Vendor-neutral report measures the 2026 conversational AI stack with open data

A vendor-neutral report published on Ethora's research pages and announced in a late-August dev.to post sets out to measure what the conversational AI stack actually looks like in 2026, using primary data rather than analyst commentary or vendor-run surveys. The study draws on GitHub, Hugging Face, npm and PyPI figures, search demand, sampled AI-assistant answers, open job postings and talent data, and releases every dataset plus the collection scripts under a CC BY licence, with a Zenodo DOI, so each figure can be checked or challenged.

A five-layer model of the stack

The report's framing device is that every conversational AI product, whether a side-project bot or an enterprise assistant in a regulated industry, is assembled from the same five layers: interface, orchestration, model, data/memory and infrastructure. Naming those layers matters, the authors argue, because the interesting changes are happening at different layers at different speeds.

Agents took over the vocabulary

According to the report's search data, US searches for "ai agents" climbed from roughly 60 per month in early 2023 to about 50,000 per month in 2026 — nearly a 700-fold increase. The term now organises the whole category: it dominates search behaviour, shows up in the most active repositories, and "AI agent" already ranks among the top three role categories in the hiring data.

MCP became infrastructure in about a year

The Model Context Protocol had effectively no search volume before late 2024, yet reached tens of thousands of monthly searches within roughly eighteen months, and its developer libraries already sit among the most-downloaded packages the report tracked. Watching an interoperability layer go from nothing to taken-for-granted plumbing in that span, the authors note, is rare.

Self-hosting momentum, with an asterisk

On GitHub, the most-starred projects in the space are the ones for running models and chat interfaces on infrastructure you own: Ollama, llama.cpp, vLLM and Open WebUI. Judged by stars alone, self-hosting would look victorious. The authors resist that conclusion: this is momentum among developers and early adopters, while most enterprises are still adopting cloud-hosted AI, let alone running production loads on their own hardware. Whether commercial self-hosting follows, they write, is precisely what a re-run of the report in a year should reveal.

Discovery is shifting into AI answers

Of 28 buyer-intent search terms tracked, 24 now trigger a Google AI Overview. And when the authors sampled what AI assistants actually recommend, those recommendations were built disproportionately from a handful of third-party "best of" lists and aggregators rather than vendors' own pages. The report's conclusion for tool builders: being present on the lists that models read is becoming a discipline separate from ranking in classic search.

Eighteen systems, mapped by capability

The report also includes a "periodic table": a capability matrix of eighteen notable systems, checked against current documentation, arranged by layer and shaded by adoption. Two findings fall out of it. Tool use and MCP support are now table stakes, and proactivity — acting on a schedule or a trigger rather than only replying — is the sharpest distinction between a chatbot and an agent. No system scores highly across control, conversation and agency simultaneously, which the authors place as the category's current frontier.

Where the builders are

On talent, the report finds the US leads on absolute demand and supply, India is the biggest net exporter of practitioners, and small, densely concentrated hubs — Switzerland and Singapore — top the per-capita tables. The supply view is LinkedIn-based, which under-counts China; the authors flag the limitation and corroborate it against the Stanford AI Index.

Why it matters

Conversations about conversational AI usually run on vendor marketing and anecdotes. This report substitutes countable signals — search volume, package downloads, repository stars, job postings — and, unusually, publishes the raw data and scripts behind them under an open licence. Its findings are also actionable: if agents are the organising concept and MCP the default plumbing, builders know where interoperability expectations are heading, and if AI answers increasingly mediate product discovery, marketing budgets aimed at classic SEO are targeting a moving object. The methodology's openness may prove the most durable contribution: it establishes a measurable baseline, and a repeat run next year will show with numbers whether developer-led self-hosting turned into commercial reality and which of these curves bent first.