How ChatGPT, Perplexity, Gemini, and Google AI Overviews Choose Sources

The source list is not a leaderboard

People ask which pages ChatGPT, Perplexity, Gemini, or Google AI Overviews “rank.” That wording is tidy, but it hides the messy part. An answer is usually assembled through several decisions: understand the request, decide whether current information is needed, search or retrieve candidates, rank and filter them, read selected passages, compose an answer, and connect claims to citations. Those decisions are related to search ranking, but they are not the same as a ten-blue-links result page.

The vendors document pieces of this pipeline, not a shared recipe. OpenAI says ChatGPT Search may rewrite a question into targeted searches and may use third-party search providers. Google documents AI Overviews as a Search feature with links to supporting web content, while Gemini API documentation describes Google Search grounding as a tool that supplies search results and source metadata. Perplexity’s developer documentation exposes a search system with ranked results, domain filters, multiple queries, and content extraction. Those are useful facts. “This page gets a score of 0.83” is not.

Step one: decide whether to retrieve

A model can answer from its learned parameters, but that is a poor fit for questions about today’s election, this month’s pricing, a new API release, or a local opening hour. Search-enabled products can route those requests to retrieval. The route may be triggered by the words in the question, the conversation context, an explicit request to search, or a system decision that the answer needs current evidence.

This is why the same prompt can behave differently across products. Perplexity is built around search and presents citations as part of its answer experience. ChatGPT can search the web when its system decides that browsing is useful. Gemini can use Google Search grounding in supported experiences and APIs. Google AI Overviews are an answer layer inside Google Search, not a general-purpose chatbot with an identical retrieval stack. Do not turn a product’s visible citation style into a claim about every model behind it.

Step two: turn one question into search work

Natural-language questions are often underspecified. “What changed in the SDK?” contains a product, a time window, a version, and an implied comparison. A retrieval system can issue several narrower searches: the official release notes, the current documentation, and coverage explaining the practical impact. OpenAI explicitly describes query rewriting for ChatGPT Search; Perplexity documents multi-query search in its Search API. For Google’s systems, the public documentation describes grounding and links, but not a complete query-planning trace for every AI Overview.

That distinction matters for publishers. You are not writing only for the exact sentence a customer types. You are writing for the related formulations an engine may generate: a definition query, a comparison query, a recency-qualified query, or a query containing an entity disambiguation term. Clear terminology and explicit relationships make a page eligible for more of those retrieval paths without stuffing synonyms into every paragraph.

Step three: retrieve candidates, then rank them

Retrieval is the broad net. Ranking is where the system decides which candidates deserve attention. Familiar search signals still matter: textual relevance, links and reputation, page quality, language, location, freshness where appropriate, and whether the content can be fetched and understood. A search result can be relevant without being the best passage for an answer, so an answer engine may apply another ranking or selection step after fetching content.

The public evidence supports a careful statement: answer engines use search or retrieval systems and rank candidates, but the vendors do not publish one cross-product formula for source selection. Perplexity’s Search API documentation is unusually concrete about returning ranked results and supporting domain filtering. Google Search Central documents how crawling, indexing, and ranking work at a high level, not how a particular AI Overview chooses every citation. Treat claims about hidden “authority scores,” fixed top-three slots, or guaranteed passage lengths as speculation unless a vendor has documented them.

Relevance is not the same as authority

A page can be authoritative but answer the wrong question. An official product homepage may be the strongest source for what a product is, while its release notes are stronger for what changed this week. A government dataset may beat a blog for a statistic, while a specialist tutorial may better explain implementation details. Selection is query-dependent. “Be authoritative” is directionally right advice; “publish the most relevant evidence for the claim you want associated with your brand” is more operational.

The passage itself also matters. Retrieval systems work with documents, snippets, chunks, and extracted text. A page that buries the answer under a hero slogan, renders core facts only inside an image, or relies on “as described above” creates extra interpretation work. That is an informed inference from how retrieval consumes content, not a published promise that a particular engine rewards a 40-word answer. Still, it is a rational engineering target: make the claim legible before asking a model to summarize it.

Grounding: selected text becomes evidence

Grounding is the bridge between retrieved content and generated prose. The model receives search results, fetched passages, or another tool’s structured output and is instructed to answer using that material. Google’s Gemini documentation describes Search grounding responses with grounding metadata, including queries and supporting web chunks. The metadata is not just decoration: it gives an application a way to inspect what search supplied to the model.

