Your website chatbot has free will. This one does not.
A website answer engine that cannot make things up
It cannot generate, so it cannot invent. A clear question routes to a real page, an ambiguous one returns a short menu, and anything your site does not cover is declined, not answered. No model runs when it answers, so every reply is reproducible and you can read exactly why it landed where it did. It can still be unsure, and when it is, it asks or declines rather than guessing. No fabrication, which is not the same as no mistakes.
Try it, and try to break it
The engine runs on doloop's own pages below. A clear question routes. An ambiguous one returns a short menu. Something off topic gets a flat refusal instead of an invented answer. Skeptics should poke the last one first.
The first three route straight to a page, the fourth is genuinely ambiguous so it returns a menu, the last is declined. Same input, same answer, every time.
Three things it can do, and nothing else
A generated chatbot has an unbounded worst case: it can say anything, and you find out when a customer screenshots it. This has a floor under its worst behavior. The ceiling is lower, it cannot write you a poem or reason through a brand new question, but for a site that cannot afford a public wrong answer, a lower ceiling with a hard floor is the right trade.
Running in the wild
The same engine, seeded from each site's own pages, live on four sites. The gate measures two things: can it answer the site's real questions, and does it refuse off-topic ones. One site passes both cleanly. The other three each stress one axis, which is exactly where the gate earns its keep. Open one and use the launcher in the corner.
Doloop Digital clean fit
A dense single-thesis site where every page orbits one idea. It passes both checks: it answers its real questions, and it refuses off-topic ones. The kind of site this is built for.
Open doloopdigital.com →Ekrasworks small corpus
A craft and research brand with distinct essays and a product kit, each answered with its own title, summary, and image. It routes its real questions cleanly, but the vocabulary is small enough that an off-topic question can land near a real page by chance. So it is exactly the site the abstain monitor watches: strong on reachability, the refuse axis is the one to guard, and the live gate flags it for that.
Open ekrasworks.com →Sūtrakṛt Gītā hardest corpus
700 nearly identical verses, about the worst case there is. A verse number lands the exact verse, but a broad concept makes it ask rather than guess. The hard floor on display: it defers often here, and never invents a verse.
Open gita.ekrasworks.com →Hindu Spiritual Care sensitive subject
A nonprofit where a wrong or invented answer would do real harm. So it over-declines on purpose, and that is what keeps it on the safe side of the refuse axis: it routes courses, donating, and chaplaincy, and on anything it is unsure of it says it does not have a page, rather than improvise.
Open hsciglobal.org →It stays current, and stale never means wrong
The widget sits on every page you publish, so it notices when a page changes as visitors load it, then re-reads that page from the source before anything is re-frozen. We only ever freeze content we fetched ourselves, so the answer set always traces to a page we actually read. Until a change catches up, a stale page still routes correctly or falls to a menu or a decline. It is never confidently out of date. A generated chatbot has no such floor: a stale page becomes a fluent wrong answer, with old facts asserted at full confidence and no signal that anything is out of date.
It is for the job most website chat actually is
FAQ, support deflection, navigation, "what is this, where do I, how much, what is your policy." Most of the volume and most of the visitors. It understands paraphrase and typos, costs nothing per answer, and cannot leak or invent.
Transactional jobs (check my balance, cancel my order) and open-ended back and forth. Those need a model that can act and reason. We will tell you when your site is one of these, instead of pretending otherwise.
How it gets onto your site
We seed it with you, not at you. We read your pages, build the answer set, and run a two-part readiness check before anything goes live. Can your real questions reach the right page, and just as important, do off-topic questions fail to reach one. A site can pass the first and fail the second: too small or too uniform a corpus will answer well on-topic but route an off-topic question to a real-but-wrong page. We decline those, rather than ship a bot that guesses. If it passes both, you get a single tag to drop in and the launcher appears in the corner. No model runs when it answers, so it costs nothing per question, and the suggestions even track the section you are reading.
Today this is a hands-on setup with us, not a paste-and-walk-away install. The self-serve version waits until that readiness check is proven against real visitor traffic on sites we did not author.
Talk to us about your siteWant the honest version on one page, including where a generative chatbot still wins? Read the spec sheet.
Deterministic by construction. Each answer is a pure function of your page set and a published version, reproducible at that version. No model runs on the answer path. Off topic gets a graceful handoff, manipulation attempts get a flat refusal, and nothing is ever invented.