AutoAEO

Measure and optimize your visibility in AI search.

Measure how often AI assistants cite, recommend, and mention your content — then iteratively optimize it to rank higher. Tests against Claude, GPT, and Gemini simultaneously.

---

Why

Search is moving to AI. When someone asks ChatGPT, Claude, or Gemini a question, your content either shows up or it doesn't. Traditional SEO metrics don't capture this.

AutoAEO gives you a score (0–100) for how visible your content is to AI assistants, then rewrites it to score higher — in a loop, until it converges.

How it works

┌──────────────────────────────────────────────────────────────┐
│                           AutoAEO                            │
│                                                              │
│  1. Generate queries     "best project management tools"     │
│  2. Ask each model       Claude, GPT, Gemini (with search)   │
│  3. Extract citations    domains cited as sources            │
│  4. Adjudicate mentions  Claude Haiku → brand + sentiment    │
│  5. Aggregate            mentioned · cited · sentiment       │
│  6. Optimize (optional)  rewrite content → re-score → repeat │
└──────────────────────────────────────────────────────────────┘

Measurement

For a given topic and URL, the tool:

1. Generates realistic search queries using Claude (informational, comparison, recommendation)
2. Sends each query to every model with their native search tools enabled
3. Extracts cited domains from each response (source URLs and inline links)
4. Runs an adjudicator (Claude Haiku) over each response to detect brand mentions and classify sentiment as positive / neutral / negative — for the target plus the top-cited competitor domains
5. Aggregates three signals into a composite score (0–100):
   • Mentioned (50%) — % of responses that mention your brand
   • Cited (35%) — % of responses that cite your domain as a source
   • Sentiment (15%) — average sentiment when mentioned (weight scales up with mention count, redistributing to Cited when mentions are sparse)
6. Reports your score plus a competitive leaderboard — every adjudicated domain ranked by composite score, so you can see how you stack up against the brands AI assistants surface for the same queries

Optimization

Extends measurement with an iterative improvement loop:

1. Scrapes your page content (Readability)
2. Searches for competitor content (Exa API)
3. Scores baseline content against competitors
4. Rewrites content with Claude to maximize AI visibility
5. Re-scores — keeps improvements, reverts regressions
6. Repeats until convergence or budget exhausted

Models

Model	Provider	Search
Claude Sonnet 4.6	Anthropic	webSearch
Claude Haiku 4.5	Anthropic	webSearch
GPT-5.5	OpenAI	web_search
GPT-5.4 Mini	OpenAI	web_search
Gemini 3.1 Pro	Google	google_search
Gemini 3 Flash	Google	google_search

Each model uses its native search integration via the AI SDK, so responses reflect real search-augmented behavior.

Quick start

1. Clone and install

git clone https://github.com/ymansurozer/auto-aeo.git
cd auto-aeo
pnpm install

Requires Bun (runtime) and pnpm (package manager).

2. Set up API keys

cp .env.example .env

ANTHROPIC_API_KEY=sk-...
OPENAI_API_KEY=sk-...
GOOGLE_GENERATIVE_AI_API_KEY=...
EXA_API_KEY=...        # required for optimize only

You only need keys for the models you want to test. Skip a provider and those models are excluded.

3. Measure

bun src/index.ts measure --topic "project management" --url "https://example.com"

4. Optimize

bun src/index.ts optimize --topic "project management" --url "https://example.com"

Commands

Command	What it does
measure --url <u> [--topic <t> | --queries-list <q>]	Measure AI visibility across all models
optimize --url <u> [--topic <t> | --queries-list <q>]	Iteratively optimize content for AI visibility
report --list	List past reports
report --session <id>	View a specific report

Options

Flag	Default	Description
--models <ids>	all	Comma-separated model IDs to test
--queries <n>	15	Number of queries to generate (when using --topic)
--queries-list <queries>	—	Comma-separated queries (alternative to --topic)
--runs <n>	3	Runs per query per model
--timeout <seconds>	60	Per-judge-call timeout in seconds
--max-iterations <n>	20	Max optimization iterations
--max-cost <usd>	—	Cost budget cap
--convergence-threshold <n>	0.02	Min improvement to continue (2%)

Output

All output goes to the output/ directory:

• report-{session}.html — measurement or optimization report with score breakdowns and a competitive leaderboard
• optimized-content-{session}.md — rewritten content (optimize only)
• experiment-log.jsonl — every judge call with full metadata

Architecture

src/
├── cli/                  CLI layer (Commander.js)
│   └── commands/         measure, optimize, report
├── measurement/          Measurement runner
├── optimization/         Optimization loop, content rewriter, search simulator
├── queries/              Query generation (Claude)
├── models/               Model registry, judge (AI SDK)
├── scoring/              Citation extraction, brand adjudicator (Claude Haiku), aggregation
├── experiment/           JSONL experiment logging
├── scraper/              Content extraction (Readability)
├── search/               Competitor search (Exa)
├── report/               HTML report generation
├── config/               Zod schema, defaults
└── utils/                Cost tracking

Dependencies

Package	Purpose
AI SDK	LLM calls with native search tools
Exa	Competitor content search
Readability	Content extraction
Commander	CLI framework
Zod	Config validation
async-caller	Parallelized API calls with concurrency control

License

MIT