Generative Engine Optimization with Python (Andreas Voniatis)（Z-Library）

【One-Line Pitch】 A data-science playbook for marketers and analysts who want to measure—and improve—how often AI assistants like ChatGPT, Gemini, Perplexity, and Claude mention and cite their brand. It replaces "publish more" instincts with Python pipelines, sampling, and dashboards. 【Book Arc】 - **Opening (~0%–5%)**: Frames the shift from SEO to GEO—AI engines synthesize and cite rather than rank—and previews a seven-chapter structure from topic research through measurement and future-proofing. - **Early (~5%–16%)**: Establishes why a single visibility score is misleading and introduces the four metrics to track: mentions, citations, position, and sentiment. - **Early (~16%–32%)**: Sets up the measurement pipeline's inputs—API keys, brand/domain alias maps, target and competitive prompts, sample counts, and rate-limit pacing. - **Middle (~32%–53%)**: Builds the extraction layer: URL/domain parsing, mention detection, position scoring, format classification, and a HuggingFace sentiment classifier scoped to brand-mention sentences. - **Middle (~53%–63%)**: Writes platform-specific adapters (OpenAI, Gemini, Claude, Perplexity) with retry/backoff and bounded conversation history, normalizing each API's citation output into a common record. - **Late (beyond ~63%)**: Excerpts do not cover the aggregation, visualization, and dashboard-export stages, nor the other six chapters listed in the table of contents. 【Book Arc】 (note: the excerpt set is heavily concentrated in one chapter; the arc above reflects what is actually present) 【Key Takeaways】 - **AI search is a citation problem, not a ranking problem** (Opening): Engines select and synthesize sources by information-gain standards, so keyword tactics miss the target; success means being mentioned and cited inside generated answers. - **Track four metrics, not one score** (Early): Mentions, citations, position within recommendation lists, and sentiment each reveal different failure modes—being listed among "companies to avoid" is a vivid example of why sentiment matters. - **Generative output is stochastic, so sample repeatedly** (Early): Because identical prompts yield different answers, the pipeline issues multiple requests per prompt (10 by default, up to 100 if fast) to estimate rates rather than trusting a single response. - **Brand aliases are the linchpin of detection** (Early–Middle): Models rarely print domain names, so a domain-to-alias map plus word-boundary regex drives mention detection and position scoring; getting aliases right determines data quality. - **Use a trained sentiment model, not word lists** (Middle): A lightweight DistilBERT SST-2 classifier handles negation, irony, and domain phrasing that hard-coded positive/negative banks would miss, and runs only on brand-mention sentences to control cost. - **Normalize heterogeneous APIs behind adapters** (Middle): Each platform exposes citations differently (inline URLs, grounding metadata, search-result blocks), so per-platform query functions return a shared record shape with retries and history trimming. - **Bound cost and rate limits deliberately** (Middle): Sleep intervals per platform, exponential backoff on 429 errors, and a six-turn history cap keep an 880-call run feasible and reproducible. - **Measurement lets you falsify best practices** (Early): A repeatable framework shows whether GEO activities actually move metrics—and tests the stream of online advice rather than accepting it. 【Reading Tips】 - **Deep-read the pipeline chapter**: The code walkthrough is the book's practical core; type out the helper functions rather than skimming, since alias mapping and position scoring are easy to get subtly wrong. - **Skim the setup boilerplate**: API key configuration and import blocks are mechanical; focus attention on parameter choices (sample size, sleep seconds, history cap) and why they exist. - **Treat the prompt lists as a template**: The construction/NetSuite example shows how to segment research-stage questions from commercial queries—adapt the structure to your own domain. - **Plan for cost before running**: Paid API subscriptions and hundreds of calls per run are prerequisites; prototype with fewer prompts and samples first. 【Coverage Limits】 The available excerpts cover only the measurement chapter (Chapter 6 in the final numbering), and even there the aggregation, visualization, and dashboard stages are largely absent. Chapters 1–5 and 7 are listed in the table of contents but marked unavailable, so this guide cannot describe their content.