How audit data is collected
Every Magpire Audit and every data point in the aggregate Magpire Index Report starts the same way: a set of buyer-intent prompts submitted to four AI engines, with the responses parsed, deduplicated, and scored through a consistent pipeline.
AI engines tested
We test the four AI engines that accounting-firm buyers are most likely to use when searching for professional advice. Each engine receives the same prompt set so results are directly comparable within the scoring model.
ChatGPT (GPT-4o)
Primary general-purpose AI. The most widely used engine for buyer research queries in professional services.
Prompts: City + specialism queries, question-form variants ("best accountant in [city]"), and service-line prompts.
Claude (Sonnet)
Increasingly used for professional and analytical queries. Claude tends to provide more detailed, citation-heavy responses.
Prompts: Same prompt set as ChatGPT to ensure comparability across engines.
Gemini
Google's AI, tightly integrated with Google Search data. A citation here often reflects both organic ranking and AI-specific visibility.
Prompts: Same prompt set, plus location-anchored queries that activate Google's local knowledge graph.
Perplexity
Purpose-built answer engine with explicit citation rendering. Perplexity's cited sources are surfaced inline — the most transparent signal of the four.
Prompts: Same prompt set. Perplexity's responses include source URLs, which we cross-reference with the firm's domain.
Pipeline overview
- 1
Prompt generation
For each firm, we construct prompts combining their city, target specialism, and common buyer question patterns (e.g. "best accountant for [specialism] in [city]"). The same base prompt set is sent to all four engines.
- 2
Response collection
Each engine's response is captured verbatim — the full text of the answer including any firm names, URLs, and citations the engine chooses to surface.
- 3
Entity extraction
We parse each response to identify cited firm names, domains, and specific recommendations. A firm is recorded as "cited" by an engine if the response names the firm or includes its domain.
- 4
Scoring
Citation data flows into the five-category scoring model. The final Magpire Index (0–100) is the weighted composite of all categories.
- 5
Aggregation (public data)
Individual audit results are never published. Aggregated, anonymised statistics feed the Magpire Index Report and Benchmarks hub — each published cell contains at least five firms.
← Back to Methodology · Weightings & rationale
