Komo Review

The pitch for AI search is that you get an answer instead of ten links. It is a genuine improvement for a large class of questions, and it introduces a new failure that ordinary search did not have.

A list of links makes its uncertainty visible — you can see that the sources disagree, or that nothing on page one is quite right. A synthesised answer resolves that disagreement silently and hands you one confident paragraph.

Komo's design responds to that in two ways: citing sources on every answer, and building an explicit route to actual people for the questions where no amount of searching will work.

Komo: Search That Cites Its Sources

• Three modes — Web for ordinary questions, People for reaching humans, Deep research for longer work.
• Firsthand finds the right people, then gets answers for what AI cannot answer, which is an unusually honest feature to build.
• Answers with sources is the stated promise, printed in the footer rather than buried in marketing.
• Komo Fast is the default model, with deep research gated behind a paid plan.
• An Earn programme and saved search history, with a private-search positioning.

What Separates An Answer Engine From Search?

Komo's footer states its promise plainly: answers with sources. That combination is the definition of the category.

Traditional search retrieves and ranks. It shows you where information lives and leaves the reading, the reconciling and the judgement entirely to you, which is slow and honest.

An answer engine reads several sources and synthesises. That is faster, and it moves the judgement — which source to believe, how to reconcile a contradiction — inside a system whose reasoning you cannot inspect.

Citations are the mitigation, and they only work if you use them. An answer with sources that nobody clicks is functionally an answer without sources, and the convenience of synthesis is precisely what discourages checking.

The practical discipline is to click through on anything you will act on or repeat. That is the difference between AI search making you faster and making you confidently wrong — the same discipline grounded professional tools like CoCounsel build their entire product around.

An answer with sources that nobody clicks is functionally an answer without sources.

Web, People And Deep Research Modes

Komo's query box exposes three modes, and making the user choose is better design than inferring it.

Web is the default and covers the ordinary case: a question with an answer that exists somewhere, needing a fast synthesis rather than a research project.

Deep research, which the interface notes requires a paid plan, is the longer form — more sources, more time, a structured output. It suits questions where you would otherwise have spent an afternoon, and it is wasted on questions you could answer in a sentence.

The reason explicit modes beat automatic detection is cost and expectation. A system guessing that your question needs deep research will sometimes spend several minutes on something trivial, and sometimes give a one-line answer to something that deserved twenty sources.

People is the third and the genuinely unusual one, and it is worth treating as a separate product rather than a mode.

Firsthand And Asking Actual People

Komo describes Firsthand as finding the right people, then getting answers for what AI cannot answer.

That phrase — what AI cannot answer — is a rare admission in this category, and it points at a real and permanent limit.

Language models can only surface what has been written down. An enormous amount of useful knowledge never was: what a tool is actually like to use daily, why a company made a decision, whether a supplier is reliable, what went wrong on a project nobody published a post-mortem about.

For those questions the honest answer is not a better search — it is a person who was there. Building a route to that inside a search product acknowledges the boundary instead of generating a plausible guess past it.

The obvious questions are about incentives and verification. How people are found, whether they are compensated, and how you know their expertise is real all determine whether this is genuinely useful, and the pages we read do not answer them. The Earn programme visible in the interface suggests some form of participant reward, which makes those questions more pressing rather than less.

• Click at least one source on anything you will act on or repeat.
• Choose the mode deliberately — deep research is wasted on simple questions.
• Use People-type routing for knowledge that was never written down.
• Ask how contributors are found and rewarded before relying on their answers.
• Keep a normal search engine for navigation, which answer engines do badly.

Where Answer Engines Are Worse Than Search

It is worth being specific about where synthesis is the wrong tool, because the failures are systematic rather than occasional.

Navigation is the clearest. When you want a specific page — a company's pricing, a login, a documentation section — you want the link, and an answer engine will describe the page to you instead.

Recency is the second. Anything that changed in the last few hours may be summarised from cached or older sources, and an answer engine rarely signals that its material is stale.

Contested questions are the third and most important. Where sources genuinely disagree, a list makes the disagreement visible and a synthesised answer flattens it into one position, usually the one most commonly stated rather than the one best supported.

And anything requiring judgement about source quality is worse, not better, because the weighing happened invisibly. Research-specific tools such as Felo face the same limits — this is a property of the category, not a fault in any one product.

Privacy, History And What Gets Stored

Komo is positioned around private AI search, with a History section and saved searches visible in the interface.

Those two things are in tension, and it is worth understanding how rather than assuming either.

Search history is unusually revealing. What somebody searches for over a year describes their health, finances, employment, relationships and interests more completely than almost any other record they generate.

Saved history is genuinely useful — being able to return to research you did three weeks ago is a real feature — but it means the record exists, which is the thing privacy positioning usually implies does not.

The questions to answer are concrete: whether history is retained by default, whether it can be deleted, whether it is used to personalise or train anything, and what the Earn programme does with participation data. None are addressed on the pages we read, and all are worth checking before making it your default engine.

The alternative posture is to keep the index yourself. Locally run assistants such as Khoj answer over your own documents without a history sitting on somebody else's infrastructure, which is a genuinely different trade rather than a better one — you gain control and give up the web-scale corpus a hosted engine reaches.

Using Komo For Research That Holds Up

The way AI search goes wrong is not dramatic. You read a fluent answer, absorb it, repeat it in a meeting, and only later discover the source said something narrower.

This sequence keeps the speed while preserving the checkable chain.

1. Pick the mode before typing — Web for facts, deep research for questions worth an afternoon.
2. Read the answer, then open at least one cited source before believing it.
3. For contested topics, run the same question twice and look for disagreement between answers.
4. Use a normal search engine when you want a specific page rather than an explanation.
5. Route questions about lived experience to people rather than to a model.
6. Check the history and retention settings before making it your default.

Answer Engine Terms

The vocabulary of AI search, and what each term actually implies.

undefined

undefined

undefined

undefined

undefined

undefined

Komo: Modes And Company