We spent the last few weeks trying to answer a simple question with numbers instead of vibes: what does the conversational-AI stack actually look like in 2026, and what's changing?
The result is a free, vendor-neutral report built entirely from primary data. No analyst opinions, no vendor surveys. We pulled from GitHub, Hugging Face, npm/PyPI, search demand, sampled AI-assistant answers, open-job counts, and talent data, and every figure links to its full underlying dataset. The data and the collection scripts are open under CC BY, so you can reproduce or argue with any number.
Here's what stood out.
The stack is five layers
Every conversational-AI product, from a weekend chatbot to a regulated enterprise assistant, is assembled from the same five layers: interface, orchestration, model, data/memory, and infrastructure. Naming them makes the rest of the picture legible, because different things are happening at different layers.
Agents ate the vocabulary
US search for "ai agents" grew from roughly 60 searches a month in early 2023 to about 50,000 a month in 2026. That's close to a 700-fold rise, and "agent" is now the word the category organises around. It shows up in the search data, in the most-active repos, and directly in hiring ("AI agent" is already a top-three role category).
A whole protocol layer standardised in a single year
The Model Context Protocol ("mcp server") had effectively no search volume before late 2024, and reached tens of thousands of monthly searches within about eighteen months, with its developer libraries already among the most-downloaded packages we tracked. It's rare to watch an interoperability layer go from nonexistent to assumed-infrastructure this fast.
Self-hosting is real, with an honest asterisk
The most-starred open-source projects in the space are the ones that let you run models and chat on your own infrastructure: Ollama, llama.cpp, vLLM, Open WebUI. If you only looked at GitHub, you'd conclude self-hosting has won.
We tried hard not to overclaim here. This is developer and early-adopter momentum. Most enterprises are still adopting cloud-hosted AI, let alone running production loads on infrastructure they own. How far real commercial self-hosting follows is a trend worth tracking, not a done deal, and it's exactly the kind of thing a re-run of this report a year from now will show.
Discovery is moving to AI answers
Of the 28 buyer-intent search terms we tracked, 24 now trigger a Google AI Overview. And when we sampled what the assistants actually recommend, they build those recommendations disproportionately from a handful of third-party "best of" lists and aggregators, not from vendors' own pages. Being present on the lists the models read is becoming a distinct discipline from ranking in classic search. If you build developer tools, this is the shift to internalise.
The Periodic Table
We also verified a capability matrix of eighteen notable systems against their current documentation, laid out by layer and shaded by adoption. Two findings fell out of it: tool use and MCP are now table stakes, and proactivity (acting on a schedule or trigger, not just replying) is the cleanest line between a chatbot and an agent. No single system is strong on control, conversation, and agency at once, which is where the category's frontier sits.
Where the builders are
Briefly, on talent: the US leads on absolute demand and supply, India is the largest net exporter of practitioners, and by concentration the small dense hubs lead (Switzerland and Singapore top the per-capita tables). Our supply view is LinkedIn-based, so it under-counts China, which we flag in the report and corroborate against the Stanford AI Index.
It's all open
The point of doing this from primary data was to make it checkable. Every table links to its full data, the methodology and every source are documented, and the datasets plus the scripts that produced them are public:
- Report: https://ethora.com/research/state-of-conversational-ai-2026-q3/
- Data + scripts (CC BY): https://github.com/dappros/state-of-conversational-ai
- DOI: https://doi.org/10.5281/zenodo.22131590
If you spot something wrong, or a source we should add for the next edition, I'd genuinely like to hear it. That's the whole idea.
Top comments (0)