How ChatGPT and Perplexity Choose Which Pages to Cite

You run a Shopify store. Someone types "best merino base layer for cold-weather running" into ChatGPT, and the answer names three brands with links. Yours is not one of them, and you have no idea why. The ranking playbook you know (backlinks, domain authority, page one on Google) does not explain it, because the engines that produce these answers do not rank pages the way Google's ten blue links do. They rewrite the question, fetch candidates from an index you may not be in, cut those pages into passages, and cite whatever the model quoted. This guide walks through that pipeline step by step, with the numbers from the primary studies (the Princeton GEO paper, Ahrefs, Semrush, Seer Interactive, Peec AI) and the official docs from OpenAI, Perplexity, and Google. By the end you will know which parts of the machine you can influence and which parts to stop worrying about.

What happens between a prompt and a cited answer?

Every AI answer engine runs a four-step retrieval-augmented pipeline: rewrite the prompt into search queries, retrieve candidate pages from an index, extract the passages that answer each query, and generate a response that cites the pages whose passages it used.

The engines describe this themselves. OpenAI's help center says ChatGPT search "typically rewrites your query into one or more targeted queries" that it sends to search providers, then may send "additional, more specific queries" after reading the first results, according to the Searching the web with ChatGPT article. Google calls its version query fan-out: "a set of concurrent, related queries generated by the model to request more information and fetch additional relevant search results," per the Google Search Central optimization guide. Google's Robby Stein put it more plainly in a July 2025 interview covered by Search Engine Journal: ask for things to do in Nashville with a group and the system "may think of a bunch of questions like great restaurants, great bars, things to do if you have kids, and it'll start Googling basically."

Two consequences follow from the shape of this pipeline. First, your page competes on sub-queries you never see, so a page that answers one narrow question well can be cited for a broad prompt. Second, the model only cites what it read, and it only reads passages. The rest of this guide is about winning steps 2 and 3.

Which search index does each engine pull from?

ChatGPT searches through third-party providers with Bing as the documented partner, Perplexity searches its own index built by PerplexityBot, and Google AI Overviews and AI Mode draw from Google's core index.

The distinction matters because you cannot be cited from an index you are not in. OpenAI's crawler documentation lists OAI-SearchBot as the bot "used to surface websites in search results in ChatGPT's search features" and adds that "sites that are opted out of OAI-SearchBot will not be shown in ChatGPT search answers." That crawler supplements, rather than replaces, the partner index: Seer Interactive's February 2025 analysis of more than 500 citations across roughly 100 queries found that 87% of SearchGPT citations matched Bing's top organic results, while only 56% matched Google, with a median Google rank of 17.

Perplexity's crawler documentation describes PerplexityBot as "designed to surface and link websites in search results on Perplexity," respecting robots.txt and not used for model training. A second agent, Perplexity-User, fetches pages on demand when a user asks and "generally ignores robots.txt rules" because the visit is user-initiated. The Search Control Guide for Perplexity's API shows the retrieval layer directly: up to 20 results per query (default 10), each with a title, URL, snippet, publication date, and last-updated timestamp.

Google is the simplest case. Per Google Search Central, a page qualifies for AI Overviews and AI Mode if it is "indexed and eligible to be shown in Google Search with a snippet." Standard Googlebot does the crawling. Google-Extended is a separate control for AI training and does not affect AI Overviews.

Citation counts come from Qwairy's analysis of 118,000 responses between January and March 2026, as reported by Whitehat SEO. The YouTube figure is from Ahrefs' March 2026 study, covered below.

Do these engines rank pages or extract passages?

They retrieve pages but select passages, so the unit of competition is the paragraph, not the URL.

Classic search scores a document and returns a link. A generative engine goes one step further: after retrieval it pulls the chunks of text most relevant to each sub-query, feeds those chunks to the model, and the model writes an answer conditioned on them. Perplexity's API makes the extraction step visible with a search_context_size parameter that ranges from "brief passages most relevant to your query" at low to "detailed, query-relevant content" at high, along with a max_tokens_per_page budget. When budgets are tight, only the most relevant passage from each page survives.

