Kioku: AI search share-of-voice checker
Are you in the AI's memory? Kioku (Japanese for "memory") shows whether AI assistants recommend your product. Tell Pip (the bird who runs the checks) which brand to check and, if you like, what to focus on ("BNB Chain, more specifically RWA, trading and AI"). Pip drafts the brief (what the brand is, its category, audience, website, rivals and focus areas) for you to confirm or correct in plain words. Pick how many questions to ask, from 1 to 100. Pip shows what the check will cost and how long it will take, and once you say go, asks ChatGPT, Claude, Gemini and Perplexity the questions your buyers ask, streams each answer into the chat as it lands, and builds a report:
- Share of voice, overall and for each assistant
- By topic: share of voice, rank, rivals, cited sources and next steps for each research topic
- Ranking: average position, mention rate and first-place answers
- Cited sources: which domains the assistants trust, including your site and rivals' sites
- Assistant by assistant: how each one treats you and who it favours
- Prompt by prompt: every question and answer, filterable by topic and by where you're missing, with the full text one click away
- Next steps: rule-based recommendations, plus sentiment, competitor gaps and the prompts you're missing from
It works with no API keys. Any assistant without a key answers with realistic demo text and fictional rivals, so you can try the whole flow straight away. Add keys to ask the real assistants.
Run it locally
Requires Node.js 20+ and pnpm.
pnpm install
pnpm dev
Open http://localhost:47321. The sample report lives at /report/demo.
Locally you don't need Redis: run state, the daily budget and rate limits are kept in the server process and reset when it restarts. To use a real Redis locally (for example the one your Vercel project uses), put its REST URL and token in .env.local as UPSTASH_REDIS_REST_URL and UPSTASH_REDIS_REST_TOKEN, or pull them with vercel env pull .env.local.
Connect real assistants
Copy .env.example to .env.local and add any keys you have, then restart pnpm dev. Each assistant switches to live answers on its own. The rest stay in demo mode, and the report labels which is which.
| Assistant | Key | Optional settings |
|---|
| ChatGPT | OPENAI_API_KEY | OPENAI_MODEL (default gpt-4.1-mini), OPENAI_WEB_SEARCH |
| Claude | ANTHROPIC_API_KEY | ANTHROPIC_MODEL (default claude-haiku-4-5), ANTHROPIC_WEB_SEARCH |
| Gemini | GEMINI_API_KEY (or GOOGLE_API_KEY) | GEMINI_MODEL (default gemini-2.5-flash), GEMINI_WEB_SEARCH |
| Perplexity | PERPLEXITY_API_KEY | PERPLEXITY_MODEL (default sonar) |
With direct keys, web search is off by default for ChatGPT, Claude and Gemini, and Perplexity always searches. Set OPENAI_WEB_SEARCH, ANTHROPIC_WEB_SEARCH or GEMINI_WEB_SEARCH to true to get cited sources, at a higher cost per check. OpenAI web search needs a search model such as gpt-4o-mini-search-preview, and Anthropic web search must be enabled for your organisation. Set SOV_FORCE_MOCK=true to ignore all keys.
One key for everything: OpenRouter
Set OPENROUTER_API_KEY and every assistant without its own key runs through OpenRouter. A direct key always wins for that assistant.
| Assistant | OpenRouter model (default) | Override |
|---|
| ChatGPT | openai/gpt-6-luna | OPENROUTER_CHATGPT_MODEL |
| Claude | anthropic/claude-haiku-5.5 | OPENROUTER_CLAUDE_MODEL |
| Gemini | google/gemini-3.1-flash-lite | OPENROUTER_GEMINI_MODEL |
| Perplexity | perplexity/sonar | OPENROUTER_PERPLEXITY_MODEL |
Through OpenRouter, ChatGPT, Claude and Gemini search the web by default, like the real assistants. OpenRouter's web plugin uses each provider's own search, and the report reads the sources from the answer's URL citations. Set OPENAI_WEB_SEARCH, ANTHROPIC_WEB_SEARCH or GEMINI_WEB_SEARCH to false to turn it off for that assistant. Search costs extra: Gemini's native search is the priciest of the three. Perplexity always searches on its own. OPENROUTER_BASE_URL points everything at another OpenAI-compatible gateway.
Keys stay on the server. The browser only learns whether each assistant is live, and which model it uses.
How Pip drafts your brief
With a key, Pip starts by working out which brand you mean (POST /api/intake/brief). A small, fast model drafts the whole brief from what it already knows: what the brand is, its category, audience, websites (up to three it's sure of), up to five rivals, and the focus areas you named, phrased so they make sense for that brand. Pip asks "Do you mean BNB Chain, the Layer 1 blockchain behind the BNB token?" and shows the brief as a card. This takes about a second and a half.
- A name that could mean several things ("Mercury") gets a "which one?", with the options as quick replies.
- A brand the model doesn't know, where you said what it is ("Fernly, a plant care app for new plant parents"), is drafted from your words, with well-known rivals of that kind.
- A bare name the model doesn't know is looked up on the web through OpenRouter's web plugin (Exa, three results, about a cent). The lookup only counts if the results describe exactly the name you wrote, since a similar name is a different company. If it finds nothing, Pip asks what the brand is in a few words and drafts from that.
여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗
