Otter vs Fireflies vs Granola vs Scribbl: We Tested All Four on 5 Meetings (2026)

Getting notes out of a meeting isn't hard anymore. Most AI note takers record the call, write down who said what, and give you a summary when it ends. So the differences between them show up in everything around the notes.

We compared four popular ones, Fireflies, Granola, Otter and Scribbl, on seven things people tend to ask about:

1. Does a bot join your call?
2. Do people know they're being recorded?
3. Does it train AI on your meetings?
4. Who sees your notes?
5. Do you keep the video?
6. How good are the notes?
7. What happens after the call?

We picked these because they're the things that change your day to day. You might weigh them differently. And it doesn't cover everything: price, languages, security certifications and support are worth comparing too, and we didn't test those.

The notes we tested ourselves, on five recorded meetings. For the rest, we went through each company's own help pages, policies and pricing, as of October 2026. Plans and features change, so check those before you choose. For each one, we show how we checked it and what we found.

The results at a glance

	Fireflies	Granola	Otter	Scribbl
Bot joins the call	By default; desktop app records without one	Never	By default; desktop app records without one	Never
Trains AI on your meetings	No	Yes by default (anonymized), can turn off	Yes (de-identified)	Not for anyone other than you
Notes shared by default	Recap emailed to everyone on the invite	Private until you share	Depends on auto-share settings	Private until you share
Keeps the video	Paid plans, bot only	No audio or video	Enterprise plan	Every plan
Coverage of the researchers' notes (5 meetings)	69%	77%	77%	78%
Mistakes (5 meetings)	about 28	about 6	about 18	about 6
Ranked first, blind side by side (of 80)	0	32	0	48

Does a bot join your call?

The first thing we looked at is whether each tool sends a bot to your call.

For a long time, that's how most AI note takers worked. You'd invite a notetaker to the meeting, it would join like another participant, and it would record from inside the call. More recently, a lot of tools have added ways to record from your own computer, with nothing joining the call. Some platforms have also started limiting bots. Microsoft Teams, for example, now holds outside note-taker bots in the lobby until the organizer lets them in.

To check this, we went through each company's help pages.

Fireflies joins with a bot by default. Its desktop app, and its Chrome extension for Google Meet, record without one, but don't keep the video.

Granola doesn't use a bot. It's a desktop app that listens to your computer's audio, so it works with whatever call app you use. You start it by clicking into a note, and it works whether or not you're the host.

Otter joins with a bot by default. Its desktop app records without one.

Scribbl doesn't use a bot. It works on Google Meet through a Chrome extension, and on Zoom and Teams through a desktop app for Mac and Windows. It starts when you join, and it works whether or not you're the host.

So Granola and Scribbl never use a bot, and Fireflies and Otter offer a way to record without one.

Do people know they're being recorded?

Next, we looked at consent: do the other people in the meeting know they're being recorded?

Recording laws are different depending on where people are, and in some places everyone on the call has to agree, not just the person recording. Note takers are also in court over this. Otter is being sued in a class action, In re Otter.ai Privacy Litigation, that says its notetaker recorded people without the consent the law requires. On August 13, 2026, a federal judge let the wiretap and privacy claims go forward. The case hasn't been decided.

To check this, we looked at each company's help pages and at the reporting on the Otter case.

Each of these tools has some way to signal that a meeting is being recorded, but how it works differs by tool, by meeting platform and by your settings. None of them can make sure everyone actually heard it.

So none of the four stands out here. With any of them, the most reliable thing you can do is say it out loud at the start: "Mind if I record this so I have notes?"

Does it train AI on your meetings?

Next, we looked at whether each company uses your meetings to train its AI.

Meetings often include pricing, customer names or hiring decisions. Some people don't mind de-identified data being used to improve a product. Others want a clear no.

To check this, we read each company's privacy policy and help pages.

Fireflies says your data is never used for AI training.

Granola uses anonymized data for training by default, and you can turn that off in its settings.

Otter's privacy policy says it trains its AI on de-identified audio recordings and transcripts.

Scribbl says it doesn't use your meetings to train AI for anyone other than you.

So Fireflies and Scribbl don't train shared AI on your meetings, Granola does unless you turn it off, and Otter does with de-identified data.

Who sees your notes?

Next, we looked at who can see your notes after the meeting, by default.

Sharing notes automatically can be useful. The whole team stays on the same page without anyone forwarding anything. It can also send notes from a sensitive call, like a pricing negotiation or an interview, to people you didn't intend.

To check this, we looked at each tool's defaults in its help pages.

