# Where ChatGPT, Perplexity, Claude and Gemini get their answers > ChatGPT favors Wikipedia, Perplexity favors Reddit. Tables are cited 2.5×, fresh content 3.2×. Here's where engines pull citations from and how to make a page 'extractable'. _Source: https://seomatrix.ai/blog/ai-citation-sources/ · Updated: 2026-07-03_ --- ## Different engines, different sources Rule one: track platforms separately, because they cite differently. The AIO vs AIM difference matters in practice. **AI Overview** is closer to classic search: ~60% of sources from the original query, the top 3 taking ~40% of slots, often single-language sources, one meaning of an ambiguous word. **AI Mode** is vector-based: only ~15% from the original query, a flatter source histogram, multilingual sources, all meanings of a word surfaced. The implication: presence on Wikipedia and topical Reddit/forums isn't "PR for show" — it's a direct channel into specific engines' citations. ## What actually gets cited: extractability A page now has two independent jobs — to **rank** and to be **"ready to be pulled."** The answer engine chops the page into H2-level blocks itself, so every block has to survive that chopping. The formula for a self-contained block (an Extractable Trust Block): - **Claim** — an unambiguous direct answer in the first sentence, no preamble. - **Explanation** — why it's so. - **Boundary of applicability** — where it works and where it doesn't. - **Example or source** — factual confirmation. > **The killers of extractability** — Unexplained pronouns ("it", "this method"), references like "as stated above", "below we'll cover." The Trust classifier flags categorical claims with no boundary of applicability as *low cite-safety* and excludes them from generative results. What boosts citation, per 2026 data: - An **atomic 40–60-word answer with numbers** right after the subheading. - **Tables are cited 2.5× more often** — structured data is easier to extract. - **FAQPage markup** gets pulled into citations roughly **2.3× more often**. - A **direct answer in the first two sentences** of a block is exactly what the engine "pulls." - H2/H3 subheadings as self-contained question-anchors; after a table/list, a one-sentence takeaway in plain text. And important context: **ChatGPT cites only ~15% of retrieved pages** — the other 85% are evaluated and discarded. Structure and extractability decide whether you land in that 15%. ## Freshness: the 90-day cliff AI search has a documented **~3-month citation "cliff"**: content not updated for ~90 days falls rapidly out of answers. - **50% of all AI-cited material is under 13 weeks old** (Salespeak) — a "three-month shelf life." - An analysis of **35M AI Overviews**: AI-cited content is on average **25.7% fresher** than ordinary organic. - David Baum: a 60-day refresh cycle on the top 20 pieces → **7 of 20 entered AI citation** after the first update, despite not being cited before at all. - Net effect: fresh content is cited roughly **3.2× more often**. > **Put the date in the visible HTML** — AI crawlers read HTML, not your CMS database. Without a visible "Updated [month year]" the freshness signal doesn't transmit. And gaming it with a date swap doesn't work — Google distinguishes "artificially refreshed" content and penalizes it. You need substantive edits: new stats with a year, a new example, a new quote. ## How to check it on your own site We built this into the audit and the visibility probe: - **A 4-engine probe** — Perplexity, Google AIO, **Claude and Gemini** (via OpenRouter): per query you see who's cited and from which sources (for Gemini, with the real domain resolved behind the grounding redirect). - **Third-party presence** — a check of the brand's presence on Wikipedia (ChatGPT's #1 source). - **Answer-capsule / structure-audit** — an offline check of whether each section has a self-contained atomic answer. - **Freshness-audit** — catches stale dated facts in the article body. > **Check a page** — Run a URL through the [free check](https://seomatrix.ai/audit) — you'll see [the page's citation-worthiness](/tools/geo-check/) and what's stopping the engine from "pulling" it. ## What to do 1. **Write blocks that survive being cut out.** Direct answer in the first sentence + boundary of applicability + example. 2. **Give structure.** Tables and FAQPage are the most-cited formats; follow a table with a takeaway sentence. 3. **Be present where the engine sources from.** Wikipedia for ChatGPT, topical communities for Perplexity. 4. **Keep a registry and refresh on a 60-day trigger.** Substantively, with a visible date in the HTML — not rewriting from scratch. ## Sources - **AI citation factors (2026 data)** — [Digital Applied: AI Search Citation Ranking Factors](https://www.digitalapplied.com/blog/ai-search-citation-ranking-factors-2026-data-study), [PPC Land: 23 factors that get content cited by AI](https://ppc.land/23-factors-that-actually-get-your-content-cited-by-ai-search-engines/). - **Extractability and cite-safety (AEO)** — GEO-HowTo / DrMax SEO-channel 2026 (posts 13.01, 17.02, 05.03.2026): the self-contained-block formula, the Trust classifier, LLM Memory Anchors. - **The 90-day freshness cliff** — DrMax SEO-channel 2026 (post 29.05.2026, citing David Baum); third-party: Salespeak, the 35M AI Overviews analysis. - **AIO vs AIM — different engines** — US v. Google case records (AI Overviews architecture: customized Gemini + Fast Search); GEO-HowTo 2026, Kurdyukova–Volovich talk (AIO vs AIM measurements). ## FAQ ### Do different AI engines really cite from different sources? Yes. In 2026 data, ChatGPT's top citations are dominated by Wikipedia (~47.9%), while Perplexity's are dominated by Reddit (~46.7%). And Google's AI Overview (AIO) and AI Mode (AIM) work differently: AIO is closer to ordinary search (≈60% of sources from the original query, often single-language), while AIM is vector semantic search (≈15% from the original query, multilingual sources). Track engines separately. ### What makes a passage 'extractable'? Self-containment: a block must keep 100% of its meaning if cut out of the article. The formula is a claim (a direct answer in the first sentence) + explanation + boundary of applicability + an example or source. The killers of extractability are unexplained pronouns ('it', 'this method') and references like 'as stated above'. ### How important is content freshness for citation? Critical. AI search has a documented ~90-day cliff: content not updated for ~3 months is cited far less. 50% of all AI-cited material is under 13 weeks old (Salespeak); fresh content is cited roughly 3.2× more often. But swapping only the date without substantive edits is penalized.