Four AIs, Four Rulebooks: What Each System Thinks Makes an Expert Credible

Marketers talk about 'AI search' as if it were one thing. It is not. OpenAI publishes its rulebook openly. Anthropic publishes something closer to a moral charter. Google publishes a marketing policy, while its real rulebook lives in a 170-page guidelines document used by human raters. Perplexity publishes nothing and lets the ranker answer the question on every query. These differences are not academic. They decide whether your LinkedIn essay gets cited, whether your YouTube transcript appears in an AI Overview, or whether your contrarian take lands in a Claude answer or gets smoothed into generic consensus. A well-cited expert in 2026 is not someone who found the one trick. It is someone who understood that there are four different tricks, each required by a different system. The evidence of divergence is concrete. A Yext analysis of 6.8 million citations across 1.6 million AI responses in October 2025 found Gemini drew 52.15% of citations from brand-owned websites, while ChatGPT drew 48.73% from third-party directories such as Yelp and TripAdvisor. The same brand, the same query, two completely different citation sources. This is not random noise. It reflects two systems running on two different ideas of what credibility means.

OpenAI's Model Spec is the most explicit public document of its kind. It is organized as a chain of command: Platform, Developer, User, Guideline. The document includes a dedicated epistemic section titled 'Seek the truth together,' and the sub-rules read like a philosophy syllabus. For factual questions, the Spec is direct: 'For factual questions (e.g., Is the Earth flat?), the assistant should focus on evidence-based information from reliable sources, emphasizing positions with the strongest scientific support.' For ethical or political questions, the model should 'generally present relevant context, including laws, social norms, and varying cultural perspectives, without taking a stance.' Three things are worth noting about this architecture. First, OpenAI treats truth-seeking as a procedure. Follow the steps, weigh the sources, defer where scientific support is strongest. Second, there is no domain whitelist in the Spec. No 'prefer Wikipedia,' no formal hierarchy of sources by type. Third, the document changes. The April 2025 version and the December 2025 version already differ in measurable ways. This is a living document, not a charter carved in stone. For experts, the practical implication is that alignment with scientific or institutional consensus is rewarded. Contrarian positions are not penalized outright, but they do not get structural encouragement either. Clear sourcing and documented uncertainty are the two strongest signals you can send to an OpenAI-powered system.

Where OpenAI codifies a method, Anthropic codifies a temperament. Claude's Constitution, published January 2026, contains the most aggressive epistemic language of any major AI specification. Four direct quotes matter more than any paraphrase. 'Claude tries to have calibrated uncertainty in claims based on evidence and sound reasoning, even if this is in tension with the positions of official scientific or government bodies.' 'Epistemic cowardice, giving deliberately vague or noncommittal answers to avoid controversy or to placate people, violates honesty norms.' 'Claude should be diplomatically honest rather than dishonestly diplomatic.' 'Sometimes being honest requires courage. Claude should share its genuine assessments of hard moral dilemmas, disagree with experts when it has good reason to, point out things people might not want to hear.' That fourth line is structurally significant. Anthropic is not merely permitting disagreement with official consensus. It is framing the refusal to disagree, when reasoning supports disagreement, as a failure of honesty. This is a different architecture from OpenAI, which nudges toward 'strongest scientific support,' and from Google, whose AI Overviews ranker cannot meaningfully dissent because its authority signals point toward institutional bodies. The implication for experts is asymmetric. A contrarian position backed by clear reasoning and cited evidence has a higher chance of being surfaced by Claude than the same position would have on ChatGPT or Gemini. Short, sourced, well-argued dissent gets structural encouragement on Claude that it does not get anywhere else.

Google publishes a Generative AI Prohibited Use Policy and a Gemini API Usage Policy. Both read like terms-of-service documents. They forbid 'misleading claims of expertise or capability in sensitive areas, for example in health, finance, government services, or the law.' They say nothing about how to weigh sources. Google's operational rulebook is the Search Quality Rater Guidelines, updated 23 January 2025. This document governs both classical Search and AI Overviews. The 2025 update sharpened E-E-A-T in three concrete ways: Experience became a first-class signal alongside formal credentials; the Your Money Your Life scope expanded to cover elections, institutions, and public trust; and AI-generated content was not auto-penalized but must demonstrate editorial review, avoid invented references, and provide concrete informational value. The citation data shows what this architecture produces in practice. A Surfer SEO study of 46 million AI Overview citations in August 2025 found government sources receive an 11.75x citation multiplier. Health answers are dominated by NIH, Healthline, and Mayo Clinic. YouTube appears in 23% of AI Overviews cross-industry and up to 93% in gaming verticals. The YouTube number is strategically important. OpenAI and Anthropic do not retrieve YouTube by default. Google AI Overviews does, at a rate that makes video transcripts a first-class content asset for anyone competing in verticals where YouTube already dominates.

