• How do you measure social media's effect on AI answers?
• Do AI models learn about brands from social posts?
• Is citation share the right metric for social media in AEO?

Direct answer

Citation share is the wrong scoreboard for social. A citation is the link an engine chooses to show, and that choice differs by engine and can flip quickly: Reddit's share of ChatGPT's citations dropped within days in August 2026 (Axios). What a buyer acts on is the answer: whether it names your brand, recommends it, and describes it correctly.

Social content reaches the answer through at least 3 paths that need no link. Retrieval: ChatGPT, Claude, and Grok retrieve far more pages than they cite, and part of what they say rests on pages they never show. Third-party naming: most sources ChatGPT and Google AI Overviews cite for product questions are other people's pages, and the shortlist draws on what they say. Model knowledge: with web search off, the brands models from OpenAI, Google, and Anthropic recommend most prominently track how much people search for and discuss them. So value social by the answer, measured by intent and persona across engines and repeated runs.

Why I'm writing this

When Reddit's share of ChatGPT's citations fell in August, the obvious reading was that social had stopped mattering for AI answers. I think that reading mistakes the receipt for the purchase.

I work on how SolCrys measures AI answers, and the first test I apply to any metric is whether it moves with the outcome. Citation share often doesn't: it counts what an engine displays, and it scores a cited thread that names 3 competitors and not you as a win for social. This essay lays out the 3 paths, what to measure for each, and what they mean for a social team's plan.

The scoreboard problem: a citation is a receipt

Citation share answers one narrow question: which URLs did this engine show? It's engine-specific, it's volatile, and it can't see the pages an engine retrieved but didn't show.

In a 30-day sample of AI answers SolCrys collected across B2B and consumer categories, ending Oct. 16, 2026, Google AI Overviews cited social and community platforms far more often than any other engine we track, led by YouTube, then LinkedIn, Reddit, and Facebook. Gemini cited YouTube and Reddit but almost never LinkedIn, ChatGPT rarely cited any social platform, and TikTok almost never appeared. A September 2026 audit of consumer product questions by researchers at Maastricht, Utrecht, and Zurich universities found the same split: ChatGPT's displayed sources were led by technology publishers, while Google AI Overviews most often displayed reddit.com and youtube.com (Uberti-Bona Marin et al.).

So one social program can look strong on Google AI Overviews and absent on ChatGPT on the same day. That describes how each engine displays sources, not whether the program changed what either engine says.

Path 1: Retrieval, or what the engine finds but doesn't show

ChatGPT, Claude, and Grok retrieve far more pages than they cite, and part of what they say rests on pages they never show. In conversation histories donated by 613 users of ChatGPT, Claude, Grok, and DeepSeek, researchers at the Max Planck Institute for Software Systems and other institutions found that ChatGPT cited about 13% of the URLs its searches returned. Across ChatGPT, Claude, and Grok, in both the donated histories and controlled runs, search results that were never cited supported 14–53% of the claims in the answers. Reddit and YouTube dominated the results returned to ChatGPT and Grok, and were absent from Claude's (Amani et al., September 2026).

So a thread or video can supply the use cases, comparisons, and objections in an answer without appearing under it. And a drop in citations alone can't tell you whether an engine stopped retrieving a platform or only stopped showing it; that needs the pages it consulted, not the pages it displayed.

Path 2: Third-party naming, or who the cited pages name

Most of what ChatGPT and Google AI Overviews cite for product questions is other people talking about brands, and the answer's shortlist draws on it. Google says its generative AI features "can show what's being said about products and services across the web, including in blogs, videos, and forum discussions" (Google's May 2026 guide). Those features retrieve pages from Google's Search index through its core ranking systems, so on Google, answer engine optimization (AEO) is an operating layer on top of SEO. When OpenAI launched shopping research in ChatGPT, Modern Retail reported it was trained to prioritize "organic content, including review sites and even Reddit" and to avoid sites that are "overly commercial" (Modern Retail). In the September audit, brands' own sites were about 1 in 6 of the domains Google AI Overviews displayed.

SolCrys's reading of cited LinkedIn posts found the same for category questions: they came almost entirely from people with no tie to the brands being compared (LinkedIn AEO guidelines). A post doesn't need to be cited to matter if it's the reason a cited review, roundup, or thread names you.

Path 3: Model knowledge, or what the model already believes

Before an engine searches, and when it doesn't search at all, the model answers from what it learned in training, and that tracks how much a brand is discussed. A 2023 study found that a model's ability to answer a fact question relates to how many training documents mention the entities involved (Kandpal et al., ICML 2023). In September 2026, researchers at Northwestern University and Boston University asked 6 models from OpenAI, Google, and Anthropic to recommend brands in 5 consumer categories with web search off. The brands named most prominently tracked Google search interest most closely, then online brand conversation, more closely than conventional brand popularity or ad spend (Malthouse et al.). The study compares brands at one point in time; it doesn't test what happens when conversation about a brand grows.

This path is slow, since it moves when models are retrained, but durable: it shapes how a brand is described when no page about it is in view. It concerns training, not ChatGPT's per-user memory feature.

