No black box. Here's the whole method.
Every number in a Hoss report comes from a measurement you could reproduce yourself, given enough patience. This page describes exactly how we build the prompt suite, run the tests, score the answers, probe your site, and re-measure it all monthly — including the things we deliberately don't do.
Six steps, in order, every time.
Build the prompt suite
We write the questions your buyers actually ask, in their language — not your positioning. Discovery ("who's the best ___ in ___?"), service-specific ("where should I get ___?"), comparison ("best alternatives to ___?"), and trust questions specific to your category. The suite is fixed once built, so this month's numbers are comparable to last month's.
Typical suite: 60–100 prompts, tuned to your market and services.
Run it across engines, multiple times
Each prompt runs across ChatGPT, Claude, and Gemini — and Perplexity where it matters for your market. AI answers vary run to run, so single runs prove nothing; we run each prompt multiple times and record every answer verbatim, including which sources the engine cited.
Score every answer
Each answer is scored against three defined metrics — mention, recommendation, citation share (defined below) — plus a competitor tally: who was named instead, how often, and which sources powered them. Every claim an engine makes about your business is fact-checked against your own site and records, and errors are graded by severity.
Probe the site like a machine
The engineering half. We request your key pages as each search and AI crawler — Googlebot, GPTBot, ClaudeBot, ChatGPT-User, PerplexityBot — and record exactly what each receives: status codes, firewall challenges, rate limits. We diff what an AI can read against what a human sees, inventory your structured data, and check the machine-readable layer (canonical URLs, sitemaps, llms.txt, feeds).
This is the layer most agencies never test — and where the most decisive findings usually live.
Connect causes to effects
Findings get linked, not listed. A team page that blocks AI crawlers explains an engine saying it "couldn't verify your credentials." A missing service page explains a zero-visibility prompt. Every miss in the report traces to a specific, fixable cause with evidence attached — and the causes get prioritized by revenue impact and effort.
Report and walkthrough
You receive the report — like this sample — plus a working session where we walk through every cause against your live AI results. The report always includes a prompt you can run yourself, because we never ask you to believe a chart you can't reproduce.
Metrics with definitions, not vibes.
Every scoreboard we publish uses these three, defined the same way every month. If a metric can't be defined precisely, we don't report it.
Mention rate
The share of tested answers where your business appears at all — named anywhere in the response. This measures whether the engines know you exist in this context.
Recommendation rate
The share of answers where the engine actively directs the buyer to you — names you as a pick, not a passing reference. This is the number that correlates with revenue, and it's usually much lower than mention rate.
Citation share
Of all business mentions across your suite — you plus competitors — the fraction that are you. This measures your share of the conversation, and tracks competitor movement over time.
Here's how a single row gets recorded.
From our fictional sample report — one prompt, three engines, one scored row. A full audit is 60–100 of these, run multiple times each.
Recommended Cedar Point Dermatology and The Glow Collective. Cited a local directory.
Recommended Cedar Point Dermatology and Arbor Springs Plastic Surgery.
Listed Bluebonnet among "other options" after recommending Cedar Point and Glow Collective.
What we deliberately don't do.
✕Guarantee citations or rankings
No one controls what AI engines say — anyone guaranteeing placement is selling you something else. We raise the probability, and the monthly scoreboard shows whether it's working.
✕Promise clicks or traffic
AI answers often keep buyers inside the conversation — the win is presence in the answer, not a click you can count. We're upfront that attribution here is genuinely murky, and we measure what's measurable.
✕Trick the engines
No hidden text, no prompt injection, no manipulation schemes. Those get sites burned when engines adapt. Everything we do makes your business more legible and more credible — durable by design.
✕Report single-run results
AI answers vary between runs and drift week to week. One screenshot proves nothing — which is why every number we publish comes from repeated runs of a fixed suite, re-measured monthly.
Audit first. Everything else follows from its findings.
We don't sell from a menu. The audit produces the fix list; the fix list scopes the work; the retainer keeps the scoreboard honest.
The audit
The full teardown described above: prompt suite, multi-engine testing, technical probe, causal analysis, report, and walkthrough.
The implementation
A fixed-scope engagement built from your audit's prioritized causes — crawler and schema repairs, entity work, authority pages, citation development.
The retainer
Monthly re-measurement of the same suite, continued page and citation work driven by the gaps, competitor tracking, and factual-error monitoring.
See the method applied to your business.
A 30-minute call. We'll show you what the prompt suite looks like for your market and what the audit covers.
Book a callOr see the deliverable first: view the sample report