The practical effect is that a 3,000-word guide with one crisp paragraph on "merino vs synthetic base layer warmth per gram" can beat a 400-word product page that mentions warmth in passing. Semrush's January 2026 study, which compared 304,805 URLs cited by ChatGPT, Google AI Mode, and Perplexity against 921,614 URLs ranking on Google, found the biggest gaps were in text qualities that make passages extractable: clarity and summarization were 32.83% more common on cited pages, E-E-A-T signals 30.64%, Q&A format 25.45%, and section structure 22.91%. Promotional tone ran the other way at 26.19% less common.

This is also why on-page structure carries more weight in AI search than in classic SEO. The sibling guide on how to structure a blog post so AI search cites it covers the exact heading and paragraph patterns; the short version is that every H2 should read as a question and every first sentence beneath it should read as the answer.

What did the Princeton GEO paper actually measure?

The GEO paper (Aggarwal et al., arXiv 2311.09735, published at KDD 2024) found that adding citations, quotations, and statistics to a page raised its visibility in generative engine answers by 30 to 40%, while keyword stuffing produced little to no improvement.

The authors built GEO-bench, a benchmark of 10,000 queries drawn from multiple sources including Perplexity's Discover feed, then applied nine content edits to source pages and measured how much of each page a generative engine used in its answer. Two metrics were tracked: Position-Adjusted Word Count (how many words from the source appeared, weighted by position in the answer) and Subjective Impression (a seven-part rating of relevance, influence, uniqueness, and click likelihood). The abstract on arXiv reports that "GEO can boost visibility by up to 40% in generative engine responses."

The per-method results, from the paper's Table 1 and Section 4:

Two findings matter most for a small brand. First, the paper reports that "lower-ranked websites, which typically struggle for visibility, benefit significantly more from GEO." In the multi-site experiment (Table 2), Cite Sources produced a 115.1% visibility increase for the source ranked fifth in the SERP, while the top-ranked source's visibility fell by 30.3%. Quotation Addition and Statistics Addition showed the same pattern at 99.7% and 97.9% for rank five. Second, tactics that work in classic SEO transferred poorly: the authors write that keyword stuffing "offers little to no improvement on generative engine's responses."

How much does freshness matter?

Freshness is a retrieval signal before it is a citation signal: the index decides which recent pages become candidates, and a page that never becomes a candidate cannot be cited regardless of quality.

The engines expose this. Perplexity's Search API returns a date and last_updated field for every result and offers a search_recency_filter, which means recency is a first-class retrieval input, not an afterthought. OpenAI's help article on ChatGPT search shows a query rewrite that appends the year ("CCR8 immunotherapy drug development 2025"), so the fan-out step itself often asks for fresh material.

Third-party measurement points the same way. Gander's Q1 2026 analysis of 194,077 cited URLs describes a "1-year half-life" in which content loses roughly 40 to 60% of its AI visibility per year, with two-year-old content at about a third of peak, and notes that "the retrieval layer is far more recency-biased than the citation layer." Google's own guidance does not set a freshness rule, but Ahrefs attributes the January 2026 drop in top-10 overlap partly to Gemini 3 expanding queries more aggressively, which pulls in newer pages from sub-query SERPs.

For a Shopify blog the implication is operational, not philosophical. Visible dates, a real dateModified in your Article schema, and a refresh cadence for your top-cited posts are what keep a page in the candidate pool. Citevolt's Growth tier ($497/mo) includes two content refreshes per month for exactly this reason: a post that was cited in March and untouched since is a post that is decaying.

Why do Reddit and YouTube show up so often?

Reddit and YouTube dominate AI citations because their pages match the shape of fan-out sub-queries: a question in the title, first-hand answers beneath it, constant freshness, and near-total index coverage.

