ChatGPT Search and Perplexity choose citations through a four-stage retrieval pipeline called query fan-out. When a buyer submits a prompt, OpenAI and Perplexity rewrite that conversational question into multiple short keyword queries, retrieve candidate web pages, split each page into passages of 75 to 100 tokens, and attach inline links to the highest-scoring chunks. Neither ChatGPT nor Perplexity returns the same answer on every run because probabilistic token sampling, fan-out query shifts, account memory, and IP location change which passages win.
Kevin Indig's February 2026 Growth Memo analysis of 1.2 million verified ChatGPT citations and Botspot's dataset of 6,005 bot.radar runs from October 13 to November 12, 2026 show that citation selection follows measurable structural patterns. Pages that front-load entity-rich definitions, semantic HTML tables, and structured AI Appendices earn inline links far more consistently than dense B2B blog posts.
How do query fan-out and passage chunking work in ChatGPT and Perplexity?
Query fan-out is the retrieval step where an AI answer engine expands one conversational prompt into multiple short keyword searches before generating a response. Instead of matching a 25-word buyer question verbatim against a web index, ChatGPT Search and Perplexity decompose that prompt into targeted sub-queries covering product categories, pricing tiers, integrations, and competitor comparisons, then fetch candidate URLs in parallel.
- Prompt decomposition into sub-queries. OpenAI's ChatGPT Search and Perplexity parse the user prompt and emit two to five terse keyword queries, adding category synonyms, comparison terms, and the current year 2026 when recency matters.
- Web index retrieval. ChatGPT Search pulls candidate URLs through Microsoft Bing and OpenAI's
OAI-SearchBotcrawler, Perplexity queriesPerplexityBotand its own search index, and Google AI Overviews and Google AI Mode draw from the core Google Search index. Cleanrobots.txtrules and an `llms.txt` index help crawlers find canonical pages without wasting crawl budget. - Passage chunking at 75 to 100 tokens. Retrieval pipelines strip site navigation and slice page HTML into passages of roughly 75 to 100 tokens, or 60 to 80 English words, encoding each passage into a dense embedding index for semantic similarity scoring.
- Reranking and grounded synthesis. A reranker scores candidate chunks on semantic fit, entity specificity, and freshness before the language model writes its answer and attaches inline citations to passages carrying verifiable claims.
Why does passage chunking break dense B2B blog posts?
Passage chunking breaks dense B2B blog posts because retrieval models score isolated 75-word slices rather than reading a 2,000-word article from top to bottom. When a paragraph opens with a bare pronoun instead of naming its product and category, that slice loses its subject once separated from the heading above. Chris Green tested passage retrieval in a June 2025 controlled experiment using 100-token chunks with 20-token overlap on Google Vertex text-embedding-005 and HNSW indexing.
Chris Green rewrote three articles into dense prose, structured headings with lists, and question-and-answer passages across four chunking pipelines: token-based, LangChain recursive, HTML-aware h1 through h4, and semantic breakpoints. Question-and-answer chunks achieved the highest query-to-chunk cosine similarity in every scenario, structured headings with lists finished close behind, and dense prose scored lowest every time. Green's three-article, five-query sample is illustrative and measures retrieval similarity rather than final citations, yet the chunking math explains why self-contained passages win.
Does ChatGPT give the same answer to everyone?
ChatGPT does not give the same answer to everyone, and ChatGPT rarely returns identical brand order or citations across two consecutive clean runs. Answer variance is a built-in property of retrieval-augmented language models driven by probabilistic token sampling, dynamic query fan-out, account memory, and locale routing. Pasting a buyer prompt into a personal ChatGPT Plus account once and sending a screenshot to leadership measures a sample size of one.
- Probabilistic token sampling. ChatGPT, Perplexity, Gemini, and Claude sample output tokens from probability distributions at non-zero temperature, shifting brand order, phrasing, and cited URLs between runs even when the prompt stays identical.
- Fan-out query variation. Because the model writes its own search sub-queries on each turn, one run queries
"B2B AI visibility software 2026"while the next run queries"ChatGPT citation tracking tools pricing", retrieving different candidate pages. - Account memory and chat history. Logged-in ChatGPT Plus, Team, and Enterprise workspaces retain user memory and prior prompts about a marketer's own company, biasing answers toward brands discussed earlier in that account.
