How ChatGPT and Perplexity decide sources to cite is retrieval plus a short sourced passage. ChatGPT search may query Bing. This URL holds 1.04 percent share of that query in Bing AI Performance (29 Aug 2026).
How ChatGPT and Perplexity decide sources to cite is not a lottery. ChatGPT search rewrites your prompt into targeted queries and may send them to Bing. Perplexity retrieves live pages at query time. In both cases, a page is cited when a short, sourced passage matches a sub-question the engine actually ran.
I am Amelia S. Gagne, CEO of Kief Studio and co-author of The Crossroads of AI Integration. I measure this on my own site. Bing Webmaster AI Performance for ameliasgagne.com (export 29 August 2026) shows this URL already earning citations on the query how ChatGPT Perplexity decide sources to cite, at 1.04 percent share of a very large implied pool. Share that low on your own subject is a warning: the query is huge, and one page cannot own it. This article is the mechanics page. Sibling posts own how to get a small business cited, how assistants recommend a business, and why volume loses.
OpenAI's Help Center is explicit. ChatGPT search "typically rewrites your query into one or more targeted queries" and sends those to search partners. The same article names Bing (and Shopify for some commercial results) as a partner, and says ChatGPT may send follow-up queries after it reads the first batch. Enterprise and Edu workspaces are even plainer: disassociated search queries can be shared with the Bing search engine. If your page is not in Bing's index, ChatGPT search cannot retrieve it in that mode.
That rewrite step is fan-out. One user question becomes several engine questions. Independent measurements in August 2026 (Search Engine Journal; Promptwatch) describe ChatGPT stacking extra fan-out queries, including a sharp rise in site: operators around 8 August. Official pages, documentation, and named domains get asked directly. Forum threads get asked less. If you want this page cited, its title and first paragraph have to match a fan-out query, not a slogan you invented.
Perplexity is a different machine with the same job. It searches the live web on the query in front of it, reads a set of pages, and cites a smaller set. The overlap with Google's top 10 is incomplete. Ahrefs has reported that only a small fraction of URLs cited by assistants rank in Google's top 10 for the original query. Organic rank still matters for Google AI Overviews. It is not a prerequisite for Copilot or ChatGPT search. I have measured that split on client properties: a page can earn hundreds of Copilot citations with almost no Bing organic impressions.
Retrieval is not citation. Engines gather more pages than they name. A page can be fetched and still dropped because the claim the model needs is buried, hedged, or unsourced. Put the answer in the first 40 to 60 words of the article and of each H2. Keep subject, claim, and number in one sentence so a short reading window still contains a complete fact.
Google AI Overviews pull from Google's index. Pages Google already trusts have an advantage. FAQPage schema, a direct answer near the top, and a consistent entity help. This is the one surface where classic SEO still gates the citation. If you only chase Copilot, you can skip ranking. If you want AI Overviews, you still need crawlable, indexed, reputable pages.
Do not treat "rank everywhere, cite everywhere" as one job. Getting a small business cited by ChatGPT is a GEO checklist. This page is the selection logic. Zero-click search is not the enemy when the engine reads you aloud. Measure the channel you care about. Bing AI Performance for Copilot. Manual prompts for ChatGPT, Perplexity, Gemini, and AI Overviews.
Entity recognition is why a named author with a Person schema, a stable sameAs graph, and a publication history beats an anonymous "team" byline on the same words. That is entity SEO: the Knowledge Graph has to know who wrote this. I publish as Amelia S. Gagne, CEO of Kief Studio, with the same name on LinkedIn, this site, and the book page. Inconsistency fragments the entity.
Answer structure is why writing for featured snippets still transfers. A heading that restates a sub-question, followed by a 20 to 40 word self-contained answer, is what both snippets and assistants lift. Creative titles that do not match any query die in fan-out.
