Abhord Product Guide (Refreshed Edition, January 2026)
What’s new in this edition
- Broader multi-model coverage: updated presets reflect newer 2025–2026 model families and regional variants.
- Cleaner normalization: default de-duplication across LLM outputs and weighted share-of-voice (SOV) options by model prevalence.
- Aspect-based sentiment: sentiment now supports entity- and feature-level facets (e.g., “pricing,” “UX,” “support”).
- Guided surveys: quick-start templates for brand perception, category discovery, and competitor benchmarking.
- Alerts and baselines: out-of-the-box spike detection and weekly SOV baselines for competitor tracking.
- Better governance: project roles, API key scoping, and audit trails to keep teams compliant.
1) Initial setup and configuration
- Create a workspace and project
- Workspace: for your organization; invite teammates with roles (Admin, Editor, Viewer).
- Project: one per brand or initiative. Name it clearly: “Brand X – 2026 Q1 GEO.”
- Connect LLM providers
- Bring your own keys for the providers you plan to survey. Set per-provider rate limits and cost caps.
- Tip: enable at least three families (e.g., one US-centric, one EU-centric, one open-weight via API) to reduce single-model bias.
- Define entities and synonyms
- Add your brand, products, features, and canonical competitors.
- Map synonyms/aliases and common misspellings. Group sub-brands under a parent when you want unified reporting.
- Configure compliance guardrails
- Turn on PII suppression and profanity filters in prompts and outputs.
- Enable logging redaction for sensitive entities (e.g., internal code names).
- Choose your default taxonomy
- Use the “Brand/Competitor/Category/Features” taxonomy. This powers mention extraction, facet sentiment, and SOV.
- Set baselines and alerts
- Historical window: last 30–90 days is a good starter baseline.
- Alerts: enable “mention spike,” “negative sentiment surge,” and “SOV drop” notifications in Slack/email.
2) Run your first survey across LLMs
- Pick a template
- Start with “Brand Perception (General Market)” to validate entities and extraction.
- Draft prompts strategically
- Use unbranded discovery prompts: “What are the leading solutions for [your category] and why?”
- Add branded perception prompts: “What are common pros/cons of [Brand] for [use case]?”
- Include grounding variants: “Cite sources where possible,” and an anti-hallucination nudge: “If unsure, say so.”
- Select models and sampling
- Choose 4–6 LLMs across providers/regions. Start with 50–100 generations per model for statistically useful early reads.
- Enable “temperature sweep” (e.g., 0.2, 0.6) to capture both deterministic and creative variance.
- Run with normalization on
- Turn on cross-model de-duplication and answer clustering. This reduces identical or near-identical answers from similarly trained systems.
- QA the pilot
- Review 25–50 outputs across models. Confirm mention extraction highlights the right entities and that synonyms resolve correctly.
- Adjust prompts or synonyms, then scale to your full sample.
3) Interpreting results: mentions, sentiment, share of voice
- Mentions
- What it is: total extracted references to your entities across model outputs within the time window.