optifeed-radar

name

optifeed-radar

description

Measure whether the models behind ChatGPT, Perplexity, Gemini, or Claude recommend a brand or specific named products, audit a site's AI-readiness, generate buyer questions, inspect cited sources, or compare saved AI-visibility runs. Use for AI visibility, GEO, AEO, AI-SEO, brand recommendation, product recommendation, competitor share-of-voice, and readiness-audit requests. Run Optifeed Radar locally through its CLI or MCP server. The readiness audit uses no API keys or AI calls; visibility and product checks use the user's own provider keys and spend their API credit.

Optifeed Radar

Use Optifeed Radar to collect inspectable, point-in-time evidence about what AI engines recommend. Keep readiness findings separate from recommendation visibility: they answer different questions and have different scores.

Choose a surface

• Prefer the Optifeed Radar MCP tools when they are available in the current AI agent. They return structured output and apply non-interactive cost caps.
• Otherwise run npx optifeed-radar <command> with Node 20 or newer.
• From an Optifeed Radar source clone, run npx tsx src/cli/index.ts <command> instead.

Follow the workflow

1. Normalize the target to a bare hostname such as example.com. Do not pass a path, query string, credentials, or an unrelated URL.
2. Start with the free audit unless the user asks only for a paid visibility run. Use audit_store over MCP or npx optifeed-radar audit <domain>. Report the AI-readiness score as readiness, not as the AI Visibility Score.
3. Decide what evidence the user needs:
   • Use a brand check to test whether unbranded buyer questions surface the brand and to measure position and competitor share of voice.
   • Generate the buyer-query pack first when the user wants to review or edit the sample before a full run. Query generation uses a small number of API calls.
   • Use a product check only for products the user explicitly names.
   • Use saved-run diff or sources commands when no new engine calls are needed.
4. Before spending API credit, inspect the available engines with npx optifeed-radar config when using the CLI. Get explicit approval for the engines, mode, and maximum cost unless the user's request already specifies all three. Never expose, print, or persist key values.
5. Default a first paid brand check to quick mode, the requested engines, and a conservative --max-cost. Add --yes only after approval so an AI agent can run the CLI unattended. Treat the cap as a strong bound, not a guaranteed exact ceiling: one in-flight call per engine can finish after the cap is reached, and Radar reports any overshoot.
6. Present the score with its sampling context, engine coverage, actual spend, skipped engines, grounded versus parametric split, and partial or capped status. Link conclusions to the raw answers or cited sources in the output.
7. Describe a single run as a dated estimate. Use at least two saved runs before claiming movement, and disclose when prompt or engine sets changed.

Map requests to commands and tools

Intent	CLI	MCP
Free site readiness	audit <domain>	audit_store
Brand recommendation visibility	check <domain>	check_visibility
Review buyer questions	queries <domain>	generate_buyer_queries
Named-product visibility	shopping <domain> --products "A, B"	shopping_check
Change between saved runs	diff <domain>	get_snapshot_diff
Cited domains and share of voice	sources <domain>	Read the saved check output
Key and state status	config	Not exposed

Useful brand-check flags include --quick, --engines openai,perplexity, --grounded, --max-cost 0.50, --json, --report report.html, --fail-under 50, and --yes. The MCP brand check accepts domain, optional engines, quick, and max_cost; it defaults to a $0.50 cap.

Handle products carefully

• Ask for the product list; do not infer or import it. Radar checks at most 10 named products per run.
• Treat input order as a stable tie-break only, never as the merchant's ranking.
• Add a descriptor for an opaque name, such as Aria 2: quiet home espresso machine, so product questions target the right category.
• Start with two or three products and a maximum cost. The MCP product check defaults to $0.20 per product.
• Explain that each product is tested on its own shelf. Product scores do not mean one listed product directly beat another.

For richer CLI input, use a YAML file:

products:
  - name: Aria 2
    aliases: [Aria II]
    descriptor: quiet home espresso machine
  - name: Presto X
    descriptor: fast dual-boiler espresso machine

Then run:

npx optifeed-radar shopping example.com \
  --products-file products.yml \
  --max-cost 0.50 \
  --yes

Configure provider access

Set at least one of OPENAI_API_KEY, ANTHROPIC_API_KEY, GOOGLE_API_KEY, or PERPLEXITY_API_KEY for buyer-query generation, brand checks, or product checks. Put keys in the calling process environment or a .env file in the working directory. Exported values take precedence. The free audit needs none.

Radar runs locally and has no Optifeed-hosted backend. Each provider key is sent only to its matching provider. Do not paste keys into a prompt, command output, report, snapshot, or source file.

Keep conclusions honest

• Treat scores as estimates from generated buyer questions, not panel data or a guaranteed ranking. AI engine answers vary between runs.
• Distinguish actual retrieval from requested grounded mode. A model can decline to search, and Radar reports retrieval answer by answer.
• State that provider APIs can differ from consumer chat interfaces.
• Do not claim continuous monitoring; Radar performs point-in-time checks.
• Do not claim catalog discovery or product-feed linting. Those capabilities are on the roadmap and waitlist at optifeed.com.
• A non-zero exit from check means the run measured nothing, not that the brand scored badly. All buyer prompts come from one judge call, so a failure there - no API credit, a rate limit, an unusable response - leaves nothing to ask and the run stops before querying engines. The reason is printed; report it and stop. Never present an unmeasured run as a low or zero visibility result. Declining the cost prompt also aborts, but exits zero.

Measured July 2026, a quick brand check cost about $0.09 on one engine, $0.41-$0.46 across four parametric engines, and $0.85-$1.09 across four with grounding. A measured two-product, four-engine grounded shopping run cost $0.70. Those figures predate the August 2026 switch of the OpenAI engine to a pinned reasoning model, whose reasoning tokens bill as output: measured 2026-08-13, a quick check on OpenAI alone cost $0.2297, roughly double that engine's earlier share. Provider prices and model behavior change, so use all of these only as dated planning ranges and rely on Radar's current estimate and actual-spend report.

More at optifeed.com