Felo Review

Most AI search tools are built by English-speaking teams for English-speaking users, and the consequence is invisible until it matters.

Ask about a Japanese regulatory change or a Chinese manufacturing standard and you will often get an answer assembled entirely from English-language coverage of the thing, rather than from the thing itself.

Felo comes at search from the other direction. It is a multilingual platform first, and the rest of what it does — slides, agents, documents, an API — is built around the assumption that the research and the deliverable may not be in the same language.

Felo In Five Points

• Multilingual is the product, not a setting — Felo describes itself as a free multilingual AI search and creation platform, and the language reach is what most distinguishes it.
• Search plus Slides in one flow: Felo Slides turns a research answer into a presentation without an export-and-rebuild step.
• An agent mode and an LLM playground sit alongside search, for multi-step tasks and for comparing model behaviour.
• LiveDoc and an API extend it past the chat box into documents and into your own applications.
• There is a free tier, which lowers the cost of finding out whether the multilingual difference matters for your sources.

What Multilingual Search Actually Changes

Felo describes itself as a free multilingual AI search and creation platform, and that word does more work than it appears to.

There are two different things a tool can mean by multilingual. The weak version is a translated interface: the buttons are in your language, and the search still runs across predominantly English sources. The strong version is retrieval across languages — the tool finds Japanese sources for a Japanese question, or finds them for an English question and answers you in English.

The second is genuinely useful and much harder to build, because it requires the retrieval layer to work across scripts and search conventions that differ substantially from English.

The practical test is easy to run yourself, and worth running before committing. Ask about something where the authoritative sources are not in English — a regulation, a local market, a company that files in another language — and look at what it cites. If the citations are in the source language, the retrieval is real. If they are all English-language secondary coverage, you have the weak version.

For anyone working across Asian markets specifically, this is not a nice-to-have. The gap between what an English-first tool surfaces about Japan and what actually exists in Japanese sources is large enough to change conclusions.

The comparison to run is against whatever answer engine you already use. Perplexity is the sensible baseline, precisely because it is strong enough on general questions that anything replacing it has to win somewhere specific.

There are two things a tool can mean by multilingual: a translated interface, or retrieval that actually crosses languages. Only one of them changes your answers.

Felo Slides And The Research-To-Deliverable Gap

Felo Slides generates presentations, and its position inside a search product is the interesting part rather than the slide generation itself.

Think about what research for work actually ends in. Very rarely a satisfied curiosity; far more often a deck, a memo or a document that someone else reads. The search tool answers the question and then hands you a wall of text you must restructure by hand.

That restructuring is where most of the time goes, and it is also where most of the sourcing gets lost. Facts move from an answer with citations into slides without them, and the deck arrives looking more certain than the research was.

Generating slides directly from the answer closes that gap mechanically. Whether it closes it well depends on something you should check: does the deck keep the citations, or does it strip them? A sourced deck is far more defensible than a confident one.

Expect to edit heavily regardless. Generated slides are a reasonable first structure — the argument in order, the sections identified — rather than a finished deliverable, and treating them as finished is how errors reach an audience.

If the deck rather than the research is the hard part, a dedicated generator such as Gamma produces a better-looking result — at the cost of the research and the slides living in different places.

Agent Mode, LiveDoc And Where Each Fits

Felo's product set runs wider than search — an agent mode, an LLM playground, LiveDoc and an API — and knowing which to reach for saves time.

Search answers a question. Agent mode is for tasks with several steps that depend on each other, where the second search depends on what the first returned, and it is worth the extra latency only when the task genuinely has that shape.

The LLM playground is for comparison rather than production. It is where you find out whether a given model handles your kind of question well before building anything on it, and that is a more useful habit than most people practise.

LiveDoc extends the platform into documents, which matters if your output is prose rather than slides. The API is what makes any of it repeatable — the difference between a person doing research and a process that runs on a schedule.

Most users will live in search and Slides. The rest is worth knowing about at the point where the manual version starts repeating itself.

How To Evaluate Felo Against What You Already Use

The free tier makes evaluation cheap, and the way to use it is to test the thing that differentiates rather than the thing everyone has.

Every AI search tool answers general questions acceptably. Comparing them on that produces a tie and tells you nothing.

Run the multilingual test first, because it is the only question where the answer could be decisive.

1. Take three real questions where the authoritative sources are not in English, and run each through Felo and through whatever you use now.
2. Compare the citations rather than the prose — count how many are in the source language versus English secondary coverage.
3. Take one answer through to Felo Slides and check whether the citations survive into the deck.
4. Time the search-to-deck path against your current process, including the manual restructuring you normally do.
5. Test one multi-step research task in agent mode to see whether the extra latency buys a better answer.
6. Decide on the differentiator, not the baseline — if the multilingual retrieval and the slide path are not better for your work, the general answer quality will not carry it.

What Does Felo's Site Leave Unanswered?

Two things are worth knowing before you build a workflow on this, and neither is stated on the marketing pages.

The first is which models sit behind which features. Felo offers an LLM playground, which implies multiple models are available, but the site does not say what powers ordinary search or Slides. That matters for anyone with a policy about where data goes, and for anyone whose answer quality expectations are model-specific.

The second is how the free tier is bounded. Free tiers on AI search products almost always carry limits on searches, model access or output volume, and the shape of those limits determines whether free is a trial or a usable state. The site presents a free entry point without our being able to establish where it stops.

There is a corporate footer with Enterprise, Blog, Tools, Privacy and Terms sections, which is more transparency than many AI startups offer and the right place to start if data handling is a constraint for you.

None of this is disqualifying. It is the ordinary due diligence for a tool that will hold your research questions, which are frequently more sensitive than the answers.

For academic work the gap is wider still. A literature tool like Elicit retrieves from indexed papers rather than the open web, which is a different and considerably more reliable corpus for that job.

• Confirm which model backs search and Slides before routing anything confidential through them.
• Find where the free tier stops before designing a workflow that assumes it does not.
• Read the privacy terms if your research questions are themselves commercially sensitive.
• Check whether Slides preserves citations, since an unsourced deck is a liability.
• Test the multilingual retrieval on your own sources rather than trusting the description.