Fireflies emails the recap to everyone on the calendar invite by default, and anyone with the link can open it. You can change that in its settings.

Granola keeps your notes private until you share them.

Otter has auto-share settings that can send notes to other people, depending on how they're set.

Scribbl keeps your notes private until you share them.

So by default, Granola and Scribbl keep notes private, Fireflies shares the recap with everyone on the invite, and Otter depends on its auto-share settings.

Do you keep the video?

Next, we looked at what each tool keeps from the meeting, and whether you get the video.

Whether this matters depends on you. AI notes are a summary, so a recording lets you go back and check what was actually said. Some people would rather not keep recordings at all.

To check this, we went through each company's pricing and help pages.

Fireflies keeps video on its paid plans, when its bot joins the call.

Granola keeps the transcript, but not the audio or video.

Otter keeps video on its Enterprise plan.

Scribbl keeps the video on all of its plans, on Google Meet, Zoom and Teams, and its notes are timestamped, so you can click a point and jump to that moment.

So if you want the video, Scribbl keeps it on every plan, and Fireflies and Otter on some. If you'd rather not keep recordings, Granola is built that way.

How good are the notes?

Next, the notes themselves. This is the part we tested directly.

We put the same five meetings through all four tools, about two hours of audio in total. They're recorded meetings from the AMI Meeting Corpus, a public research library at the University of Edinburgh, between 14 and 37 minutes long, so anyone can check the results. We wrote down how we'd score them before running four of the five meetings.

5

public meetings, 14 to 37 minutes each

103

items in the researchers' own notes

80

blind side-by-side AI judgments

Check 1: graded against the researchers' own notes

The researchers who recorded each meeting also wrote their own summary of it: the decisions, action items and problems. We used every sentence of those summaries as an answer key, 103 items across the five meetings. Three AI models, Claude Opus 5.5, Claude Fable 5.1 and GPT-6.1 Sol, checked every item against the transcript, each one seeing only one tool's notes, with no tool names. An item the notes got right scored 1 point, and one they got partly right scored half. The graders also listed every mistake they found: wrong facts, the wrong person, and anything the notes made up.

Meeting	Length	Fireflies	Granola	Otter	Scribbl
IS1008d (product design, final meeting)	24.5 min	60%	61%	78%	81%
IB4003 (planning an office move)	34 min	86%	80%	60%	67%
ES2008b (product design)	37 min	56%	80%	67%	80%
IS1001c (product design)	24 min	57%	77%	91%	79%
TS3012a (product design, kickoff)	14 min	86%	87%	87%	83%
Average		69%	77%	77%	78%

Scribbl covered 78% of the researchers' points on average, Granola and Otter 77%, and Fireflies 69%. That's close to a three-way tie, and each of the four had at least one meeting where it scored highest.

The mistakes separated them more. Across the five meetings, Granola's notes had about 6 mistakes, Scribbl's about 6, Otter's about 18, and Fireflies' about 28.

Kind of mistake (5 meetings, averaged across graders)	Fireflies	Granola	Otter	Scribbl
Wrong fact	7.0	2.7	2.7	3.7
Wrong person	4.0	0.3	3.7	1.0
Action item nobody agreed to	9.3	0.3	0.7	0.7
Other content that didn't happen	4.0	0.7	5.7	0.0
Decision described differently from how it happened	3.7	2.0	5.0	1.0
Total	28.0	6.0	17.7	6.3

Check 2: compared side by side, blind

Then, as a second check, we had four AI models compare the notes side by side, blind: Claude Opus 5.5, Claude Fable 5.1, Claude Sonnet 5.5 and GPT-6.1 Sol. Each one got the transcript and four sets of notes labeled A to D, without knowing which tool wrote which, and scored them on accuracy, completeness, decisions and action items, and usefulness, 25% each. Each model judged each meeting four times with the order shuffled, so every tool sat in every slot once. That's 80 AI judgments in all. It's not a human panel.

Meeting	Fireflies	Granola	Otter	Scribbl
IS1008d	0 first, 5.3	3 first, 7.5	0 first, 5.2	13 first, 7.9
IB4003	0 first, 6.4	11 first, 8.0	0 first, 4.7	5 first, 7.7
ES2008b	0 first, 5.0	2 first, 7.7	0 first, 5.4	14 first, 8.3
IS1001c	0 first, 5.4	16 first, 8.7	0 first, 6.2	0 first, 7.7
TS3012a	0 first, 5.5	0 first, 7.2	0 first, 5.7	16 first, 8.1
All 80	0 first, 5.5	32 first, 7.8	0 first, 5.4	48 first, 7.9

Each cell is how many judgments ranked that tool first, then its average overall score out of 10.

