🥔 Potato

English · 中文

Not a crawler, not an "AI-ranking truth detector", not a recommendation engine. Potato is a reproducible proxy measurement — every result represents only this engine, under this exact config.

Potato measures how visible your brand is inside Claude's web-search answers. It asks Claude a fixed set of frozen questions, collects every brand mention and source citation, then scores them with deterministic rules — no AI judge, no guesswork. The result is a reproducible, auditable report where every number carries a confidence interval.

1. Download AI-Visibility-Easy-Tool-windows.zip from the Releases page (right-hand side of this repo).
2. Unzip the whole folder first — don't run it from inside the .zip.
3. Double-click AI-Visibility-Easy-Tool.bat.
4. A small black console window opens and your browser launches the wizard automatically. (Keep that window open while you use the tool; close it to stop.)
5. Follow the 3 steps on the page:
   • Brand — your brand name, official domain, and category (optional: aliases, competitors).
   • Draft & review — generate the question set (a free template = $0, no key, or AI-drafted with your own key), edit anything you like, and run the built-in quality check.
   • Run & open report — choose the free mock preview ($0) or a real run against Claude, then open the HTML report with one click.

The portable build ships an official, python.org-signed Python (not a homemade .exe), so Windows Smart App Control won't block it — no code-signing certificate needed. The free preview needs no key and costs nothing.

_{Developers / macOS / Linux: pip install -e ".[dev]", then aivis gui (wizard) or aivis run --yes ($0 mock). Build the portable zip with python tools/build_portable.py. (The project is named Potato; the CLI command and Python package are aivis.)}

_{A single-file, offline HTML report (no external links, no scripts). Above: a $0 mock demo with 5 neutral brands. A live example: examples/single_brand_report.html.}

• Mention coverage — how often an answer names your brand by a clear name.
• Citation validity — share of cited links that are actually reachable.
• Owned vs earned citations — cited from your own domain vs third-party pages (counted separately, never blended).
• Share of voice — your share of all mentions across the measured brands.
• Stability — how consistent the result is across repeated rounds.
• Prominence — how early / prominently your brand shows up.
• Presence matrix — for each question, the strongest evidence reached: mentioned → cited → verified.
• Evidence chain — every cited URL: owned or earned, reachable or not, whether the page really mentions the brand, and the check method.
• A ranking with Wilson 95% confidence intervals, plus a health light.

• Deterministic scoring, no AI judge. Every counted number comes from fixed rules, not a model's opinion. AI is used only to draft questions (which you approve and freeze) and to narrate the report — never to decide the metrics.
• Conservative counting. Only a clear, unambiguous brand name scores. Abbreviations and ambiguous names are recorded but not counted — the tool never rounds up in your favor.
• Three evidence tiers — mentioned → cited → verified — and only verified enters the core score. "Verified" means Potato actually fetched the cited page and confirmed it is reachable and really mentions the brand.
• Independent citation check. Potato re-fetches the third-party pages Claude cited and checks them itself (dead links, mismatches) — it does not take Claude's word for them.
• Clean room. Every probe is a fresh call with no history and no personalization; region, language and temperature are pinned — the same question under the same conditions, every time.
• Honest uncertainty. Every ratio is shown with a Wilson 95% interval. When two brands' intervals overlap, Potato says "not distinguishable" instead of inventing a precise ranking.
• No leading the witness. 24 frozen questions (8 discovery / 6 comparison / 6 scenario / 4 brand-defense); the first 20 never name any brand, and your own brand is drafted blind — the AI is never told which brand is yours.
• We pinned the search tool that actually returns citations. A newer web-search tool version silently dropped inline citations — we caught it empirically (0 vs 12 citations) and pinned the version that keeps them.
• Injection-safe. Fetched web pages and AI answers are treated as untrusted data — never fed back to the model as instructions; HTML is sanitized.

Some categories are answered differently depending on location. Potato pins a single region and language (e.g. US / en-US) and runs every probe clean-room, so a run is one consistent national-level view — not a blur of different cities, and not your personalized results. Different regions are measured as separate strata and are never averaged together into a single misleading number.

Right now Potato connects to Claude (with web search) only, so what you see is your visibility on Claude specifically — not "AI in general." But the engine is provider-agnostic by design: every model sits behind one adapter interface, with no model-specific branches in the pipeline. Other AIs can be added later — or you can write your own adapter — without rewriting the engine.

• The free preview (mock) needs no key and costs nothing.
• A key is needed only for the two steps that actually call Claude: AI question drafting (optional) and a real measurement.
• You bring your own Anthropic API key, and Anthropic bills you directly.

• The author charges you nothing — ever. Potato is open source (MIT).
• A full real check is hard-capped to a budget you set — on the economy (Haiku) tier it typically lands around $5 or less, with higher-quality tiers (Sonnet / Opus) costing more. Potato estimates the cost before it starts and stops before it can overrun. That money is Claude usage paid to Anthropic, never to the author.

• Runs only on your machine (127.0.0.1). There is no cloud server, and nothing is uploaded.
• Your key lives in memory only, passed to a local subprocess through an environment variable — never written to disk, never logged, never placed on a command line.
• It is sent only to api.anthropic.com (your own account) — never to any author server. Zero telemetry.
• Open source and auditable; the report's "network destinations" panel shows exactly where the run connected, so you can verify all of the above yourself.

• Measure your own brand from the CLI / Claude Skill (SETUP mode): aivis init scaffolds a focal config → aivis validate checks it → aivis estimate prices a run → aivis run executes it. See SKILL.md.
• Project map: SKILL.md · ARCHITECTURE.md · CLAUDE.md (behavioral red lines) · src/aivis/contracts/ (data contracts) · configs/demo/ (neutral 5-brand demo).
• Quality bar: contracts frozen, three-layer storage (raw is append-only), deterministic pure scoring, golden tests, CI guards (ruff / mypy / import-linter / gitleaks / pytest) — 254 tests green.

MIT.