Inside the Engines: Four Architectures, Four Levers

The shared spine

Across Google AI Mode, ChatGPT search, Perplexity, and Claude, the pipeline looks like this. The user’s typed query gets rewritten by an upstream LLM into one or more retrieval queries that better match the underlying retrieval index. Each rewritten query is dispatched against one or more retrieval back-ends: a lexical index (BM25 or equivalent), a dense vector index, sometimes an entity-graph traversal, sometimes a domain-restricted federated search. The candidate documents come back, get reranked by a cross-encoder model that scores each candidate against the rewritten query in context, and the top-K survivors are synthesized by the generative model into the answer the user sees. Citations are inserted at synthesis time, frequently with a separate citation-selection model deciding which sources to attribute.

The spine is consistent; the implementations diverge sharply.

Google AI Mode — the index you cannot beat

Google’s AI Mode is the only engine in this list that retrieves against a comprehensive, continuously updated index of the open web. The retrieval side is Google’s existing crawl + index pipeline, with AI Mode adding query-fan-out, passage extraction, and synthesis on top.

The optimization levers, in rough order of impact:

1. Be cleanly indexable by Googlebot. If classic Google does not have you, AI Mode does not have you. SEO hygiene is the entry condition.
2. Structured data and entity alignment. AI Mode appears materially more sensitive to schema.org type coverage, sameAs identifiers to canonical sources (Wikidata, Wikipedia, LinkedIn for people and organizations), and Knowledge Graph entries than classic Search.
3. Passage extractability near the top of body content. Passages that answer common questions in the first 200–300 words of the body are disproportionately surfaced.
4. Freshness for time-sensitive queries. AI Mode incorporates last-modified signals more aggressively than classic Search. Stale content that ranked well on authority alone is increasingly being passed over.
5. Authority signals. Domain authority, internal-link topical clustering, and link-graph centrality continue to matter, though the weights appear lower than in classic Search.

ChatGPT search — the just-in-time fetcher

ChatGPT does not maintain a persistent crawl of the open web. When the user’s query triggers search (which is most queries in GPT-5.x with the search tool enabled), the system runs a federated query against Bing’s index, plus its own internal scoring, identifies candidate URLs, and fetches them on demand. The retrieved page contents are inserted into the model’s context window for synthesis.

The optimization levers are different in character:

1. Allow OAI-SearchBot, GPTBot, and ChatGPT-User in robots.txt. If the fetcher cannot retrieve your URL, you are not cited. We see this routinely on enterprise sites whose CDN defaults block OpenAI’s crawlers.
2. Sub-second TTFB. The fetcher has a budget. Pages that take three seconds to first byte are often abandoned mid-fetch and never enter the synthesis context.
3. Render-free HTML. ChatGPT’s fetcher does not execute JavaScript in any reliable way. If your content appears only after client-side rendering, it is invisible.
4. Clean passage structure near the top of the body. Same as Google AI Mode, but more critical — ChatGPT’s context window is bounded and the synthesizer often pulls only the first chunk.
5. Bing-side ranking. Because the federated query runs through Bing, classical Bing SEO still moves ChatGPT citations. Bing Webmaster Tools, IndexNow integration, and structured-data alignment matter here in a way they do not for Google’s AI Mode.

Perplexity — the legible engine

Perplexity is the most observable engine in the set. It exposes its source list, often shows you the rewritten queries it ran, and tends to cite more sources per answer than any other engine. Architecturally, it runs hybrid retrieval (lexical + dense) against its own crawled index, supplemented by federated queries to specialized back-ends (academic, news, video). It also exposes a Sonar API that lets you probe its retrieval programmatically — the cheapest, most legible target for an instrumented GEO program.

Optimization levers:

1. Allow PerplexityBot and Perplexity-User. Their crawl is aggressive when allowed.
2. Passage clarity above all else. Perplexity is the most extractive engine in the set; clean, self-contained passages with explicit subject-predicate-object structure win.
3. Citation-worthy distinctive data. Perplexity systematically prefers sources with original data, named statistics, and quantified outcomes over sources that aggregate or rephrase.
4. Inline citations to authoritative sources within your own content. Perplexity weights pages that themselves cite primary sources more heavily.
5. Schema.org type coverage. Particularly Article, Dataset, ResearchArticle, and HowTo.

Claude (Anthropic) with search — the conservative reader

Claude’s retrieval, when enabled, draws from a federated set of sources weighted toward institutional credibility. The engine is more conservative than Perplexity, frequently choosing fewer but more authoritative citations, and is the engine most likely to refuse a synthesis if the retrieved sources do not agree.

Optimization levers:

1. Allow ClaudeBot, Claude-Web, Claude-User, Claude-SearchBot.
2. Institutional anchoring. Citations to .gov, .edu, peer-reviewed sources, and well-known publishers weight your content as more credible. Pages with no citation surface lose to pages with citations to canonical sources.
3. Named author attribution. Claude weights byline-attributed content with author schema higher than anonymous content.
4. Explicit dates and version metadata. Claude is more sensitive to freshness signals than ChatGPT and tends to mark stale content with caveats in synthesis.
5. Original research and data. Claude is the engine most likely to cite a small site with original data over a large site with aggregated content. This is the engine where the content-moat thesis (Chapter 30) pays off fastest.

A practical optimization matrix

Most teams that try to optimize for all four engines simultaneously end up doing nothing well. The discipline is to choose two as primary and two as secondary, based on where your buyer actually spends decision time.