Abhord Quickstart Guide (Refreshed March 2026)

This practical guide helps new Abhord users stand up a reliable LLM survey program, read the metrics that matter, and turn insights into action.

What’s new in this 2026 refresh

Model mix guidance: pair at least one frontier model with one open or cost‑efficient model to reduce bias and improve coverage.
Structured outputs by default: use JSON extraction prompts for stable mentions and sentiment; avoid free‑form when you plan to trend.
Position‑aware share of voice (SoV): weigh answers that appear earlier or in “single best” responses more heavily for decisioning.
Cadence and size updates: for directional reads, 50–100 answers per model is often enough; use Wilson intervals to judge significance.
Alerting thresholds: act on ≥5 percentage‑point SoV moves sustained across two runs, or ≥10pp single‑run spikes.

1) Initial setup and configuration

1) Create your workspace

Projects: One per brand, market, or product line.
Roles: Assign at least one Admin (settings), one Analyst (surveys/dashboards), and one Editor (prompts/entities).

2) Connect model providers

Bring-your-own keys for the models you plan to test (e.g., frontier + cost‑efficient/open models).
Set per‑provider rate limits and concurrency to avoid throttling during runs.

3) Define entities and synonyms

Brand entity: canonical name plus common variants (e.g., “Acme”, “Acme Co.”, ticker, product names).
Competitors: list each rival and their known aliases; add exclusions (e.g., “Acme Brick” if unrelated).
Save to your global dictionary so extraction is consistent across surveys.

4) Configure defaults

Output mode: JSON or structured spans for extraction tasks.
Sampling: temperature 0.2–0.4 for extraction; 0.6–0.8 for generative discovery. Keep top_p at 1.0 initially.
Token limits: cap max_tokens to prevent truncation; 512–1024 is safe for short QA.
Random seeds: fix a seed for reproducibility when supported.
Redaction: enable PII redaction if you’ll paste proprietary queries.

Pro tip: Add a staging project to test prompts and extraction rules before deploying to your main dashboards.

1) Pick a use case

Brand QA: “Who makes the best X?”, “Is [Brand] trustworthy?”, “Top alternatives to [Brand]?”
Navigational/transactional: “Where to buy…”, “Pricing for…”, “Compare [Brand] vs [Competitor].”
Informational: “What is [Category] and leading providers?”

2) Create a survey

Query set: 10–25 high‑intent prompts that reflect how real users ask. Include exact, comparative, and “best of” forms.
Models: choose 2–4 (at least one frontier + one cost‑efficient/open).
Samples: aim for 50–100 answers per model to get a stable first read.
Geography/language: specify locale to match your market.

3) Prompt design