How AI Search Engines Choose What to Cite

Every major AI assistant now sits on top of two very different information sources. One is the model's training data — a frozen snapshot of text the model learned from, with a cutoff date, no awareness of anything published after that date, and no way to verify whether a specific fact is still true. The other is live retrieval: a real-time web search or index lookup the system runs at the moment you ask a question, feeding fresh results back into the model before it writes an answer. This second mechanism is what the industry calls retrieval-augmented generation, or RAG, and the general term for citing retrieved sources back to the user is grounding.

The decision to search rather than answer from memory is not made by the language model reasoning in isolation — it's typically handled by a routing layer, a classifier, or an explicit tool-use decision baked into the product. Signals that commonly trigger a search include time-sensitive language ('today,' 'latest,' '2026'), named entities that change over time (companies, prices, people, products), factual questions the system judges likely to be outdated or unreliable from memory alone, and explicit user requests like 'look this up.' None of the six engines covered below have published a complete, authoritative list of these triggers, and the exact thresholds are proprietary. What is public and consistent across all of them is the shape of the pipeline: retrieve candidate documents, rank and filter them, pass a subset into the model's context window, then generate an answer with citations pointing back to whichever sources actually influenced the output. Getting cited means surviving every stage of that pipeline, not just ranking well in a traditional search index.

ChatGPT's live-search capability is powered by a search index (OpenAI has partnered with Bing and built its own crawling and indexing infrastructure) and is invoked either automatically, when the system judges a query needs current information, or explicitly, when a user selects a search mode. This is distinct from the model's underlying training data — when browsing is active, ChatGPT issues one or more queries, retrieves a set of web pages, and constructs its answer from snippets and page content pulled from that retrieval step, with inline citations linking to the specific pages used.

OpenAI has not published the exact ranking formula that decides which retrieved pages make it into a given answer, so any claim of a precise algorithm should be treated as speculation. What is observable from how the product behaves is that ChatGPT tends to favor pages that answer the query directly and concisely near the top of the content, pages from sites with an established topical track record, and pages that are recent when the query is time-sensitive. It also appears to cite multiple sources per answer rather than relying on a single page, and it will pull from forums, review sites, and comparison pages alongside traditional publishers when those sources contain the specific, extractable answer to the question asked.

Claude's web search is implemented as a tool the model can invoke mid-conversation, rather than a separate mode the user has to toggle. When a query would benefit from current information, or when Claude judges its training data is insufficient to answer confidently, it can issue a search, review the results, and optionally issue follow-up searches before composing a final answer — this multi-step search behavior is one of the more distinctive aspects of how Claude grounds its responses, since it can chain several queries together rather than firing a single search and stopping.

Anthropic has been explicit that Claude aims to cite sources transparently and avoid presenting unverified claims as fact, which shapes citation behavior in a specific way: Claude tends to attribute claims closely to the specific source that supports them, rather than synthesizing an answer and citing sources loosely at the end. For content to be picked up, it generally needs to state facts in a way that's easy to extract and attribute cleanly — a single page making a clear, well-supported claim is easier for Claude to cite confidently than a page that buries the same claim in caveats or scatters it across multiple sections.

Gemini's grounding is tightly coupled to Google Search itself — when grounding is enabled, Gemini issues queries against Google's search index and can return results with direct links, making it functionally different from a model that queries a generic web index built by a separate company. This gives Gemini access to the same ranking signals and index depth that power traditional Google Search results, including Google's understanding of site authority, structured data, and freshness.

In practice, this means content that already performs well in conventional Google Search — pages with strong backlink profiles, clean structured data (schema markup), clear headings, and demonstrated topical authority — has a real chance of being reused as grounding material in Gemini's answers, since the retrieval step is drawing from the same underlying index. Google has been more forthcoming than most competitors about the mechanics of grounding with search as a developer-facing API feature, but the exact weighting of signals inside a given answer still isn't public. Sites that are poorly indexed or blocked from crawling by Googlebot are unlikely to surface in Gemini's grounded answers for the same reason they wouldn't rank well in standard search.

Perplexity is built around citation as the core product experience rather than an add-on feature — every answer is constructed as a synthesis of ranked sources, with numbered citations attached to nearly every sentence. This citation-first architecture means Perplexity's source-ranking step is doing more of the visible work than in chat-first products: it typically retrieves a broad set of candidate pages, scores them for relevance and apparent reliability, and then writes an answer that stitches together the highest-ranked sources rather than generating first and citing second.

