/ai-search-rescue

name: ai-search-rescue
description: 'Use when a site has lost visibility in AI search surfaces (Google AI Overviews, Google AI Mode, ChatGPT citations, Perplexity citations) or wants to systematically work toward being cited rather than just ranked. Trigger phrases include "AI Overviews recovery", "ChatGPT zitiert uns nicht mehr", "wir sind aus den AI Overviews rausgeflogen", "generative engine optimization", "GEO strategy", "AI search visibility", "warum zitiert AI Overview den Wettbewerber statt uns", "Perplexity-Citations aufbauen". Anti-Use: NICHT für klassische SERP-Rankings — dafür existieren `seo-outreach-report` und `competitor-deep-audit`.'
user-invocable: true
argument-hint: '[domain]'
allowed-tools: [Read, Grep, Glob]
license: MIT
metadata:
  author: Max Schottke
  version: '0.5.4'
  category: marketing

AI Search Rescue

Overview

A framework for recovering and growing visibility in AI search surfaces: Google AI Overviews, Google AI Mode, ChatGPT, Perplexity, Bing Copilot, Claude.ai search. The mechanics differ from classical SERP-ranking: instead of position 1-10 of a result list, you are competing to be the source the LLM cites in a generated answer.

Two operational findings drive this skill:

1. **AI citations may lead classical SERP recovery by several weeks (unproven, observation rescinded).** In one mid-sized e-commerce recovery, AI-Mode + AI-Overviews + ChatGPT citations started moving up roughly 2–6 weeks before the Sistrix VI did. The observation was later rescinded: the rise could not be separated from a plateau that preceded further updates, and the visibility index it was compared with turned out to be a poor proxy for that site's clicks. Treat it as a hypothesis to test, not as an indicator. If you observe the same or opposite pattern in your own recoveries, please contribute the data point back to `LESSONS.md`. 2. **The optimization isn't keyword targeting, it's passage targeting.** AI surfaces extract short passages (one sentence to one paragraph) that directly answer a sub-question of the user's prompt. If your page has no extractable passages, you don't get cited even when you rank #1 organically.

When to use

• A site that recovered organic rankings but is invisible in AI Overviews
• Owner says "ChatGPT keeps recommending our competitor, never us"
• Site has thin or unstructured content that AI surfaces can't extract from
• Post-Core-Update recovery work — track AI citations as leading recovery indicator
• New domain/topic launch — set up for AI-citability from day one
• B2B / YMYL site (medical, legal, financial) — AI surfaces over-index on authoritative sources here

**Don't use for:**

• Pure SERP-position optimization (use `seo-outreach-report` + `competitor-deep-audit`)
• Local-pack / Map-pack work (different mechanics)
• Pure e-commerce product-page conversion (different lever)

How AI surfaces choose what to cite

Different surfaces, similar mechanics. The model receives a user query, retrieves candidate documents (via search index or RAG), and assembles a generated answer. Two questions determine if you make it in:

1. **Does the retriever surface your page?** This is classical SEO — backlinks, relevance, freshness, indexability. 2. **Does the LLM find an extractable passage on your page that directly answers a sub-question?** This is new — it depends on structure, clarity, and the presence of definitive, source-able statements.

You can do everything right on #1 and lose on #2. Sites that rank #3 organically but get zero AI citations almost always fail on extractability.

digraph extractability {
  "Page indexable + retrieved?" [shape=diamond];
  "Has clear topic sentence per section?" [shape=diamond];
  "Definitive statements vs hedged generalities?" [shape=diamond];
  "Author/source attribution visible?" [shape=diamond];
  "Cited by AI surface" [shape=box,style=filled,fillcolor=lightgreen];
  "Skipped — not retrieved" [shape=box,style=filled,fillcolor=lightcoral];
  "Skipped — no extractable passage" [shape=box,style=filled,fillcolor=lightyellow];
  "Skipped — too vague to cite" [shape=box,style=filled,fillcolor=lightyellow];
  "Skipped — no author trust signal" [shape=box,style=filled,fillcolor=lightyellow];

  "Page indexable + retrieved?" -> "Has clear topic sentence per section?" [label="yes"];
  "Page indexable + retrieved?" -> "Skipped — not retrieved" [label="no"];
  "Has clear topic sentence per section?" -> "Definitive statements vs hedged generalities?" [label="yes"];
  "Has clear topic sentence per section?" -> "Skipped — no extractable passage" [label="no"];
  "Definitive statements vs hedged generalities?" -> "Author/source attribution visible?" [label="definitive"];
  "Definitive statements vs hedged generalities?" -> "Skipped — too vague to cite" [label="hedged"];
  "Author/source attribution visible?" -> "Cited by AI surface" [label="yes"];
  "Author/source attribution visible?" -> "Skipped — no author trust signal" [label="no"];
}

Before any optimization: consistency first, then measurement

Observed in one case (experimental, N=1 — **Maturity:** `experimental_n1`).

A shop that contradicts itself on basic product facts across template, content and structured-data layers gets those contradictions measured as its AI visibility. Sequence:

1. **Fact base.** Collect product facts from evidenced values only. A value without evidence is null, not an estimate. 2. **Consistency check.** Check every rendered page against the fact base, structured data included. Report occurrences and root causes separately; most hits trace back to a few sitewide strings. 3. **Freeze the prompt benchmark** (fixed prompt set, checksummed). A time series without a frozen prompt set is not a benchmark. A snapshot taken now is a *diagnostic* before-state: keep it to judge the corrections, and label it as measured on a self-contradicting site. 4. **Resolve the sitewide contradictions.** 5. **Take the reference baseline** after