Every published rate comes from a dated edition file with its question bank, exact API model identifiers, response counts and coded aggregates. This page documents what the repository proves — benchmark design, collection contract, coding, validation and limitations.
Protocol
Each industry gets a curated bank of standardized buyer questions covering discovery, DIY-vs-hire, vetting, pricing, comparison, local availability, urgency and red-flag scenarios. Banks are frozen within the edition; the exact questions are published on each industry page.
The collection contract expects one API response per question and model (chatgpt: gpt-5-mini · claude: claude-sonnet-5 · gemini: gemini-3-flash-preview), using identical question wording within the edition window. Published pages disclose observed response counts against that expectation rather than silently treating missing responses as complete.
Responses are classified against the same 12 binary behaviors (definitions below). Per-model percentages are simple proportions over observed responses. The displayed equal-model mean gives each available model the same weight.
The import and test pipeline validates required fields, response coverage, percentage bounds and aggregate consistency. Missing responses are reported. This repository does not presently document double coding, reviewer counts or adjudication, so no human-review claim is made.
Each file names its edition and collection date. A future edition can update the stable page after passing the same checks, but no refresh cadence or historical trend is claimed until multiple retained editions exist.
Published study files expose the question bank, API model IDs, observed response counts, per-model behavior rates, question-level agreement and question-level disagreement. Raw answer text is not part of the current public dataset claim.
Rubric
| Behavior | Coded as present when… |
|---|---|
| Recommends hiring a professional | The answer advises engaging a professional rather than (or before) self-service. |
| Suggests DIY first | The answer proposes trying a do-it-yourself route before hiring. |
| Names specific providers | The answer names at least one identifiable brand, company or practitioner. |
| Gives price or cost info | The answer includes price points, ranges or cost frameworks. |
| Tells to check reviews | The answer instructs the buyer to consult reviews or ratings. |
| Tells to verify credentials | The answer instructs verification of licenses, certifications or affiliations. |
| Mentions case studies / portfolio | The answer suggests asking for past work or documented outcomes. |
| Mentions local proximity | The answer weighs geographic closeness in the choice. |
| Gives selection criteria | The answer provides an explicit checklist for choosing a provider. |
| Warns about red flags / scams | The answer warns about fraud patterns or warning signs. |
| Asks a clarifying question | The answer asks the buyer for more context before advising. |
| Recommends multiple quotes | The answer advises comparing several providers or bids. |
Limitations
AI outputs vary with model version, location, personalization, retrieval mode and time. Figures describe a defined sample within a defined collection window — a controlled snapshot, not an absolute or permanent ranking. Sample sizes are published with every statistic (current edition: 547 questions, 1,641 responses across 14 industries). The cross-industry ARI is an equal-industry mean, not a response-weighted population estimate. Small service panels are marked directional, and incomplete response coverage is disclosed.
Citation license
These statistics are free to cite with attribution and a link.Authority Specialist. “AI SEO Statistics (2026-07 edition).” AuthoritySpecialist.com. https://authorityspecialist.com/research/ai-seo-statistics/methodology