Clean network noise out of your engagement KPIs

The missing variable

Your dashboards log everything except the network

Speed is how fast a connection is on average. Jitter is how much it stumbles within the day — fine one second, frozen the next. On a high-jitter day every like, retweet and tap-through has to land between the stalls. Ad platforms log delivery, frequency and viewability; nothing in the stack logs the connection your audience was actually on. KASPR does: round-trip measurements across millions of internet addresses, summarised here by location and day.

And it moves like a real variable, not a constant. On any given day the shakiest places run several times the instability of the calmest, and the same place swings from day to day — which is what lets the analysis compare each location against its own normal, the only comparison the numbers below rely on.

On a typically shaky day, a location’s connection instability runs about 53% above that location’s own norm.

The finding

Same account, shakier connection, fewer likes

Take every tweet in the sample and ask one question: when the connection in the author’s location was shakier than its own norm, what happened to engagement? The comparison is apples to apples — same place, same date, and in the strictest cut the same account — so the answer is not driven by who posts, from where, or by whatever was happening on the platform that day.

In the strictest cut, a doubling of connection instability goes with roughly 0.6% fewer likes and 0.5% fewer retweets per post. Invisible on any single post. Real across a few hundred thousand of them.

Read the chart like an analyst. The looser cuts in the top rows are noisy — too noisy to call. Hold the account fixed and both metrics settle clearly below zero. The tighter the comparison, the clearer the pattern.

Stability, not speed

Stability, not speed, carries the association

Speed and stability are different products. A market can be slow-but-steady or fast-but-erratic, and average speed mostly reflects geography and infrastructure — it barely changes day to day. The instability is the part that changes from day to day — and day-to-day change is what shows up in metrics.

Put average speed and instability head-to-head in the same model and speed shows nothing, for either metric. Instability keeps its size: still clearly negative for retweets, while the likes reading turns too noisy to call once speed joins the model. Whatever drives the penalty, it tracks the stumbles, not the headline megabits.

Held to the same test, average speed shows nothing. Instability is the term that survives.

The robustness check

Run it any defensible way — the sign holds

One flattering regression proves nothing, so the grid of analyses was locked before any result was seen: every defensible combination of metric, transformation, control set and sample rule — 1,260 in all, 180 with instability as the focus. 144 of those ran to completion, and every one of them is on the chart. Nothing was dropped for its answer.

100% come back negative. 31% are individually strong enough to call, concentrated in the log-scale family, where the middle estimate is −0.0084. The raw-count versions are far noisier — heavy-tailed metrics always are — but the sign holds there too.

This is descriptive evidence and the page treats it that way. It says the pattern survives every analytic choice; it does not, on its own, prove the network causes the gap. The designs that could settle cause — built on outages — are separate, ongoing work.

100% of completed runs agree. No way of slicing the data flips the sign.

The creative side

Shaky days change what gets posted, too

Engagement is only the receiving end. Point the same comparison at the content itself — post length, tone, links, hashtags, word length, readability — and ask which of them move when the local connection deteriorates.

One does: tone. Posts written on shaky days read more negative, by 0.014 of a standard deviation per notch of instability — small, but measured cleanly enough to call. Word length leans down without clearing the bar, and the structural features do not move. The mood of what gets written, not just what gets engaged with, tracks the network.

No easy workaround

No format, audience or time slot offsets it

If this were a heavy-media problem or an audience quirk, you could route around it — swap the format, shift the slot, lean on bigger accounts. Seven workarounds were tested, one at a time: links, replies, follower count, mentions of either 2020 candidate, upbeat tone, and prime-time posting.

None of them changes the picture. Every interaction is too noisy to call, and the headline penalty keeps its size in every test. On this evidence the effect is broad-based: reformatting or retiming an individual post would not be expected to offset the network it lands on. What a team can do is know which markets and days carry the noise — and read the KPIs, and set the benchmarks, accordingly.

Seven workarounds tested. None of them changes the estimate.

From the study to your stack

The data behind it, and how to plug it in