Understanding the Internet’s structure, performance, and reliability is becoming increasingly difficult. At the same time, the Internet is playing an ever more important role in society, even as our ability to observe it directly continues to decline. Existing measurement techniques offer only partial visibility into how it behaves: path measurements show only the routes we happen to probe, routing data captures only the routes networks choose to announce, and end-user measurements are often sparse and shaped by where users happen to run tests. Each of these signals is incomplete on its own, but together they provide the main evidence we have for understanding Internet behavior. Still, this evidence alone is often not enough to answer the questions we actually care about: where connectivity is fragile, why performance degrades, which users are affected, and whether the Internet is delivering the properties we expect of it.

 

Answering these questions requires combining partial observations and reasoning beyond what we can directly see. Inference is therefore becoming an essential part of Internet measurement. But for these inferences to support meaningful auditing in practice, especially for network operators and policymakers, they must be principled: assumptions should be explicit, the accuracy and coverage of the inferred view should be quantified, outputs should be interpretable, and the remaining uncertainty should help guide where additional measurements are needed. To address this need, I present models and systems that turn sparse, heterogeneous measurements into more complete views of Internet structure, performance degradations, and their impact on users. More broadly, I argue for a shift in perspective: inference should not be treated as a disconnected supplement to measurement, but as an integral part of the measurement process itself.

 

I conclude with two future directions that I find particularly exciting: first, using causal inference to move Internet measurement beyond diagnosis and toward intervention; and second, rigorously studying how LLMs are changing how information is accessed online, with downstream consequences for traffic, incentives, and the still-open question of how the modern Internet will sustain itself financially.