Methodology and Data Availability: The State of Conversational AI 2026
The State of Conversational AI, Q3 2026. This page documents how every figure in the report was produced so that any reader can check, reproduce, or extend it. It is the companion to the report's dated provenance log.
Principles
- Primary data only. Every published number is a specific, reproducible measurement we collected ourselves, not a figure quoted from another report.
- One source line per number. Nothing ships without a dated entry (metric, source, pull date, method, file) in the provenance log.
- Proxies, stated as proxies. Stars, downloads, search volume, and job counts measure interest and activity, not revenue or market share. We say so, and we report orders of magnitude and shapes rather than false precision.
- Point-in-time snapshots. All figures are dated (August 2026).
Sources and how each was measured
| Layer of the report | What we measured | Source | Tool / method | Date |
|---|---|---|---|---|
| Open-source adoption | GitHub stars, forks, watchers, issues, license, dates for a fixed set of 35 tracked projects | GitHub REST API (authenticated) | collect_adoption.py |
Aug 2026 |
| Model layer | Hugging Face model downloads (30-day) and likes per organisation | Hugging Face Hub API (public) | collect_adoption.py |
Aug 2026 |
| SDK / package usage | npm and PyPI last-month download counts for the SDK set | npm downloads API + pypistats | collect_adoption.py |
Aug 2026 |
| Search demand | Volume, difficulty, CPC, traffic potential, and AI-Overview presence for 28 buying-intent terms (US) | Ahrefs Keywords Explorer | build_demand_outputs.py |
Aug 2026 |
| Demand trend | Three-year monthly search-volume history for headline terms | Ahrefs volume-history | build_demand_outputs.py |
Aug 2026 |
| AI-answer discovery (AEO) | Share of voice and cited domains across 15 category prompts on ChatGPT and Google AI Overviews (US) | Ahrefs Brand Radar | MCP + Python parsing | Aug 2026 |
| Employment demand | Open-role counts by keyword and country, plus a US sample | Adzuna Developer API + free listings (Remotive, Jobicy, HN "Who is hiring") | scan_jobs.py |
Aug 2026 |
| Freelance / gig market | Active AI project counts and budgets | Freelancer.com public projects API | scan_jobs.py |
Aug 2026 |
| Talent supply | AI/ML practitioner counts by country (absolute and per capita) | LinkedIn Recruiter profile counts by job title (manual extraction) | analyze_talent.py |
Aug 2026 |
| Capability matrix | The "Periodic Table" capability cells, per-cell | Current official product documentation, verified in a parallel documentation pass | manual + doc research | Aug 2026 |
Reproducibility
The collection scripts are published in the open-data repository. Given the same API access, re-running them reproduces the processed CSVs the report is built from. The report site (main page plus one full-data page per table) is generated deterministically from those CSVs by build_report_site.py.
Limitations
- Search volumes and difficulty are estimates. Google withholds exact data; these are third-party model estimates, useful for shape and order of magnitude, not exact counts.
- Stars and downloads are noisy. Stars measure developer mindshare, not production usage; download counts skew toward small and quantised model variants, and one deployment can trigger many downloads. We report family-level and project-level trends.
- Closed commercial products have no repository to count, so the adoption section reads as "where open developer attention concentrates"; commercial names are covered through search demand and assistant answers instead.
- Job-market counts overlap across keywords and vary in coverage by market, so employment and supply-vs-demand figures are directional.
- AI-assistant sampling initially covers ChatGPT and Google AI Overviews; Gemini, Perplexity, and Claude populate on a monthly cycle.
- License terms change with releases. Any per-model license cell should be verified against the current model card before it is relied upon.
Data availability
Every figure links to a full-data page listing the complete underlying dataset. Raw pulls (JSON) and processed tables (CSV) are published in the open-data repository and available on request.
Versioning and citation
This is the Q3 2026 edition (data as of August 2026). See the report's "How to cite" box for the citation and license (CC BY 4.0). Corrections and source suggestions are welcome via ethora.com/contact.