In the past two months, we've shipped eight new major product releases. Network Graph. AI-Powered Browser ID. AI Rule Builder. Cross-Device Authentication. Verified IP Location. Location 2FA. Proof of Humanity. Web Trusted Location.
That's an unusual pace for an infrastructure company. The pace comes from something specific: a decade-plus of investment in proprietary tech and data that, until recently, was genuinely hard to act on at full resolution. AI changed what we can do with it, and the backlog of possibility that had been building up is now shipping.
This post is about why that's happening, and what these launches actually have in common.
Incognia has spent more than twelve years building a physical-world identity graph, grounded in indoor location data collected across our network of over one billion devices. Not GPS coordinates. Not ISP assignment records. Actual behavioral location signals from the real world, continuously observed and refined into a persistent understanding of where each device lives, works, and moves through the day.
Nobody else has this. That isn't a marketing statement, it's a description of the industry. Apple tried it with WiFiSlam and iBeacons. Google tried it with Project Tango and Eddystone. Neither turned indoor location into a production-grade product. We did, and we've been compounding that advantage since.
Nowadays, we're embedded in over 1 billion devices across 190 countries. Which is, by far, the largest deployment of indoor location technology. On top of that, we've built a location identity technology with an incredible resolution: a false acceptance rate of less than 1 in 17 million, this is 17x better than sophisticated biometric technologies such as FaceID.
Since 2020, this dataset powered the category we're best known for: mobile fraud prevention. Device recognition, location spoofing detection, address verification, trusted location checks. All of that remains core. But the underlying data has always contained far more signal than what we were extracting from it. The bottleneck was modeling capacity, not data.
The shift is simple to describe and significant in practice. Our location graph contains patterns that human-designed models could only partially surface: behavioral signatures across time, relationships between devices at the same physical places, subtle correlations between network identifiers and real-world presence. Modern AI architectures, especially transformer-based models, are built exactly for this kind of pattern extraction across large, complex, dynamic datasets.
What we're doing now is running the same proprietary data foundation through a dramatically more capable modeling stack. The result isn't incremental improvement. It's new product surface area.
Network Graph made our persistent device identity directly visible to fraud analysts. Because the underlying ID holds through app reinstalls, different apps, and tampering, the graph shows real actors over time instead of fragmenting into new nodes every time a fraudster evades.
A few weeks later we brought the same philosophy to the web with AI-Powered Browser ID, a transformer-based approach that tokenizes browser signals and maps them into a high-dimensional vector space. It delivers a 25% improvement in re-identification over our previous generation. The result is a re-identification model that adapts as signals change, rather than relying on static fingerprints.
Verified IP Location is the clearest expression of our thesis. IP geolocation has historically been inferred from ISP assignment records, which degrade fast and tell you almost nothing about where traffic is actually coming from. We built an IP location signal the way meteorologists build weather forecasts: from a dense network of real observations. Every device in our network contributes to an inferred location for the IPs it touches, continuously. Coverage of IPs is already above 90%, with a clear path to 99%. That makes a previously fuzzy signal actionable, and it is only possible because of the scale of the underlying device network.
Web Trusted Location converts that signal into actionable risk logic at the session level. At login, it compares the current session's inferred location against the browser's established access points, producing a distance score and an age score that together distinguish a familiar access pattern from an anomalous one. In testing, 92% of sessions with complete location data were correctly classified as low risk, and bot or cloud traffic was flagged accurately, without adding friction for real users.
Cross-device authentication then ties web sessions to the verified physical location of the approving mobile device, closing the gap that social engineering exploits in traditional step-up challenges.
AI Rule Builder is the clearest example of AI operating on our own library rather than on customer data. Risk analysts can describe a fraud scenario in plain language and get back fully-formed rule logic, drawing from our library of 500+ proprietary signals. The model was trained specifically on that library, which is why it returns precision logic instead of keyword matches.
The two most recent launches are the clearest statement of where this is going. Location 2FA turns physical presence into a second factor: an authentication check grounded in being in the right place, which is the one dimension of identity that remote attackers cannot replicate at scale.
Proof of Humanity goes further. It packages our signal stack into a continuous attestation that an account is controlled by a unique, real human moving through the real world. A device farm can fake a face. It cannot fake a morning commute across ten thousand accounts. Our False Acceptance Rate on this signal is 1 in 17 million, roughly 10x to 100x better than iris recognition benchmarks from NIST, with no hardware and no user enrollment required.
There is a question that naturally follows from what I've described so far. If the dataset is this large and the models are getting more capable, what does it cost to actually run this?
Incognia processes signals from over one billion active users on roughly $300K a month in cloud costs. That is a level of efficiency most infrastructure companies at our scale do not come close to, and it is not an accident. Our engineering team has spent years building a platform that handles this volume with unusual unit economics, because from the start we knew the category we were building in would demand it.
This is its own moat. A data advantage only compounds if you can operate against it cheaply enough to monetize it at scale. Any competitor looking to replicate what we do would not only need to build a comparable indoor locationtech and dataset, they would need to build the infrastructure to serve it at the margins that make the business work. Data is one moat. The ability to run it at this unit economics is the second.
Eight products shipped in eight weeks because every one of them is built on the same foundation. We aren't spreading investment across unrelated categories. We are extracting more product surface area from a dataset that has always had more to give, and AI has made that extraction economically and technically viable for the first time.
This matters strategically for a specific reason. The questions the industry will be asking over the next several years are going to revolve increasingly around authentication in a world where AI agents are transacting on behalf of humans. A credential can be replayed. A session can be spoofed. A behavioral signal can be modeled. What cannot be faked by an agent running in a data center is actually being in the physical location where a real person lives or works.
That is the layer we have been building toward since the company was founded. Eight launches is the rate at which it is now becoming addressable.