The numbers are consistent across studies. Peec AI's March 2026 analysis of 30 million cited sources, reported by Search Engine Land, ranked Reddit as the most-cited domain across ChatGPT, Google AI Mode, Gemini, Perplexity, and AI Overviews, with YouTube, LinkedIn, Wikipedia, and Forbes completing the top five. Profound's tracking of 680 million citations between August 2024 and June 2025 shows how the mix differs by engine: Wikipedia led ChatGPT at 7.8% of citations, Reddit led Perplexity at 6.6%, and Reddit led AI Overviews at 2.2% with YouTube at 1.9%. Ahrefs found that among AI Overview citations that do not rank in the top 100, YouTube alone accounts for 18.2%.

Notice how small the leading percentages are on a per-citation basis. Even the top domain rarely holds more than a few percent of all citations on a given engine. The long tail is enormous, which is the opening for a niche store. A Reddit thread wins because someone with the exact problem asked the exact question and three people who own the product answered. Your blog can do the same thing with more authority, more numbers, and a product to link to, as long as it is written in that shape. The guide on how to get your Shopify store cited by ChatGPT goes through the keyword-to-post workflow.

How does each engine show its citations?

ChatGPT shows inline link chips plus a sources panel, Perplexity numbers every citation inline and lists the sources above the answer, and Google AI Overviews attach link cards to the passage they support.

The display format changes what a citation is worth. In ChatGPT, the model generates prose first and attaches links to the sentences it grounded in retrieved text; sources not directly used may not appear at all. OpenAI's help center notes that "search results and citations can be incomplete, outdated, or incorrect," which is a candid description of a model choosing what to show. With roughly eight citations per answer, each slot is scarce.

Perplexity is built around verifiability, and it shows. Every sentence carries a bracketed number, the source list sits at the top of the answer, and the API returns the citation list as structured data alongside the text. Because Perplexity averages close to 22 citations per response in the Qwairy dataset, a page has more chances to appear but each appearance is diluted. Position in the numbered list matters more than presence.

Google AI Overviews render links as cards beside or beneath the summary, and Google states its models "identify more supporting web pages, allowing us to display a wider and more diverse set of helpful links" than a classic result page. Because the links map to fan-out sub-queries, a page can be cited for a sub-topic it ranks for even if it does not rank for the head term. The March 2026 Ahrefs study of 863,000 keywords quantified this shift: 37.9% of AI Overview citations came from the top 10 organic results, 31.2% from positions 11 to 100, and 31.0% from beyond the top 100, down from about 76% top-10 overlap in July 2025.

What is the difference between ChatGPT search and Perplexity?

ChatGPT search is a conversational model that searches when it decides it needs to and cites sparingly, while Perplexity is a search engine that always retrieves first and cites densely, with Pro Search adding multi-step retrieval on top.

The retrieval trigger is the first difference. OpenAI's help center says ChatGPT "may choose to search the web based on what you ask," so for many product questions the model answers from parametric memory with no citations at all. Perplexity retrieves on every query. That is why ChatGPT citation rates for any given brand are low and volatile while Perplexity rates are higher and steadier.

The second difference is depth. Perplexity's default answer runs one search pass. Pro Search performs multi-step reasoning: it reads the question, formulates follow-up searches, reads more sources, and may ask a clarifying question before answering. The same split exists in Perplexity's API. The Sonar model page lists sonar as the "lightweight, cost-effective search model with grounding," sonar-pro as the "advanced search offering with grounding, supporting complex queries and follow-ups," and sonar-deep-research for "exhaustive searches" that produce full reports. Note that the Sonar chat completions endpoint is being folded into an Agent API, with support running until September 27, 2026, so tooling that tracks Perplexity citations will need to migrate.

The third difference is which index the fan-out hits. ChatGPT's rewritten queries go to Bing-backed providers, so Bing rankings predict ChatGPT citations. Perplexity's go to its own index with recency baked in. A page that ranks on Google but is thin on Bing will show up in AI Overviews and Perplexity long before it shows up in ChatGPT. Citevolt tracks all three engines weekly for each client because a brand's citation profile is almost never uniform across them.

What changes from week to week?