- Geographic and language routing. A buyer prompting from
en-USin New York and a buyer prompting fromfi-FIin Turku reach different regional search results, local currencies, and much thinner Finnish training corpora.
How should marketing teams measure volatile AI answers instead of taking screenshots?
Marketing teams measure volatile AI answers by running a fixed cohort of buyer prompts repeatedly inside clean, unpersonalized sessions across engines and locales. bot.radar executes scheduled runs across ChatGPT, Google AI Overview, Google AI Mode, Gemini, Perplexity, Claude, and Copilot, including tracked diagnostic prompts such as 192d3adc and e967d22b, and reports three statistical metrics. Visibility rate is the share of runs naming a brand, citation rate is the share linking to a domain, and share of voice is a brand's share of all brand mentions, giving growth teams and B2B brands defensible trend lines.
What do ChatGPT and Perplexity cited pages have in common?
Pages cited by ChatGPT and Perplexity share five measurable traits: front-loaded conclusions, high proper-noun entity density, explicit subject-first definitions, question headings, and semantic HTML tables. On February 16, 2026, Kevin Indig published "The science of how AI pays attention" in Growth Memo, analyzing 1.2 million verified ChatGPT citations alongside positional data on 18,012 citations and paragraph-level metrics across 1,000 heavily cited pages. Paired with research from Chris Long at Nectiv, AirOps, and Chris Green, Indig's findings show how cited chunks differ from average web copy.
| Citation factor | Study and sample | Measured finding | Practical page rule |
|---|---|---|---|
| Document position, ski ramp | Kevin Indig, Growth Memo, Feb 2026, 18,012 ChatGPT citations | 44.2% of citations come from the first 30% of a page, 31.1% from the middle, and 24.7% from the last 30% | Place 3 to 6 Key Takeaways and the direct answer in the top 20% of the page, plus a closing summary |
| Proper-noun entity density | Kevin Indig, Growth Memo, Feb 2026, 1,000 heavily cited pages | 20.6% entity density in cited text vs 5% to 8% in Brown Corpus and Penn Treebank English | Name specific brands, products, people, standards, years, and metrics instead of generic nouns |
| Definitive wording | Kevin Indig, Growth Memo, Feb 2026, 1.2M ChatGPT citations | 36.2% of cited sentences use explicit definitions vs 20.2% of non-cited text, a 1.8x lift | Write subject-first "X is Y" definitions near the first mention of every core concept |
| Question headings and echo | Kevin Indig, Growth Memo, Feb 2026, 1.2M ChatGPT citations | Cited text is 2x more likely to contain a question mark, and 78.4% of question citations come from headings | Phrase H2s as literal buyer questions and repeat the heading subject in the first words of the answer |
| Semantic HTML tables | Chris Long, Nectiv, B2B SaaS set vs 25,000 Google ranking URLs | 30% of top ChatGPT-cited B2B SaaS pages contain a <table> tag vs 13% of Google ranking pages, 2.3x | Publish comparisons, specs, and pricing in real HTML <table> tags rather than images or div grids |
| Passage chunk formatting | Chris Green, June 2025 test, 4 chunkers on Vertex text-embedding-005 | Q&A chunks scored highest on query similarity across all 4 chunkers; dense prose scored lowest | Keep answers within 75 words of their heading, cap paragraphs at 120 words, and cut pronoun openers |
| Commercial content recency | AirOps, August 2025, 4,000+ ChatGPT-cited pages across 900 queries | 53.4% updated within 6 months overall; 60%+ within 6 months and 83% within 12 months on commercial queries | Review commercial SaaS pages every 3 to 6 months with substantive edits and matching dateModified schema |
Why do entity density, front-loading, and HTML tables lift ChatGPT citations?