Structured data does not magic a citation into existence. It reduces guesswork. Article, Person, Organization, FAQPage, and BreadcrumbList tell a crawler what the page is. JSON-LD structured data is the format worth implementing. Invalid markup is worse than none. Google's Rich Results Test is the check, not a plugin's green badge.
Freshness is real and easy to fake badly. A timestamp bump with no new sentences does not fool a model that is reading the body. Substantive updates, a visible updated date, and a new sourced fact do. Princeton, Georgia Tech, Allen AI, and IIT Delhi's GEO paper (arXiv:2311.09735) found that adding citations, quotations, and statistics can raise visibility in generated answers, with the gain depending on the query. That is still the best public experimental evidence we have. It is not a promise that stuffing numbers works. The numbers have to be true and attached to a named source.
Thin pages that promise a complete guide and deliver a glossary. Sales copy dressed as research. Entities that cannot be verified off-site. Claims that contradict higher-authority sources without evidence. Keyword-only posts with no decision and no number. Those are not "penalized" in a Google-algorithm sense. They are simply not selected when the engine has a cleaner passage from someone it can name.
I study behavioral psychology, and the reader-side shift is simple. People used to scan ten links. Now a model reads for them and returns a short answer with a few names. If you are not in that answer, you were never in the running. That is why how AI assistants decide which businesses to recommend is a different query from this one: recommendation is a naming event. This page is the retrieval event that precedes it.
Keep the slug if the URL already has citations. Change the title and H1 to the query Bing or ChatGPT actually run. Rewrite the first 50 words so they contain the answer and the keyword. Add three outbound links to primary sources (.gov, standards, vendor-help that is the actual documentation). Add FAQ pairs whose first sentences stand alone. Then ping IndexNow so Bing can recrawl. That is the same work we run for clients on LTFI managed operations: structure and freshness as ongoing work, not a launch checkbox.
Do not add more FAQ questions to a page that already lost. On ameliasgagne.com, pages with more numbers and more FAQs have lost to shorter decision pages that matched the grounding query. The control is query match plus outbound authority, not schema volume. Why more content loses in answer engine optimization is the sibling for that point. E-E-A-T is not a checklist is the sibling for the entity work.
ChatGPT search rewrites the prompt into targeted queries and may send them to Bing. Perplexity retrieves live pages at query time. Both cite a short, sourced passage that answers a sub-question, not an entire homepage.
Yes, in documented search modes. OpenAI's Help Center names Bing as a search partner and says ChatGPT may share rewritten queries with Bing. Pages missing from Bing cannot be retrieved that way.
No. Google AI Overviews lean on Google's index and trust. ChatGPT search and Copilot can cite a page that never ranks for humans. Treat AI Overviews and ChatGPT as separate channels.
Schema and answer-first rewrites can be recrawled in days if you ping IndexNow and stay in Bing. Entity recognition (consistent name, credentials, third-party mentions) still takes months to accumulate.
No. Write the answer a practitioner would actually use, in the first sentences, with named sources. Architecture changes. Voice does not. Content written to trick a model reads thin to humans and gets dropped.
An AEO audit in a week: ten real questions, who gets cited, one URL per query, extractable first sentences, schema, IndexNow. This domain: 13 cited URLs of 147 posts in Bing AI Performance (29 Aug 2026).
AEO for Massachusetts local service businesses: identical NAP, town-true pages, 20 to 40 word FAQs. SparkToro 2024: 58.5 percent of US Google searches ended without a click. Fake city mills are a trust leak.
A Bing index still matters for Copilot citations. OpenAI: ChatGPT search rewrites queries and may cite partners. This domain: 13 cited URLs of 147 posts (Bing AI, 29 Aug 2026). IndexNow with a real key file. One query, one URL.
Work With Us
Kief Studio builds, protects, automates, and supports full-stack systems for businesses up to $50M ARR.
Newsletter
Strategy, psychology, AI adoption, and the patterns that actually compound. No spam, easy to leave.
Subscribe