Citation sources shift far faster than organic rankings: an engine can drop a domain by 50 percentage points in a month when its retrieval or reranking layer is retuned.

Semrush documented the clearest example. Tracking 230,000 prompts weekly across ChatGPT, Google AI Mode, and Perplexity from July to October 2025, it found that ChatGPT's citation of Reddit collapsed from about 60% of responses in early August to about 10% by mid-September 2025, and Wikipedia fell from roughly 55% to under 20% in the same window. Perplexity and Google AI Mode did not move. Semrush's Sergei Rogulin attributed it to "an attempt to avoid over-citing on certain websites," which is a reranking policy change, not a content change.

Google moved the same way with a model swap. Ahrefs' July 2025 sample showed about 76% of AI Overview citations coming from the top 10; its March 2026 sample showed 37.9%, and the authors point to the Gemini 3 rollout in January 2026 expanding fan-out more aggressively. Three other things change on a shorter clock: which sub-queries the fan-out generates for a prompt, which fresh pages entered the index this week, and the model's own answer wording, which decides which passages get quoted.

What does not change week to week is the shape of a citable passage. The GEO paper's edits worked in 2023 on a Bing Chat clone, and Semrush's 2026 correlations on live engines point at the same qualities. Optimize for the stable layer.

What can a Shopify brand actually control?

A Shopify brand controls five inputs: crawler access, index coverage, passage quality, freshness, and entity consistency. It does not control which index an engine uses, how the fan-out is generated, or the model's reranking, so effort spent there is wasted.

1. Open the doors in robots.txt. Allow OAI-SearchBot, PerplexityBot, and Googlebot. Shopify's default robots.txt allows them; a custom robots.txt.liquid or a bot-blocking app can silently close them. OpenAI is explicit that opted-out sites are not shown in ChatGPT search answers.
2. Get indexed on Bing, not only Google. Verify the store in Bing Webmaster Tools, submit the sitemap, and check that blog URLs are indexed. Seer's 87% Bing match rate means this is the single largest lever for ChatGPT specifically.
3. Write posts as question-shaped passages. One question per H2, the answer in the first sentence, then a number, a named source, or a quotation. This is the GEO paper's Cite Sources, Statistics Addition, and Quotation Addition applied to a blog post.
4. Put the entity on the page. Brand name, product names, and the same facts (weight, material, price, sizing) stated consistently across product pages, blog, and About page. Add Article and Product schema. Google says schema is not required for AI features; it still helps the engine tie the passage to your brand.
5. Refresh on a schedule. Update the two or three posts that already earn citations every 60 to 90 days with a new number, a new date, and a visible updatedAt. Freshness keeps a page in the candidate pool.
6. Measure per engine. Run the same 10 to 20 buyer questions through ChatGPT, Perplexity, and Google AI Overviews weekly and log which URLs get cited. The free Citevolt audit does this for six queries as a starting point.

Citevolt's own baseline run on 2026-09-14 across six buyer queries found the same thing the studies do: the cited pages were long explainer and comparison articles, not product pages. That is the format Citevolt writes and publishes weekly for Shopify stores, four to twelve posts a month depending on tier, with per-engine citation tracking in the Growth and Scale plans. If you would rather build this in-house, the sibling guide on what AI SEO agencies charge lays out the cost of doing it either way.

What to do next

Start with the audit, then fix access, then fix structure.

1. Run your store through the free audit at audit.citevolt.com to see which engines already cite you and which competitors take the slots.
2. Open Bing Webmaster Tools and confirm your blog URLs are indexed. If they are not, ChatGPT cannot cite them.
3. Check yourstore.com/robots.txt for OAI-SearchBot and PerplexityBot. Remove any disallow rules a bot-blocker app added.
4. Pick your five best-selling product categories and write one question-shaped explainer per category, 1,500 words or more, with at least three sourced numbers and one quotation each. Use the structure guide as the template.
5. Re-run the same buyer questions in four weeks. Perplexity will usually move first; ChatGPT will move last. If you want this done for you at a fixed price, the Citevolt plans start at $297/mo for four posts.