The Story
A search box that answers back, and pays its sources.
Here is a strange thing about the internet in 2026: every publisher on Earth spent decades building an archive, and then generative AI came along and, functionally, offered to read that archive out loud to strangers for free. This is not great for the people who wrote the archive. Miso.ai's entire proposition is a rearrangement of that sentence. Instead of the AI eating your content, the AI becomes your content - a private answer engine, fine-tuned to your catalog, that only says things your material actually supports, cites where it got them, and, in the O'Reilly version of the deal, routes royalties back to the authors it drew from.
The company is called Miso Technologies, it answers to Miso.ai and also askmiso.com, and it is small in the way that load-bearing infrastructure is often small - roughly thirty people quietly sitting behind more than a billion searches and recommendations across the web. It is based in San Francisco, though its intellectual center of gravity is a few thousand miles east, in a research lab at Cornell Tech.
That lab matters, because it explains why Miso is a little different from the wave of "we wrapped an LLM" startups. The founders - CEO Lucky Gunasekara and CTO Andy Hsieh, joined by co-founder Scott Lloyd - came out of Professor Deborah Estrin's Small Data Lab, where the animating question was almost the opposite of Big Tech's. Not "how much data can we mine about a person," but "how little do we actually need." Their answer, built on NSF- and MacArthur-funded work, was to analyze the content with pretrained neural networks rather than surveil the user. Miso can personalize, the company says, 100% anonymously.
If you are a marketplace or an e-commerce site, that pedigree shows up as a real-time recommendation and personalization API - semantic intelligence plus clickstream analysis, driving personalized homepages, deals, trending picks, and emails. The pitch here is a revenue lift, sitewide, without the creepiness. But the part of Miso that has made it into press-gazette panels and O'Reilly blog posts is the newer act: Miso Answers.
Miso Answers is the ChatGPT-shaped box, except housebroken. You type a question in plain language; it returns a summarized answer built strictly from the customer's own content catalog, with citations attached, in the brand's tone. The company's word for the thing it is engineered not to do is "hallucination," and the whole design is a series of guardrails against it. If it isn't in your catalog, the model doesn't say it. Boring, on purpose.
"The R in RAG stands for royalties."
— Miso.ai's framing of how publishers should profit from AI, via O'Reilly's radar
That line - a joke that is also a business model - captures the whole company. RAG, "retrieval-augmented generation," is the industry term for grounding an AI in specific documents. Miso's twist is to treat retrieval not just as an accuracy trick but as an accounting one. If the answer was generated because of a particular author's chapter, the system knows it, cites it, and can pay for it. O'Reilly Media, Miso's earliest flagship partner, made its answer engine available to roughly 2.5 million paying subscribers and described the arrangement, in a phrase startups dream about, as "really lucrative."
The other marquee deployment is with Foundry, the publisher behind a fleet of technology titles. Miso is the vendor behind "Smart Answers," a generative-AI feature that lets readers of CIO.com and its sister sites - Computerworld, CSO, InfoWorld, Network World, and the consumer-facing Macworld, PCWorld, and TechAdvisor - ask questions and get answers drawn from those publications' own reporting. It is, in effect, the newsroom's archive learning to talk.
What can you actually do with all this? If you run a content business, you can bolt a trustworthy question box onto your site - the company advertises a JavaScript SDK that gets you live in about ten minutes - and turn a passive archive into something people interrogate. If you run a store, you can hand shoppers semantic search and recommendations that understand intent instead of keywords. And if you are simply a reader, you get the increasingly rare thing: an AI answer you can check, because it shows its work.
None of this makes Miso the only player. It competes with search-and-personalization incumbents like Algolia, Coveo, Constructor, Bloomreach, and Yext, and, on the generative side, with the general drift of Perplexity, Glean, and every RAG toolkit on GitHub. Its wager is that "trustworthy, cited, and yours" beats "fast and everywhere" for the specific customers - publishers, learning platforms, catalogs - who cannot afford to be wrong or to be scraped. So far, the customers who have said anything have said it is working.
