Abhord Quickstart Guide (2026 Refresh): From Setup to Actionable AEO
This refreshed edition reflects what’s changed in the LLM landscape since last year: models are more retrieval-aware, answer engines surface citations more often, and cross-model variance has grown with new safety and style defaults. You’ll find updated recommendations on sampling, weighting, and validation to keep your insights decision-ready.
1) Initial setup and configuration
- Create a workspace and roles
- Set an org workspace, then add teammates with roles (Admin, Analyst, Viewer). Keep API keys and provider credentials scoped to the workspace, not personal accounts.
- Connect providers and answer engines
- Add your LLM and answer-engine connectors (e.g., model families from multiple vendors, plus web-enabled “answer engines”). Label each connector as “param-only” or “web-aware” so you can segment results later.
- Define entities once
- Add your brand, products, and known synonyms/misspellings. Include locale variants and ticker or category tags. Do the same for competitors you care about tracking.
- Establish canonical sources
- Upload or link to your canonical facts (product specs, pricing pages, docs). Mark them as “authoritative” so Abhord can test if engines cite or converge on these.
- Prep taxonomy and guardrails
- Sentiment scale: choose granular (e.g., -2 to +2) to reduce neutral pile-ups.
- Mentions: decide whether to count indirect references and model self-expansions.
- Safe prompts: store red-team prompts and compliance notes centrally; apply to all projects.
- Notifications and cadence
- Set weekly or monthly runs by default; enable Slack/email alerts for shifts beyond a defined threshold (e.g., +/–10% share of voice or sentiment deltas).
What’s new and recommended
- Segment by retrieval mode: keep “web-aware” engines separate from “param-only” models to understand whether fresh web content is driving outcomes.
- Calibrate sentiment per model family: new defaults mean some models skew “warmer” or “drier.” Use a small gold set to align scales.
2) Running your first survey across LLMs
- Clarify the objective
- Example: “How do major answer engines describe Product X’s pricing and key differentiators to SMB buyers in the US?”
- Draft a minimal, consistent prompt set
- Write 3–5 intent-aligned questions, each with:
- Audience and locale: “for US-based SMB owners”
- Time frame: “as of March 2026”
- Source preference: “use cited, verifiable information when available”
- Choose your model panel
- Select at least one model from three distinct families, plus one or two “answer engines.” This diversifies styles and retrieval behaviors.
- Sampling and controls
- Set n=5–10 responses per model-prompt pair for stability.
- Fix temperature for “fact” prompts (e.g., 0.1–0.3) and allow higher diversity for “discovery” prompts (0.5–0.7).
- Enable deduping and near-duplicate clustering so “templatey” answers don’t inflate mentions.
- Pilot, then launch
- Run a 10% pilot, inspect outputs, adjust prompt clarity or entity synonyms, then scale to full run.
- Save as a template
- Save the panel, prompts, and filters as “Brand Perception US/SMB” for scheduled re-runs.
Updated tip
- Include a “contrast” prompt: “What