3 paths, 3 measures

Each path has its own lever and its own proxy, so each needs its own line in the plan.

What this means for social teams

Post for the answer, not the impression count.

• Answer the questions buyers ask. Work from intents and personas, not trending topics, and answer one question per post with the claim first. Buyer-style questions by industry make a useful starting list (what buyers ask AI).
• Post as named people, with first-hand evidence. Google's guide notes that "a first-hand review provides a unique perspective based on personal experience," which a brand account rarely offers. On LinkedIn, the posts cited for category questions come mostly from independent voices, so give them something specific to repeat: a number, a method, an honest limit.
• Disclose every time. The FTC's guidance for employees is plain: "Listing your employer on your profile page isn't enough" (FTC). Put the disclosure in the post.
• Earn mentions on the pages engines already cite. For your priority prompts, find the reviews, roundups, threads, and videos each engine cites, and give their authors access, data, or an honest comparison (source-layer strategy).
• Match platform to engine for citations. In our sample, Google AI Overviews cites YouTube, LinkedIn, and Reddit; ChatGPT rarely cites any of them. See the YouTube, LinkedIn, and Reddit guides, and judge influence by the answer on each engine.

What not to do

Every path above rewards real discussion by real people, and faking it breaks platform rules and Google's policies.

• Seeded or undisclosed paid mentions. Google says "seeking inauthentic 'mentions' across the web isn't as helpful as it might seem," and its spam policies now cover "attempting to manipulate generative AI responses in Google Search" (what crosses the line).
• Engagement pods. LinkedIn's Professional Community Policies tell members not to "agree with others ahead of time to like or re-share each other's content." Engagement isn't what gets a post cited anyway: a meaningful share of the LinkedIn posts engines cite drew almost none.
• One post, many accounts. Identical text across employee profiles reads as coordination, not discussion.

How to measure social's effect on AI answers

Measure the answer, then trace it to its sources. Platform analytics count impressions and engagement, not influence on AI answers.

• Fix a prompt set by intent and persona, and freeze it. One prompt per intent and persona that matters, phrased the way that buyer asks (why intent and persona are the unit).
• Record 3 things per answer. Whether you're named, how favorably you're positioned (Recommendation Score), and whether the description is correct.
• Record the sources and who they name. For each cited page, note the platform and whether it names you, competitors, or neither. On prompts that don't name your brand, compare your mention rate when a cited page names you with your rate when none does. The second rate is presence your displayed sources don't explain; retrieval and model knowledge are the likely paths, and telling them apart takes more evidence.
• Read each engine separately, across runs. Answers vary from run to run, so judge a change across several runs, not one (how many runs a number needs).
• Credit a campaign with a control. Check whether untouched prompts on the same engine moved too, because an engine-side change moves them as well.

How SolCrys fits

SolCrys measures the answer, not just the link. Each workspace runs a frozen Golden Prompt Set, built from search demand, the questions people ask around it, and community discussion where available, on ChatGPT on the free plan and on Gemini, Google AI Overviews, Perplexity, and Claude on paid plans. For every answer it records whether you're named, grades how favorably with the Recommendation Score, and flags which cited pages mention you and which name competitors. Signals surfaces community items weekly.

That's the Loop: Measure → Diagnose → Execute → Verify. When the gap is a source problem, Workspace Deep Analysis can recommend a platform post with governance guidance: post from your own account, disclose your affiliation, use first-hand data, and skip engagement pods. Execution stays governed: your team writes and publishes, and SolCrys reruns the same prompts so you can judge across runs whether the answer moved. Start Free to see what ChatGPT says for 10 of your buyer questions.

About the author

Eason Wang is Co-Founder & CPO of SolCrys. He holds a PhD in Machine Learning and has spent 18+ years building enterprise products, starting at Microsoft Research Asia. His current focus is agentic AI workflows and the infrastructure required to run them in production. Connect on LinkedIn.

Written October 2026. For the companion argument that being cited isn't being recommended, see Cited but Not Recommended.

Sources

• Google Search Central: guide to optimizing for generative AI features (May 2026, updated July 10, 2026)
• Google Search Central: spam policies for Google web search
• Amani et al.: Characterizing Web Search by Conversational LLM Agents, arXiv:2609.19244 (Sept. 2026)
• Uberti-Bona Marin et al.: Auditing AI-Generated Product Recommendations, arXiv:2609.18729 (Sept. 2026)
• Malthouse et al.: Evaluating Brand Retrieval and Ranking in Large Language Model Recommendations, arXiv:2609.16304 (Sept. 2026)
• Kandpal et al.: Large Language Models Struggle to Learn Long-Tail Knowledge, arXiv:2211.08411 (ICML 2023)
• Axios: Reddit fades from ChatGPT citations (Aug. 20, 2026)
• Modern Retail: OpenAI's new ChatGPT shopping tool (Nov. 24, 2025)
• LinkedIn Professional Community Policies
• FTC: Endorsement Guides, What People Are Asking

FAQ

Related guides