Perplexity also offers Focus modes that constrain retrieval to specific domains (academic papers, Reddit, YouTube, or the general web), which changes which sources are even eligible for a given query. Independent observation of Perplexity's output suggests it favors sources with clear factual density — pages that state specific numbers, dates, and named entities rather than vague or promotional language — and it appears to reward source diversity, often pulling from five or more distinct domains in a single answer rather than leaning on one or two. As with the other engines, Perplexity has not published its exact ranking weights, so these are observed patterns rather than confirmed mechanics.

AI Overviews are generated directly inside Google Search results, drawing on Google's web index plus the Knowledge Graph, and are shown to a subset of queries where Google judges a synthesized, multi-source summary is more useful than a traditional list of blue links. AI Mode extends this into a more conversational, multi-turn search experience, letting users ask follow-up questions within the same session while still grounding answers in retrieved web content. Both surfaces sit on top of the same underlying Google index that traditional search rankings use, which is a meaningful architectural difference from most competitors.

Because AI Overviews and AI Mode are so closely tied to Google's core ranking systems, the same fundamentals that drive organic search visibility — crawlability, page experience, topical authority, structured data, and content that directly answers a query — carry over as prerequisites for being cited. Google has stated publicly that AI Overviews draw from pages that already rank well organically, though it has not disclosed a separate, distinct ranking algorithm specific to the AI Overview synthesis step itself. A page that doesn't rank on page one for a related query is unlikely to be cited in an AI Overview for that query, which makes traditional SEO fundamentals a floor requirement rather than a separate discipline here.

Microsoft Copilot's grounding is built on the Bing search index, giving it a retrieval foundation distinct from Google's and from OpenAI's own crawling infrastructure, even though Copilot uses OpenAI models for generation. When Copilot determines a query needs current information, it queries Bing, retrieves and ranks candidate pages, and composes an answer with citations linking back to the specific pages used — a pipeline structurally similar to ChatGPT's browsing mode, but pulling from a different underlying index with its own crawl coverage and ranking signals.

Because Bing has historically had a smaller crawl footprint than Google in some categories, and different ranking emphases in others (Bing has, for instance, been noted as placing relatively more weight on structured, well-organized content and clear on-page signals), a page's visibility in Copilot doesn't always mirror its visibility in Google-backed engines. Sites that are indexed and performing well in Bing but not prioritized in Google's index — or vice versa — can see meaningfully different citation rates in Copilot versus Gemini or AI Overviews, which is one of the clearest illustrations of why engine-specific tracking matters rather than assuming Google visibility is a proxy for all AI search visibility.

Across all six engines, four factors show up repeatedly as prerequisites for citation, even though each engine weighs and implements them differently: authority (does the source have an established track record on the topic, whether measured through backlinks, domain reputation, or repeated mention across the web), freshness (is the content current enough to be trustworthy for a time-sensitive query), structural clarity (can the retrieval and generation pipeline easily extract a clean, specific answer from the page, aided by clear headings, direct statements, and structured data), and factual specificity (does the page state concrete facts — numbers, dates, named entities — rather than vague or purely promotional language). These four factors are the closest thing to universal rules in a landscape where none of the underlying ranking algorithms are fully public.

But 'closest to universal' still isn't identical. ChatGPT and Copilot draw from different underlying indexes (OpenAI's own infrastructure plus Bing partnerships, versus Bing directly). Gemini and AI Overviews draw from Google's index and inherit its organic ranking signals almost directly. Claude and Perplexity build their own retrieval and ranking layers on top of general web search, with Perplexity in particular optimizing hard for citation density and source diversity. That means a brand can be cited constantly in Perplexity answers while being nearly invisible in Copilot, or dominate Google AI Overviews while barely appearing in Claude's search results — not because of an error, but because each engine is running a genuinely different retrieval and ranking pipeline on a genuinely different index. Treating visibility in one engine as a stand-in for visibility across all of them is the single most common mistake in early AI-visibility strategy, and it's the reason cross-engine tracking exists as its own discipline. MentioningYou monitors citation and recommendation behavior across ChatGPT, Claude, Gemini, Perplexity, Google AI, and Copilot individually, precisely because engine parity can't be assumed — it has to be measured, engine by engine, on an ongoing basis.