HVAC company · accuracy problem

Background

A second HVAC business that was already lightly present in AI — around 4% visibility — but carrying a hidden liability beneath that number.

The challenge

The models were stating inaccurate facts about it — wrong service details and outdated specifics — at the exact moment a customer was deciding who to call. Low presence was the smaller problem; being described incorrectly was the costly one.

What we did

We fixed accuracy and presence together:

• Traced where each incorrect claim originated and corrected it at the source
• Rebuilt the service and company information into clear, current, structured facts
• Earned fresh citations across ChatGPT, Perplexity, and Google AI so models had accurate, attributed data to cite
• Re-tested the same factual prompts over time to confirm the corrections held

Results

Over four months AI visibility more than tripled, from 4% to 14%, and the share of answers stating correct facts rose from roughly a third to almost all. Presence and accuracy moved together, so more visibility meant more correct visibility. Figures are rounded measured estimates, not exact daily rankings. Results vary by market, model, prompt set, and implementation.

Measurement & evidence

Market:

HVAC company · accuracy problem

Period:

4 months

Measurement:

Fixed buyer-prompt panel, repeated across ChatGPT, Claude, Perplexity, and Google AI.

Primary metric:

AI Visibility — the share of relevant answers, across repeated runs of the fixed prompt panel, that name the business. How this is measured

Starting point → latest measurement:

4% → 14%

Actions completed:

Traced where each incorrect claim originated and corrected it at the source; Rebuilt the service and company information into clear, current, structured facts; Earned fresh citations across ChatGPT, Perplexity, and Google AI so models had accurate, attributed data to cite; Re-tested the same factual prompts over time to confirm the corrections held.

Disclosed here:

Category and market type, the measurement window, the metric definition, the provider set the prompt panel was run across, the before/after values, and the source-level work completed.

Withheld:

The client's name and domain, the verbatim prompt panel, and per-run answer text — the first because clients ask for anonymity, the last two because an engagement-specific panel would identify the client. Clients receive all three in their own reporting.

Interpretation:

Measured association after the completed work; not a guarantee that every change caused the full movement.