Grounding does not make an answer infallible. A source can be stale, ambiguous, incomplete, or misinterpreted. Several pages can repeat the same unsourced claim, creating the appearance of agreement without independent evidence. A model can also make a synthesis that no single citation states verbatim. Good systems therefore need claim-to-source alignment, not merely a handful of links appended after generation.

• A source supports the claim: the linked passage actually contains the relevant fact.
• The source is fit for the claim: an official specification, filing, dataset, or firsthand report is used where appropriate.
• The date fits the question: current claims use current evidence, while historical claims preserve historical context.
• The wording preserves uncertainty: conflicting or incomplete evidence is not flattened into false certainty.

Citations are an output contract, not a quality badge

Citations make an answer auditable, but their presence alone proves little. A citation can point to a page that supports one clause while appearing beside a paragraph containing four other claims. It can identify a search result rather than the exact sentence used. It can also be absent when a response came from the model’s internal knowledge, when the product did not browse, or when its interface chose not to expose every source.

For teams measuring visibility, count more than “mentioned” versus “not mentioned.” Record the question, the exact claim, the cited URL, the passage that supports it, the date checked, and whether the source is first-party or independent. A brand mention without a supporting link is a different outcome from a precise citation to its documentation. A citation to a directory is different from a citation to the business’s own pricing page.

Freshness without the freshness theater

Freshness is a conditional signal, not a universal victory condition. “Who is the current CEO?” needs recent evidence. “How does TLS work?” benefits more from stable technical references than from a page published yesterday. Search systems can use dates, update signals, and query intent, but an updated timestamp does not make unchanged copy new. A page that quietly changes its date risks confusing both readers and systems.

Maintain dates as data, not decoration. Put publication and meaningful update dates in the page, structured data, and sitemap where applicable. Explain what changed in release notes and changelogs. Keep old URLs when they remain useful, redirect retired versions carefully, and link current documentation from the canonical product page. This creates a source trail an engine can follow instead of a pile of near-duplicate pages competing with one another.

Why source diversity shows up in good answers

A single publisher can be excellent and still be insufficient. A product claim is strongest when official documentation states the behavior, independent technical coverage tests or explains it, and a standard or regulator provides the governing context. Multiple copies of one press release are not diversity; they are duplication. Nor is diversity a license to treat every opinion as equal. The right mix depends on the claim.

Vendors do not publish a universal source-diversity quota for ChatGPT, Perplexity, Gemini, or AI Overviews. The practical inference is narrower: pages that add distinct evidence have a better chance of contributing to a robust answer than pages that merely paraphrase an existing source. Publish original measurements, clear methodology, named authorship, primary documents, and useful counterexamples. Make your page the source that adds something, not the fifth page repeating the first.

The publisher’s implementation checklist

You cannot force an engine to cite you. You can reduce the reasons it would skip you. Start with search fundamentals: allow legitimate crawlers, return a healthy page, use canonical URLs, expose important text in crawlable HTML, and keep navigation understandable. Then improve answerability: put a direct definition near the top, use descriptive headings, name the subject in each important paragraph, and state dates, units, conditions, and exceptions explicitly.

• Map each commercial or reputational claim to a page that can serve as its primary source.
• Write one standalone passage for each question customers actually ask, rather than hiding the answer in a feature grid.
• Use schema.org where it accurately describes visible content; structured data is a consistency layer, not a citation shortcut.
• Add authorship, editorial ownership, references, methodology, and update history where they help a reader judge the evidence.
• Test prompts across engines and save the citations, not just the generated wording.

The uncomfortable conclusion

There is no single AI source algorithm waiting to be reverse-engineered. ChatGPT, Perplexity, Gemini, and Google AI Overviews overlap in their use of retrieval, ranking, grounding, and citations, but they expose different tools and make different product decisions. The honest strategy is not to chase a secret factor. It is to build pages that are discoverable, relevant, current when they need to be, explicit about their evidence, and useful as independent sources.

Then measure the outcome like an engineer. Ask a fixed set of real questions, record whether the engine searched, inspect which passage it cited, compare your source with the sources beside it, and repeat after meaningful changes. If your page is not selected, the fix may be technical, editorial, evidentiary, or simply a better competing source. At least you will be debugging the right system instead of optimizing for a ranking factor nobody documented.

This article separates vendor-documented behavior from informed inference. Product interfaces, retrieval systems, and documentation can change; verify current behavior before treating an observation as a specification.