Entity density, front-loading, and HTML tables lift ChatGPT citations because concrete nouns lower language-model perplexity inside the top-of-page zones crawlers extract at the highest rate. In Kevin Indig's February 2026 Growth Memo study, 44.2% of ChatGPT citations came from the first 30% of the page, making opening sections roughly 2.5x more likely to earn a citation than middle paragraphs. Inside individual paragraphs, 53% of citations came from middle sentences versus 24.5% from first sentences and 22.5% from closing sentences, showing ChatGPT extracts whichever sentence carries the highest standalone information gain.
Heavily cited chunks in Indig's Growth Memo dataset also averaged Flesch-Kincaid grade 16 against grade 19.1 for non-cited text, alongside a balanced 0.47 subjectivity score on a 0-to-1 NLP scale. Economist-style analyst prose that pairs a checkable number with its business implication beats both dry glossary lists near 0.1 subjectivity and promotional hype near 0.9. On B2B SaaS comparison queries, Chris Long at Nectiv found real HTML <table> tags on 30% of top ChatGPT-cited pages compared with 13% across 25,000 Capterra-query Google results exported from Ahrefs, though Long notes that correlational finding reflects B2B comparison intent.
How much does content freshness really affect AI citations?
Content freshness affects AI citations most strongly on commercial software evaluation prompts, though vendor studies need careful methodological caveats. AirOps's August 4, 2025 report on 4,000 ChatGPT-cited pages across 900 queries in 15 industries found 35.2% were updated within three months, 53.4% within six months, and over 60% of commercial pages within six months. Because AirOps sells content-refresh workflows, included no non-cited control set, and printed conflicting 26.2% and 56.4% one-year staleness numbers, treat a three-to-six-month SaaS review cycle as directional hygiene and never bump dateModified without updating real facts.
How do ChatGPT and Perplexity citation rates differ in real B2B tracking data?
ChatGPT and Perplexity cite owned brand domains at sharply different rates even across identical buyer questions. Across 6,005 production runs in bot.radar from October 13 to November 12, 2026, ChatGPT inline-cited bot.spot in 1.95% of runs on 3.19% brand visibility. Across that same 30-day prompt cohort, Perplexity recorded a 0.38% owned-domain citation rate on 2.49% visibility, routing a much larger share of its footnote links to third-party software roundups, review directories, and independent comparison posts.
Across all seven engines tracked in that 6,005-run Botspot dataset, bot.spot earned an owned-domain citation 56.1% of the time when named in an answer. That 56.1% mention-to-citation conversion comes from publishing canonical entity facts, valid Article and FAQPage JSON-LD, and server-rendered AI Appendices on owned pages. As our comparison of AEO vs GEO vs SEO outlines, earning ChatGPT citations starts with structured owned documentation, whereas winning Perplexity citations requires pairing owned pages with G2, Capterra, and independent editorial reviews.
What is the summary playbook for ranking in ChatGPT and Perplexity?
The summary playbook for ranking in ChatGPT and Perplexity is a repeatable loop connecting on-page passage structure to multi-run citation measurement. Because query fan-out and token sampling prevent any static position-one ranking, B2B marketing teams improve citation odds by placing verifiable, self-contained chunks in the first 30% of key pages and tracking statistical visibility across clean daily runs.
- Front-load direct answers and an AI Appendix. Place an explicit definition and 3 to 6 factual takeaways in the top 20% to 30% of every commercial and guide page, either in your CMS templates or across WordPress, Drupal, HubSpot, and Webflow beta using bot.optimizer.
- Format H2s as buyer questions with 75-word entity-echoing answers. Write headings as literal prompts, repeat the heading subject in the opening words, cap paragraphs at 120 words, and eliminate pronoun-led section openers.
- Anchor claims with named entities and HTML tables. Target 15% to 20% proper-noun density by naming publishers, dates, sample sizes, and competitors, and render feature or study comparisons in semantic
<table>tags. - Build third-party corroboration alongside owned pages. Pair crawlable product pages and `llms.txt` discovery with accurate profiles on G2, Capterra, and industry roundups so Perplexity and ChatGPT verify your category claims.
- Measure visibility rate, citation rate, and share of voice across runs. Track buyer prompt sets daily in bot.radar across ChatGPT, Perplexity, Gemini, Claude, Copilot, Google AI Overviews, and Google AI Mode to spot citation gaps and prioritize content refreshes.