Abhord Quickstart Guide (Refreshed Edition)

This practical guide helps you stand up Abhord, run your first cross-LLM survey, interpret results, set up competitor tracking, and turn insights into action—fast.

What’s new in this edition:

• Clearer default schemas for mentions, sentiment, and share of voice (SoV)
• Stronger guidance on prompt versioning, language/region coverage, and confidence scoring
• Updated recommendations for multi-LLM portfolios and model variance control
• Practical QA steps to reduce false positives and entity drift

1) Initial setup and configuration

1) Create a workspace

• Invite teammates and assign roles: Admin (billing + providers), Analyst (projects + exports), Viewer (dashboards).
• Set your default region(s) and language(s). Start with your core market; add adjacent locales later to compare variance.

2) Connect providers and models

• Add API credentials for the LLMs you plan to survey (e.g., at least three from different vendors to reduce single-model bias).
• Tag models by purpose: “fast-scan,” “balanced,” “deep-reasoning.” You’ll sample across tags later.

3) Define your entities

• Add your brand, products, and company as canonical entities.
• Add known aliases, misspellings, and localizations (e.g., product nicknames, ticker, prior brand names).
• Optional: Add disambiguation notes (e.g., “Not the film; we’re the CRM platform”).

4) Guardrails and privacy

• Enable PII redaction in prompts and outputs.
• Turn on deduplication of near-identical mentions and activate hallucination checks (evidence validation) where available.

5) Baseline keywords and themes

• Seed themes like “pricing,” “support,” “security,” “innovation,” and “performance.”
• Add negative keywords (e.g., unrelated companies with similar names) to cut noise.

Pro tip: Save all the above as a “Workspace Template” so new projects inherit consistent settings.

2) Run your first survey across LLMs

Goal: Ask a portfolio of models the same structured questions and capture consistent, analyzable output.

1) Start a new project

• Choose a template: “Brand Health Scan” or “Competitor Landscape.”
• Select languages/regions. Keep English/US first for a clean baseline.

2) Write task prompts

• Keep questions neutral and specific. Example:

- “List notable brands in project management software and briefly explain why each is mentioned. Return JSON with fields: brand, reason, evidence, confidence.”

• Avoid leading language (“best,” “top”) unless the research objective is rankings.

3) Define an output schema

• Mentions: entity_canonical, entity_alias, evidence_snippet, source_hint (if provided), confidence (0–1).
• Sentiment: score (-1 to +1), polarity (negative/neutral/positive), rationale_short.
• SoV: auto-computed later; ensure mentions are structured for counting.

4) Choose models and sampling

• Select 3–5 heterogeneous models. Set sample size per model (e.g., 20–50 prompts each).
• Enable randomized prompt variants and order shuffling to reduce pattern bias.
• Fix temperature within a narrow band across models; keep one “exploratory” high-temperature model for breadth.

5) Quality controls

• Add a couple of “control” prompts with known answers to sanity-check consistency.
• Turn on automatic retries for rate limits and transient errors.
• Test-run 5 prompts; inspect outputs; then scale.

Run the job and wait for completion. You’ll land on the Results dashboard.

3) Interpreting results: mentions, sentiment, share of voice

Mentions

• Definition: Count of times an entity appears in structured results after alias normalization and deduplication.
• What to check:

- Canonical