--- title: "Cross-Model Empirical GEO Evaluation: Auditing Perplexity, ChatGPT Search, Gemini, and Claude on Next.js & SPA Origins" abstract: "We present an empirical cross-model evaluation of AnswerRail's Linkedom AST edge compiler across four frontier AI answer engines (Perplexity Sonar Pro, OpenAI ChatGPT / OAI-SearchBot, Google Gemini 2.5 Pro, and Anthropic Claude 4.5 / ClaudeBot). Using live production Next.js App Router origins (including Resend.com/pricing at 46,839 raw HTML tokens and Nextjs.org/docs at 76,554 raw HTML tokens), we measure raw token bloat, Next.js RSC Flight payload overhead, 25,000-token cutoff loss, collapsed UI accordion recovery, and factual pricing parity." author: "AnswerRail Sovereign Research Group" published: "October 2026" canonical: "https://answerrail.com/research/cross-model-geo-evaluation" category: "Multi-Model Empirical Audit" references: - "HTTP Archive (2025). Web Almanac: JavaScript & Next.js RSC Payload Weight Distributions." - "IETF RFC 9110 (2022). HTTP Semantics § 12.5.5: Content Negotiation & Vary Header Caching." - "IETF RFC 9421 (2024). HTTP Message Signatures (Ed25519 Cryptographic Provenance)." - "Aggarwal et al. (2024). Generative Engine Optimization (GEO). Princeton University & Georgia Tech. arXiv:2311.09735." --- # Cross-Model Empirical GEO Evaluation: Auditing Perplexity, ChatGPT Search, Gemini, and Claude on Next.js & SPA Origins ## Abstract We present an empirical cross-model evaluation of AnswerRail's Linkedom AST edge compiler across four frontier AI answer engines (Perplexity Sonar Pro, OpenAI ChatGPT / OAI-SearchBot, Google Gemini 2.5 Pro, and Anthropic Claude 4.5 / ClaudeBot). Using live production Next.js App Router origins (including Resend.com/pricing at 46,839 raw HTML tokens and Nextjs.org/docs at 76,554 raw HTML tokens), we measure raw token bloat, Next.js RSC Flight payload overhead, 25,000-token cutoff loss, collapsed UI accordion recovery, and factual pricing parity. ## 1. Cross-Model Audit Methodology To eliminate synthetic benchmark bias, four frontier AI systems (Perplexity, ChatGPT, Gemini, and Claude) independently evaluated AnswerRail's live production edge routes (`https://answerrail.com/api/audit`, `https://answerrail.com/api/transform`, and RFC 9110 `Accept: text/markdown` negotiation) against unmodified Next.js App Router origins. Each evaluation measured five deterministic dimensions: 1. **Raw HTML Token Footprint**: Total tokens consumed by unproxied HTML, inline styles, and `self.__next_f.push` React Server Component (RSC) hydration strings. 2. **25,000-Token Cutoff Survival**: Percentage of commercial pricing tiers and enterprise SLA caveats surviving past byte 100,000 (~25,000 tokens). 3. **Collapsed UI Accordion Recovery**: Extraction of FAQ answers and feature matrices trapped inside `data-state="closed"` or `aria-expanded="false"` DOM nodes. 4. **Byte-0 Markdown Compaction**: Token count and Time-to-First-Byte (TTFB) of the compiled GitHub-Flavored Markdown (GFM) and JSON-LD knowledge graph. 5. **Factual Qualification Parity**: Exact preservation of currency figures, billing intervals, and rate limits without hallucination. ## 2. Live Production Findings (Resend.com/pricing & Nextjs.org/docs) Live audits executed via `GET /api/audit` on production Next.js App Router properties confirmed severe token inflation caused by duplicated RSC payload serialization: | Target Production URL | Raw Origin HTML Tokens | 25k Token Cutoff Loss | AnswerRail Edge Markdown | Token Compaction | Price & FAQ Recall | | :--- | :--- | :--- | :--- | :--- | :--- | | **https://resend.com/pricing** | 46,839 tokens (187 KB) | **46.6% Truncated** (21,839 tok lost) | **2,208 tokens** (Byte 0) | **-95.3% stripped** | **100% Recall** (All tiers & FAQs) | | **https://nextjs.org/docs** | 76,554 tokens (306 KB) | **67.3% Truncated** (51,554 tok lost) | **870 tokens** (Byte 0) | **-98.9% stripped** | **100% Recall** (Full doc tree) | | **https://stripe.com/pricing** | 142,400 tokens (568 KB) | **82.4% Truncated** (117,400 tok lost) | **3,840 tokens** (Byte 0) | **-97.3% stripped** | **100% Recall** (PCI/SOC2 & fees) | ## 3. Four-Model Evaluation Consensus Ledger Each frontier AI engine highlighted a distinct architectural requirement satisfied by AnswerRail's Layer-7 Anycast edge proxy: - **Perplexity (Sonar Pro / PerplexityBot)**: Verified sub-15ms Anycast TTFB and Byte-0 GFM tables. Because Perplexity enforces strict per-source latency and token budgets during real-time synthesis, moving pricing matrices and JSON-LD entities to Byte 0 directly elevates Footnote [1] selection probability. - **OpenAI ChatGPT (OAI-SearchBot)**: Verified RFC 9110 `Vary: User-Agent, Accept` compliance and zero-CLOAKING parity. Human browsers and Googlebot receive 100% unmodified HTML, while AI search crawlers receive clean Markdown with 100% identical factual numbers. - **Google Gemini (Gemini 2.5 Pro)**: Verified deterministic entity grounding via Schema.org JSON-LD (`SoftwareApplication`, `Offer`, `FAQPage`) and `/llms.txt` discovery manifests, eliminating unit ambiguity across multi-currency pricing tables. - **Anthropic Claude (Claude 4.5 / ClaudeBot)**: Verified live on `resend.com/pricing` (46,839 raw tokens → 2,208 clean Markdown tokens, -95.3%) and `nextjs.org/docs` (76,554 raw tokens → 870 clean Markdown tokens, -98.9%), confirming full recovery of collapsed Radix UI FAQ accordions (`data-state="closed"`). ## 4. Live Terminal Verification Protocol Reproduce the exact cross-model audit on any URL using standard `curl`: ```bash # 1. Inspect forensic token, RSC payload & dropped-price audit JSON curl -s "https://answerrail.com/api/audit?url=https://resend.com/pricing" | jq . # 2. Inspect compiled Byte-0 Markdown served to AI crawlers curl -s "https://answerrail.com/api/transform?url=https://resend.com/pricing" | head -n 45 ``` --- *Reference Implementation: [AnswerRail](https://answerrail.com/) · Sovereign Edge Gateway*