A guy on my crew called me from a job to tell me the app was broken. He had just run a speed test in the driveway. 84 megabits, he said, and it still would not hold a picture.
He was reading the download number. Everybody reads the download number. It is the big one, it is the one in the middle of the screen, and it is the wrong one. For live video the only figure that matters is upload speed, and on that same driveway his upload was probably somewhere around 4 or 5 Mbps, dropping under 1 the moment he got on the roof.
Cell networks are built to send, not receive
This is not a defect. It is the design.
Your phone puts out roughly 200 milliwatts at maximum transmit. The tower sector answering it runs tens of watts through a high-gain antenna mounted a hundred feet in the air. Two devices trying to talk to each other with a hundred-fold power difference is going to have a range problem in one direction, and that direction is yours.
The spectrum allocation follows the same logic. LTE gives the uplink less room. Mid-band 5G splits time between transmit and receive and weights that split heavily toward download, because the overwhelming majority of consumer traffic is download. And most 5G deployments you will stand under are non-standalone, meaning the phone's uplink often still rides the older LTE anchor carrier. That is why upload speed for live video can be flat unimpressive while the screen proudly shows 5G.
Practical translation: whatever your download number is, assume your upload is a fifth of it, sometimes a tenth, and assume it degrades faster than download as conditions get worse. Because the link that fails first is the weak transmitter, and the weak transmitter is in your hand.
What the stream actually needs
Rough numbers I trust from doing this, not from a spec sheet.
A 720p stream at 30 frames per second, good enough to actually look at a roof surface, wants somewhere around 1.5 to 3 Mbps. Go to 1080p and you are looking at 4 to 6. Drop to 480p and you can survive on 600 kbps to 1 Mbps, but you have given up most of the detail you climbed up there for.
Then add headroom. Cell capacity is shared and it moves under you while you work. I want roughly double my target bitrate available, so a comfortable 720p walk means seeing 3 to 4 Mbps up and seeing it consistently. Under about 800 kbps sustained I do not go live at all.
Consistency is doing more work in that sentence than the number is. A link that averages 3 Mbps by alternating between 6 and near zero is worse for live video than a flat 1.2. Encoders and congestion controllers can adapt to a low ceiling. They handle a ceiling that keeps collapsing much worse, and every collapse is either a freeze or a chunk of the wall you were showing somebody.
Why the driveway speed test misleads
Five separate reasons, and they compound.
It measures download by default, and the download result is what your eye lands on.
It measures a burst. A speed test takes everything the network will give it for a few seconds against a nearby well-provisioned server. You need twenty minutes of steady. Peak throughput and sustained throughput are different properties of the same link and only one of them is on your screen.
It hides packet loss and jitter, which is what actually wrecks a live stream. A link at 3 Mbps with 8 percent loss will test well and stream badly. A clean 1.2 Mbps link will test poorly and stream fine. If your test app reports loss or latency variance, those two numbers predict your experience better than the throughput figure does.
It happened at ground level, on the street side of the house, with a different set of obstructions and possibly a different sector than the one you will attach to at ridge height.
And it happened at a moment. Towers are shared. Marginal addresses can be usable at 8 a.m. and hopeless at 4 p.m. with nothing changed but the neighbors.
Reading the link like a tradesman instead of a number
Once you stop trusting the big number, a few habits replace it.
Test upload only, and test it twice with fifteen seconds between. Two similar results mean a stable cell. Wildly different results mean congestion, and the low one is the honest one.
Watch what the picture does rather than what the app claims. Degradation follows an order: bitrate first, then frame rate, then resolution. If detail is going soft while motion stays smooth, you are at the top of that ladder and you have room. If motion is stuttering, you are further down. If you are watching a 360p roof, you were out of upload two steps ago.
Move slower when upload is thin. This is the most useful thing on this page. Encoders spend bits on what changes between frames. Pan fast on a thin link and the entire frame is new information every frame, so everything degrades at once. Hold on the spot you want seen for a slow four count and watch it sharpen. On a weak link, holding still is not politeness, it is the only way to get a usable image.
Turn Wi-Fi off before you climb, because a phone clinging to the homeowner's access point at the ragged edge of range will hand you an upload figure far worse than the cellular you were about to get.
The number that should be on your dashboard
If your streaming tool shows you anything while you work, the thing you want on screen is current upload bitrate and a loss indicator, not bars. Bars describe how loud the tower is. They say nothing about whether you can push anything back. I have had four bars and no usable uplink more times than I can count, always in the same kind of place: high, open, and in view of several towers at once, none of them clearly winning.
The broader picture of how a link degrades and what to do about it is in the main piece on streaming at one bar. This post is the narrow version: upload is the constraint, upload is what you test, and upload speed for live video is a fraction of the number the speed test hands you.
Two minutes of testing the right direction saves a wasted appointment. Try a session from the worst corner of your service area and watch what upload does when you climb. You will start reading a link by feel after a handful of runs, and seeing the flow once is faster than reading about it.
InspectStream records what an inspector observed. It is not a public adjusting service and does not prepare, negotiate, or advise on insurance claims.