Consumer & Creative AI
Muse's Download Estimates Differ by 87%
Meta's Muse is growing fast, but estimates range from 2.3 million to 4.3 million downloads. Builders need conversion evidence, not install hype.
September 25 reporting on Meta’s Muse puts app downloads between 2.3 million and 4.3 million, depending on the measurement company. That 87% spread is large enough that a builder should treat the launch as a distribution opportunity to test—not an addressable-user forecast ready for a business plan.
The audience is real; its size is still an estimate
TechCrunch reports Sensor Tower’s estimate at more than 3.4 million downloads, Apptopia’s at 4.3 million, and Appfigures’ at roughly 2.3 million, as reported on Friday. Sensor Tower and Appfigures refer to Thursday’s estimates; the article does not establish an identical cutoff for Apptopia. The spread is (4.3 ÷ 2.3 − 1) × 100 = 87%, rounded. These are alternative estimates of the app’s cumulative installs, not three segments that should be added together.
Muse's download estimates differ by 87%
Download estimates reported September 25; cutoffs may differ. Not active users.
The chart’s disagreement is the story for planners. None of these estimates establishes how many people completed a useful task, came back, connected a business service, or made a purchase. It also does not establish which measurement firm is right; methodology and measurement cutoffs may differ. Choosing the largest total because it makes a connector investment easier to approve would replace uncertainty with a sales assumption.
A second calculation gives the launch some scale without pretending to measure retention. Meta’s original announcement is dated September 8. Combining that date with TechCrunch’s September 24 measurement cutoff yields 16 elapsed calendar days. More than 3.4 million Sensor Tower-estimated installs divided by 16 is more than 212,500 installs per elapsed calendar day. This is a rough calendar-normalized launch average, not an observed daily series or a precise hourly measurement window.
It is still consequential. An app can deserve developer attention before anyone knows its durable audience size. TechCrunch also reports that Sensor Tower measured a 27% increase in daily active users on Wednesday after Connect. That is evidence of a short-term change in engagement, but no absolute daily-user total is supplied there. It cannot be converted into a retention rate by dividing it into cumulative downloads.
Distribution has a visible commercial engine. The same reporting describes Meta promoting Muse across its properties and buying advertising elsewhere. That can bring users to a new service quickly. It does not tell a third-party builder how many will discover a particular integration or authorize a particular action. Developers need a funnel for their own product, not a generic multiplier borrowed from Meta’s reach.
The underlying product is more than a chat surface. Meta’s launch description says Muse can keep working after the app closes, connect to services chosen by the user, and seek approval before sensitive actions such as sending email or making purchases. Those are vendor descriptions of intended controls, not results from an independent evaluation performed for this article.
Build a connector experiment, not a platform dependency
The practical opportunity is to make one useful service reachable from an agent. TechCrunch’s Connect coverage reports more than 1,500 connector applications in less than a week, alongside announced retail and productivity integrations. An application is not an approved connector, and an approved connector is not proof of demand. The number establishes developer interest; it does not establish an equal share of Muse’s audience for each participant.
Start where the value proposition already exists. A merchant could test whether agent-originated shoppers complete the same purchase journey as other customers. A software provider could test whether a narrowly scoped integration helps existing users finish a known task. Those are proposed experiments, not claims that Meta currently exposes every attribution field a builder would want. Request the measurement and approval requirements before committing engineering time.
The cost side is not fully public. Meta says Muse is free for most needs with subscription plans for heavier use. The Connect reporting quotes an intention to earn a small fee from transactions over time, but supplies no universal merchant fee schedule. Builders should not assume that free consumer access means free distribution, fixed future economics, or zero support burden for the service on the other end.
Operational costs include maintaining the connector, handling failed actions, and reconciling what the user approved with what the destination system received. Price those activities using the business’s own records. The fetched material does not support a standard dollar saving per Muse user or a conversion-rate forecast, and the install estimates cannot supply the missing inputs.
The model announcement offers a plausible reason for better agent behavior without settling those economics. Meta says Muse Spark 1.3 improves long-horizon work and confirmation before consequential actions. Those claims are relevant to task completion. They are not evidence that a particular third-party workflow succeeds at an acceptable rate, and model-level improvements should not be substituted for integration testing.
Privacy also has a rollout boundary. Meta describes Muse Secure VM and a separate Sentinel component in the launch post, while saying a Confidential VM would arrive later in the year. Do not describe that future protection as already deployed. Review the actual access requested by a connector and the audit information available to users. A platform’s security narrative is useful context; it does not remove the developer’s responsibility to minimize its own permissions.
Our earlier analysis of Meta’s transcription price and missing word timestamps showed how one missing capability can change the economics of an apparently attractive product. The equivalent question here is whether the integration supports the specific action, review, and recovery behavior the business needs. Today’s Copilot lead similarly separates agent access from a defensible operating budget.
The verdict is to allocate a bounded experiment, not rebuild around the platform yet. Expand when attributable completed work, repeat use, and support costs justify it. Pull back if installs fail to become useful activity or the commercial terms undermine the margin. Better retention data and published transaction economics would change this assessment more than another download milestone. An 87% measurement spread is not a reason to ignore Muse; it is a reason to keep the investment reversible.