How to interpret AI sentiment scores for your brand (Mar 2026 Update 2)

Abhord Quickstart Guide (2026 Refresh)

This practical guide gets new users from zero to insights in under an hour. It reflects platform changes and recommendations as of March 2026.

What’s new in this edition

• Expanded model panel: native support for leading proprietary and open‑weight LLMs, plus per‑model weighting and cost caps.
• Better accuracy: semantic mention detection now uses entity graphs and cross‑model agreement scoring; fewer false positives.
• Aspect‑level sentiment 2.0: configurable taxonomy with intensity scoring and “mixed/ambivalent” handling.
• Share of Voice (SoV) v2: channel‑weighted SoV and author‑unique smoothing to reduce spam/repost bias.
• Always‑on competitor tracking: rules‑based alerts and weekly executive summaries.
• Team workflows: Slack/Teams sharing, Jira/Linear task sync, read/write roles, and data‑retention controls.

1) Initial setup and configuration

1. Create a workspace and project

• Go to Workspaces > New. Name it (e.g., “North America – 2026”).
• Add a project for each brand or market slice (e.g., “Abhord Core US”).

1. Connect data sources

• Click Data Sources > Connect. Start with:

- News/web (top publishers), forums/communities, app stores, YouTube captions, GitHub issues, Reddit, X.

• Set lookback window (e.g., 90 days) and refresh cadence (hourly or daily).
• If you bring internal data (tickets, NPS verbatims), enable PII masking.

1. Define entities and synonyms

• Entities > Brands/Products. Add:

- Primary brand name(s)

- Product lines, codenames

- Common misspellings and symbols (e.g., “A‑bhord”, “Abord”)

• Add exclusion rules for ambiguous terms (e.g., “sage” as herb vs. product codename).

1. Configure the model panel

• Models > Selection. Choose 3–5 LLMs for triangulation.
• Set weights (default equal). If cost‑sensitive, down‑weight highest‑cost model to 0.5.
• Enable “Cross‑model consensus” to suppress outlier classifications.
• Regionalization: set language coverage (auto‑detect on; add Spanish and French if applicable).

1. Sentiment and taxonomy

• Taxonomy > Aspects. Start with: Pricing, Performance, Reliability, UX, Support, Privacy/Trust.
• Sentiment scale: use −3 to +3 intensity. Set Neutral band to −0.5..+0.5 to reduce over‑classification.
• Turn on “Mixed” sentiment for multi‑aspect mentions.

1. Governance and retention

• Roles: Viewer (read), Analyst (query/edit), Admin (sources/models).
• Retention: 18 months default; shorter for regulated teams.

2) Run your first survey across LLMs

Goal: a fast “Brand Perception Pulse” covering the last 30–90 days.

1. Choose a template

• Surveys > New > Brand Perception Pulse (recommended start).

1. Define scope

• Timeframe: last 60 days.
• Channels: News, Reddit, App stores, X.
• Market/location: United States (or your primary region).

1. Prompting and seeds

• Use the default prompts; add 5–10 seed queries (e.g., “Abhord pricing complaint”, “Abhord vs speed”).
• Enable “Semantic expansion” to capture near‑matches and long‑tail phrasing.

1. Sampling

• Sample size: start with 1,500–3,000 items; stratify by channel (e.g., 30% Reddit, 30% News, 25% X, 15% App).
• Deduplication: turn on near‑duplicate collapse at 85% similarity.
• Cross‑LLM settings: majority‑vote classification; require 0.6 consensus to accept, else route to “Needs Review.”

1. Run and monitor

• Click Run. Progress shows cost, items processed, and model agreement (Cohen’s kappa).
• If kappa < 0.45 on any aspect, pause and adjust aspect definitions or add examples.

Example starting config

• Models: 4 total; equal weights; cost cap $50 for the run.
• Neutral band: −0.5