Perplexity publishes no model spec. Its trust model is operational: a six-stage retrieval-augmented pipeline in which sources are admitted to a curated index, reranked across three tiers, and attached to answers as structural citations rather than retrofitted as footnotes. Perplexity does not declare what makes an expert credible. It lets its ranker answer that question on each query. Third-party technical analyses suggest the ranker weights freshness, structural quality, topical authority, engagement signals, and vertical-specific directories heavily. The vertical routing is particularly pronounced. B2B queries lean on LinkedIn and G2. Healthcare queries lean on Zocdoc. Travel queries lean on TripAdvisor. Reddit use is heavy across many verticals but not universal. A Peec AI study of 30 million sources in March 2026, covering ChatGPT, Gemini, Google AI Mode, Perplexity, and AI Overviews, confirmed the divergence. Reddit is the single most-cited domain across AI search overall. ChatGPT preferred Wikipedia, Reddit, and editorial outlets such as Forbes. Google AI Mode leaned toward Facebook and Yelp for local queries. Perplexity emphasized Reddit, LinkedIn, and G2 for business queries. For experts in B2B categories, Perplexity's routing logic is unusually exploitable. A complete LinkedIn profile with written posts, combined with a presence in vertical directories relevant to your category, addresses the two strongest ranker signals simultaneously. No other system weights this combination as heavily.

Five specific divergences emerge from the citation data. Brand-owned content: Gemini weights it heavily (52.15% of citations). ChatGPT barely registers it, preferring third-party sources. User-generated content: ChatGPT, Perplexity, and Google AI Overviews lean on Reddit as a primary source. Gemini does not. Directory type: ChatGPT favors horizontal directories such as Yelp and TripAdvisor. Perplexity favors vertical directories such as Zocdoc, G2, and LinkedIn. Gemini favors neither consistently. Expert dissent: Anthropic actively rewards calibrated disagreement with expert consensus. OpenAI is neutral-to-positive on it. Google structurally suppresses it because its authority signals point toward institutional bodies, not individual voices. Video transcripts: Google AI Overviews surfaces YouTube content between 23 and 93 percent of the time depending on vertical. OpenAI and Anthropic do not retrieve YouTube by default. The meta-picture is clear. Each system is optimizing a different layer. OpenAI is optimizing procedural truth-seeking. Anthropic is optimizing epistemic character. Google is optimizing a classical search ranker that has been re-dressed. Perplexity is optimizing a curated retrieval stack. These are not four implementations of one idea. They are four different ideas about what it means for an AI to know something.

The trick of writing one 'AI-optimized' piece and being cited everywhere does not work. The systems are too different. But four moves address all four architectures in parallel. First, publish in multiple archetypes. A long-form article on an authoritative-looking owned domain satisfies Gemini's preference for brand-owned content. A Reddit discussion with real engagement satisfies ChatGPT, AI Overviews, and Perplexity. A LinkedIn post combined with a complete profile satisfies Perplexity's B2B routing. A YouTube video with a dense description satisfies AI Overviews in most verticals. A Wikipedia entry that references your work is the bonus layer that both OpenAI and Google weight, without either system being able to say they are weighting you directly. Second, make credentials visible. The 2025 E-E-A-T update specifically requires thorough author bios, verified credentials, and in Your Money Your Life categories, expert-reviewed content. Most experts bury their credentials three pages deep or inside a generic 'About' section. Put them above the fold on every page. Third, write with a three-tier reranker in mind. Perplexity's pipeline in particular rewards structural quality: semantic H2 and H3 density, schema markup, explicit publication dates, anchored citations, and topical consistency across related posts. The same structural signals improve selection rates in Google AI Overviews. These are not legacy SEO techniques. They are readability signals that a machine can parse without needing to evaluate your meaning. Fourth, give Claude something to defend. Because Anthropic's Constitution explicitly licenses Claude to disagree with official consensus when reasoning supports it, a well-argued contrarian take is more citable on Claude than a vanilla consensus summary. McKinsey publishes its own contrarian takes in McKinsey Quarterly for exactly this reason. Harvard Business Review built an entire citation economy on defensible heterodox arguments. The same logic applies to individual experts: state your non-consensus position clearly, source it, argue it, and let Claude's architecture do the rest. The meta-move under all four is treating AI visibility as a distribution problem rather than a content problem. One idea, four archetypes, four different systems reading it. The expert who does this does less total writing and gets cited more consistently.