Abhord Quickstart Guide (March 2026 Refresh)
This practical guide helps new Abhord users go from zero to value in a single session. It reflects the March 2026 refresh with updated workflows, clearer metrics, and new recommendations.
What’s new since the last edition
- Unified Model Orchestrator: simpler way to run the same survey across multiple LLMs in one pass.
- Real‑time Mentions Stream: faster ingestion with smarter de‑duplication across sources.
- Sentiment 2.0: expanded tone categories (positive, negative, mixed, neutral) and stance detection (pro, anti, unsure).
- Normalized Share of Voice: apples‑to‑apples model normalization so models with higher verbosity don’t dominate.
- Competitor Watchlists: reusable entity/keyword bundles with alert thresholds and weekly rollups.
- Better Governance: project‑level roles, data retention controls, and audit trails.
1) Initial setup and configuration
- Create a workspace
- Name it after your brand or product line (e.g., “Northstar Analytics”).
- Add teammates with roles: Admin (billing + settings), Analyst (builds/runs surveys), Viewer (reads dashboards).
- Connect data sources
- Knowledge inputs: website sitemap, documentation, product feeds, FAQs.
- Tracking inputs: brand terms, product names, key people, approved synonyms and misspellings.
- Tip: Add canonical entity IDs (e.g., your Wikidata ID) to improve entity resolution.
- Choose models for analysis
- From the Model Catalog, select at least 3–5 diverse LLMs (general, open‑source, search‑augmented) to reduce bias.
- Set defaults: temperature (0.2–0.4 for surveys), max tokens, and a cost cap per run.
- Configure governance
- Turn on data retention limits (e.g., 180 days) if your org requires it.
- Enable PII redaction in ingestion and exports.
- Calibrate brand dictionary
- Add do‑not‑confuse terms (e.g., “Acme” ≠ “ACME Logistics”).
- Include competitor variants and international names for better recall.
2) Run your first cross‑LLM survey
Goal: Measure how top LLMs describe your brand, category, and competitors.
- Start a new Survey
- Template: “Brand + Category Perception.”
- Audience: “General LLM” (default).
- Models: pick at least 4 (e.g., one leading closed model, one fast API model, one search‑tuned, one open).
- Define prompts and tasks
- Core prompt example:
- “In one paragraph, who is [Brand], what do they offer, and who are their top competitors? Use concise, factual language.”
- Follow‑ups (multi‑turn):
- “List three reasons someone might choose [Brand].”
- “List three reasons someone might not choose [Brand].”
- Evaluation rubric (auto‑scored):
- Accuracy (0–5), Specificity (0–5), Helpfulness (0–5).
- Sampling and controls
- Runs per model: 25–50 for reliability (more for volatile prompts).
- Randomization: enable order randomization of brand/competitors to reduce position bias.
- Determinism: set temperature low (0.2–0.3) for baseline; run an additional high‑variance pass (0.7) for idea mining.
- Execute
- Set a budget cap; enable “fail open” retry once per model.
- Monitor run status; pause any model exceeding cost/time limits.
- Review outputs
- Use the “Answer Set” view to scan representative responses.
- Flag hallucinations; add missing facts to your knowledge inputs for the next run.
3) Interpret results: mentions, sentiment, share of voice
- Mentions
- Direct mention: exact brand/entity (e.g., “Northstar Analytics”).
- Indirect mention: brand handle, acronym, product codename.
- Unique mention rate = unique mention count ÷ total samples; helps spot over‑counting.
- Action: Low direct mentions across models often signals weak entity grounding—add structured data (FAQ, HowTo, Organization schema), strengthen About pages, and ensure consistent naming.
- Sentiment and stance
- Sentiment 2.0 tracks tone (positive/negative/mixed/neutral) and stance (pro/anti/unsure).
- Look for divergence by model: if one model skews negative, inspect its knowledge retrieval citations and time horizon.
- Action: Address recurring negatives (pricing confusion, feature gaps) with explicit, verifiable content and FAQs.
- Share of Voice (SoV)
- Definition: SoV = brand mentions ÷ total mentions across selected entities (normalized by model verbosity).
- Read the Model‑normalized SoV chart first; then view the Raw SoV to understand absolute chatter volume.
- Action: If a competitor dominates in “reliability” mentions, craft targeted pages and snippets that assert and evidence your reliability (SLAs, uptime dashboards, audits).
4) Set up competitor tracking
- Build Watchlists
- Entities: your brand, up to 8 competitors, and key category terms.
- Variants: include product lines, regional names, and ticker symbols.
- Queries and contexts
- Core intents: “best [category],” “[brand] vs [competitor