Abhord Quickstart Guide (2026 Refreshed Edition)
What’s new in this edition
- Cross‑model surveys now support consistent sampling and deduped aggregation, reducing variance when you compare LLMs side‑by‑side.
- Improved mention clustering merges near‑duplicates across models and time, so Share of Voice (SOV) is cleaner.
- Sentiment v3 improves handling of nuanced or mixed opinions; neutral isn’t a catch‑all anymore.
- Competitor tracking includes templates and alerting presets; setup is faster and more consistent.
- Recommendations updated for 2026: prioritize parity testing across LLMs, apply entity‑level exclusions, and instrument “closed‑loop” impact tracking on your pages.
1) Initial setup and configuration
- Create a workspace and roles
- Invite teammates under Settings → Workspace. Assign Roles: Admin (billing + settings), Analyst (create surveys, edit taxonomies), Viewer (dashboards only).
- If your org uses SSO, connect it first so permissions map cleanly.
- Define your entities (brands, products, people)
- Go to Library → Entities. Add your primary brand plus up to 5 competitors, and any product/model names.
- Add synonyms and common misspellings. Example: “Acme Solar,” “AcmeSolar,” “Acme PV.”
- Seed your tracking taxonomy
- Topics: support, pricing, performance, reliability, integrations, comparison, alternatives.
- Exclusions: generic phrases that trigger false positives (e.g., “apple pie” if tracking Apple).
- Choose LLM coverage
- Start with at least 3 engines to triangulate: one GPT‑family, one Claude‑family, one open‑weights or regionally popular model.
- Toggle “Consistent Sampling” so each run queries the same intents and result counts per model.
- Connect integrations (optional but recommended)
- Analytics (to measure impact): GA4/Looker.
- Publishing/ops: CMS webhook, Slack/Teams for alerts, Jira/Linear for follow‑ups.
- Privacy and guardrails
- Enable PII scrubbing for prompts and harvested outputs.
- Set data retention appropriate to your compliance posture.
2) Run your first survey across LLMs
- Create a new survey (Surveys → New)
- Objective: “How do LLMs describe Acme vs. competitors on reliability and price?”
- Audience: “General users seeking product recommendations.”
- Add intents (think of these as questions you’d ask an LLM)
- Examples:
- “Best [category] for small businesses?”
- “Top alternatives to [Your Brand]?”
- “Is [Your Brand] worth it in 2026?”
- Write 6–10 intents that reflect your funnel (discovery, comparison, purchase).
- Configure engines and sampling
- Select 3–5 LLMs; set per‑engine sample size (e.g., 50 responses per intent per model).
- Language: start with English; add locales only if you can act on them.
- Set cadence: One‑time now, then schedule weekly for drift detection.
- Run a pilot
- Run 1–2 intents first to validate: check for false positives, spammy outputs, or brand‑unsafe content.
- Refine exclusions and synonyms, then run the full set.
- QC and annotate
- Use “Quick Triage” to mark off‑topic or AI‑hallucinated mentions; this trains the clustering.
3) Interpreting results: mentions, sentiment, share of voice
- Mentions
- Definition: Any surfaced text span where an entity from your library appears (directly or via synonym).
- Practical tip: Use the “Confidence” filter > 0.7 for executive reporting; keep 0.5–0.7 when exploring.
- Sentiment
- Abhord classifies at mention‑level and aggregates to intent, model, and time.
- Categories: Positive, Mixed, Neutral, Negative (with intensity scores).
- What changed: Mixed is now separate from Neutral. Expect fewer “Neutral” buckets for comparison content.
- Share of Voice (SOV)
- Formula (normalized): Your brand mentions / total valid mentions in the same topic window,