Salience

Track how often your brand appears in answers from ChatGPT, Claude, and Perplexity — and what you'd need to do to be recommended more.

Salience is a local-first, single-binary tool that benchmarks your brand inside LLM answers. It queries each configured model with the prompts your customers actually type, detects every brand mention in both the answer text and the model's cited sources, classifies sentiment, and shows you a per- prompt picture of where you're winning, losing, and invisible.

It also probes the LLMs for why competitors are winning and what you'd need to do to flip those recommendations — turning a measurement tool into an optimization tool.

---

Why this exists

Customers ask LLMs for product recommendations now. If ChatGPT recommends your competitor and never mentions you, that's a traffic source going dark. Traditional SEO tools measure Google rankings — they have no idea what's happening inside an LLM's answer.

Salience measures exactly that gap, locally, with statistical rigor, across every script and language a brand might appear in.

What it does

• Benchmark — sends each prompt to each provider N times, persists every response to a local SQLite file.
• Detect — finds every brand mention in the answer text and the model's cited URLs. Works in Latin, CJK (Japanese, Chinese, Korean), Cyrillic, Arabic, Hebrew, Thai, and Devanagari scripts. Script transitions count as word boundaries, so Toyotaの車 correctly detects Toyota.
• Classify — three-way sentiment per mention (positive / neutral / negative) using weighted keyword markers in English, Japanese, Spanish, French, and German.
• Quantify — Wilson 95% confidence intervals on every rate. Cells with n < 10 are flagged so you don't over-claim from noise.
• Attribute — every cited URL is grouped by domain and categorized (review site / forum / wiki / news / vendor blog / etc.). The source gap report surfaces domains driving competitor mentions that you don't co-occur with.
• Explain — probes the LLM with follow-up calls ("Earlier you mentioned X, why?") and aggregates the stated reasons per competitor.
• Advise — asks the LLM ("What would my brand need to demonstrate to earn this recommendation?") and persists a ranked action plan per losing prompt.
• Compare — salience diff between any two runs, with statistical significance flagged when 95% CIs no longer overlap.
• Visualize — a local web dashboard (salience serve) with light and dark themes, live SSE updates while a bench runs, and a project picker for tracking multiple brands at once.

Quick start

# Build (Go 1.25+)
make build

# Set up your first project
./salience init                  # writes salience.json + .env template
$EDITOR salience.json            # add brand, competitors, prompts, providers
$EDITOR .env                     # add OPENAI_API_KEY / ANTHROPIC_API_KEY / PERPLEXITY_API_KEY

# Take a measurement
./salience bench -dry-run        # cost preview
./salience bench                 # run it

# Look at the results
./salience report                # Markdown summary
./salience serve                 # open the dashboard at http://127.0.0.1:7878

The dashboard supports creating, editing, and deleting projects from the UI, and lets you trigger benchmark / explain / advise runs with cost-preview modals. Power users keep the CLI; everyone else lives in the dashboard.

Subcommands

Command	What it does
salience init	Write a starter config + .env template
salience bench	Query every provider for every prompt, persist results
salience report	Markdown / HTML / JSON / CSV report from stored data
salience runs	List every persisted run
salience show -run N	Inspect samples + mentions + sentiment + context for one run
salience diff -from A -to B	Per-cell movement between two runs with significance
salience sources	URL / domain leaderboard and the source-gap to-do list
salience explain	Probe the LLM for why it recommended each competitor
salience advise	Ask the LLM what you'd need to do to win each losing prompt
salience playbook	One Markdown doc combining sources + explain + advise
salience serve	Local web dashboard with live updates
salience project	Manage workspaces (list / new / show / edit / export / import / delete)

Architecture

salience/
├── main.go                          CLI dispatcher + signal handling
├── internal/
│   ├── cli/                         subcommand glue (init, bench, report, …)
│   ├── config/                      config schema + .env loader + validation
│   ├── detect/                      brand-mention detection, i18n, sentiment
│   ├── envfile/                     minimal .env loader (no overwrite)
│   ├── pricing/                     per-model USD pricing table
│   ├── provider/                    OpenAI / Anthropic / Perplexity over net/http
│   ├── report/                      rate math, confidence intervals, renderers
│   ├── runner/                      fan-out, retries, resume, cost estimate
│   ├── server/                      dashboard HTTP server + SSE + embedded UI
│   ├── sources/                     URL / domain attribution analysis
│   └── store/                       SQLite schema + queries (modernc.org/sqlite)
└── .github/workflows/ci.yml         test + cross-compile on push

Supported providers

Provider	Models	Notes
OpenAI	gpt-4o, gpt-4o-mini, gpt-4.1, gpt-4.1-mini, o3-mini	Web-search tool requested when available
Anthropic	Claude 3.5 (Sonnet, Haiku), Claude 4 (Haiku 4.5, Sonnet 4.5/4.6, Opus 4.5/4.7)	Web-search tool block (web_search_20250305)
Perplexity	sonar, sonar-pro, sonar-reasoning, sonar-reasoning-pro	Native search built in

All providers accept endpoint overrides via OPENAI_BASE_URL, ANTHROPIC_BASE_URL, PERPLEXITY_BASE_URL — useful for proxies and local LLM gateways.

Cost protection

Every paid action (bench, explain, advise) supports:

• -dry-run — preview the call count and estimated cost without spending
• -max-cost USD — abort if the estimate exceeds your cap
• -yes — skip the interactive confirmation prompt (for cron / CI)

The dashboard mirrors all three: clicking Run benchmark opens a cost- preview modal you must confirm before any provider is called.

Project status

Salience is alpha but operationally usable today. It is dogfood for zistica and the rough edges are documented in CHANGELOG.md. API surface may still change.

License

Salience is released under the GNU AGPLv3.

• Free for personal and internal use — self-host, modify, extend.
• Hosted SaaS / commercial integrations: the AGPL requires you to open-source any modifications you offer over a network. If that doesn't fit, reach out about a commercial license.

See LICENSE for the full text.

Contributing

See CONTRIBUTING.md. Bug reports, prompts for new provider adapters, and i18n sentiment-marker contributions are especially welcome.

Security

See SECURITY.md. Do not file public issues for security vulnerabilities.

---

Built and maintained by zistica · Fukuoka, Japan.