Scribbl's notes ranked first in 48 of the 80, and Granola's in the other 32. Their average scores were close, 7.9 and 7.8 out of 10. Scribbl came out ahead on three of the meetings and Granola on two. Fireflies and Otter didn't rank first in any.

What the grading and the judges found for each one

Fireflies' notes were long and covered the most on one of the meetings. They also had the most mistakes, and about a third of them were action items nobody agreed to. In one meeting it assigned a task to a "Christian." There's no Christian in the meeting.

Granola's notes had the fewest mistakes, with Scribbl's about the same, and were the cleanest to read. They usually don't say who said what, and the next steps often don't have owners.

Otter scored highest on one meeting. Its mistakes were mostly details the meeting didn't contain and decisions described differently from how the meeting ended.

Scribbl's notes said who argued for what and who owns each next step. It listed fewer action items than Fireflies, and the graders found almost nothing in its notes that the meeting didn't contain.

So in this test, Scribbl and Granola came out on top for the notes, with Scribbl ranking first more often side by side.

Limits of this test

• Five meetings, all in English, all from one research corpus. Four are product design meetings where people play assigned roles.
• All grades come from AI models, not people.
• All four tools ran on their free plans. Paid plans may write different notes.
• Granola has no file upload, so the audio was played through a Mac while Granola recorded.
• Every change from the written plan is listed in the protocol in the test data.

What happens after the call?

The last thing we looked at is what each tool does after the call.

A summary used to be the end of the job. Now a lot of tools can also take the next step, like drafting the follow-up email, updating your CRM or creating tasks. Whether you want that depends on how you work.

To check this, we went through each company's pricing pages and integration docs, and looked at three things: taking the next step, sorting your meetings, and asking AI across them.

Fireflies connects to CRMs, task tools and chat apps. It can sort meetings with rules for Enterprise admins, for meetings going forward. And it connects to Claude and ChatGPT.

Granola offers follow-up emails and CRM updates on its Business plan, and tasks through Zapier. You file meetings into folders yourself. And it connects to Claude and ChatGPT.

Otter puts CRM notes on its Pro plan and follow-up emails on Enterprise. It sorts meetings by its own meeting types. And it connects to Claude and ChatGPT.

Scribbl drafts follow-up emails and sends notes to CRMs, task tools and Slack, with monthly limits on its free and Pro plans. It sorts your past and future meetings into groups you describe in a sentence. And on Pro, it connects to Claude and ChatGPT, where you can ask about one of those groups at once.

So across the three things we checked, Fireflies and Scribbl covered the most. Scribbl was the only one that automatically sorts past meetings into groups you describe.

Where they landed

Granola and Scribbl never use a bot and keep notes private by default. Fireflies and Otter offer ways to record without a bot. Fireflies and Scribbl don't train shared AI on your meetings. On consent, none of them stands out.

Scribbl keeps video on every plan. In our five-meeting notes test, it ranked first most often, with Granola close behind. Along with Fireflies, it covered the most after the call.

Granola tied Scribbl for the fewest mistakes in the notes test. It's built for people who'd rather not keep recordings. Fireflies offers a lot of integrations. Otter's desktop app records without a bot, and it matched Scribbl and Granola on how much of each meeting it covered.

Which one fits depends on what you care about most, and on the things we didn't test. For a wider list of tools, see our guide to the best AI note takers.

How we tested

• Meetings: five meetings from the AMI Meeting Corpus (University of Edinburgh and partners, CC BY 4.0): IS1008d, IB4003, ES2008b, IS1001c and TS3012a, 14 to 37 minutes each. The corpus was picked because the researchers wrote their own notes on each meeting, which gave us an answer key we didn't write.
• Plans: all four tools ran on their free plans.
• Input: the same audio file per meeting for every tool. Fireflies and Otter got it through their upload screens. Granola has no upload, so the file was played through the Mac while Granola recorded. Scribbl got it through the same steps a desktop Zoom recording takes.
• Check 1: each tool's notes graded against the researchers' notes and the transcript by Claude Opus 5.5, Claude Fable 5.1 and GPT-6.1 Sol, one set of notes at a time, with no tool names.
• Check 2: all four sets of notes judged side by side, labeled A to D, by Claude Opus 5.5, Claude Fable 5.1, Claude Sonnet 5.5 and GPT-6.1 Sol, four times per meeting with the order rotated.
• Everything else: each company's own help pages, policies and pricing, read in October 2026.
• Data: the audio, each tool's notes, the answer keys, every prompt, every grader's answer and the scripts that recompute every number are on GitHub.