Home › Topics › Digital Experience Monitoring: From Device Health to User Happiness › The 'It Feels Slow' Gap

Closing the “It Feels Slow” Gap: From Feelings to Measurements


There are two stories about your network, and they do not agree. In the first, told by the people who use it, "it has been slow all week," the shared drive "hangs," and the video calls "keep freezing." In the second, told by your monitoring system, every link is under half its capacity, every server is bored, and the response-time graphs are flat. I have sat in the meeting where these two stories collide. And I have said the sentence that ends it badly: "everything looks fine on our side."


Both stories are true. That is the whole difficulty. The user really did have a bad week, and the graphs really are flat. The two are measuring different things from different places at different moments. The gap between them is not a credibility contest to be won. It is a measurement gap to be closed. Closing it is a repeatable process rather than an act of diplomacy. A ticket that says "it feels slow" is not a complaint about your competence. It is the only sensor in the building that measures the right thing, reporting in the only units it has.


By the end you will be able to capture a complaint as data instead of a mood. You will be able to line it up against endpoint and infrastructure timelines and read the four-square matrix of reported versus measured. You will also be able to build a feedback loop that tells you whether a fix actually landed. And you will be able to report improvements in words users recognize as their own.


In this guide

  • Two true stories
  • Capturing the complaint: time, place, app, device
  • Matching the complaint to the data
  • The reported-versus-measured matrix
  • A worked example: one complaint, end to end
  • Building feedback loops
  • Reporting improvements in user terms

Two true stories


Your graphs are honest about what they measure. A utilization graph shows the five-minute average load on a link, from the perspective of the switch. A server response-time check shows how quickly the server answered a synthetic request sent from the monitoring machine in the server room. Neither of those is a lie. Neither of those is the user's experience, either. The user experienced a specific journey — open the file, wait, wait, give up. It happened on a specific laptop, on specific Wi-Fi, at a specific moment. The user judged it against what they expected. Three things separate that from the graph.


The first is vantage point. The graph was drawn from the server room. The user was at the far end of the second floor or at a kitchen table. The device and the last few metres of the connection do not appear on any graph you have. That is where a great deal of slowness lives. The second is time. A five-minute average smooths a thirty-second stall into nothing, and the stall is precisely what the user remembers. The third is expectation. A two-second pause is fine for a report and unbearable for a keystroke, and the graph does not know which one was happening.


There is also a human cost to getting this wrong. When IT says "everything looks fine," the user hears "I do not believe you." The next time something is slow they will not bother reporting it. Silence is not success; it is the sound of people giving up on you. The process below is designed so that you never have to say that sentence again. You will have something better to say: "let me line your Tuesday up against the data."


Capturing the complaint: time, place, app, device


A complaint becomes data the moment it has four fields. Time: when it happened, as precisely as the person can manage, and when it started if it is ongoing. Place: which building, floor, room, or "home." App: which application or journey — not "the internet," but "opening the CRM search page" or "saving to the shared drive." Device: which machine, by name if possible, so its endpoint data can be found. Two optional fields make everything easier. They are what it felt like ("twenty seconds instead of two," "the call froze twice") and who else ("the person next to me was fine"). The intake conversation itself is about what to ask, in what order, and how to avoid leading the witness. It is the subject of the slow-complaint triage guide. This guide picks up once the fields are filled in.


The most practical way to make this happen is to put the fields into the ticket template. That way, the help desk asks the same questions every time without having to remember them:


SLOWNESS REPORT
When did it happen?         Tue 10:05 - 10:20 (approx.), still happening? no
Where were you?             Building A, 2nd floor, north end desks
What were you doing?        Opening ledger pages in the accounting app
What did it feel like?      ~20 s per page instead of ~2 s
Which device?               LT-DANA-04 (laptop), on Wi-Fi
Anyone nearby unaffected?   Yes - colleague under the AP was fine
Anything else changed?      Unplugged from dock to walk to a meeting at 10:00

Notice that nothing on that form asks the user to diagnose anything. They are not asked whether it was "the network" or "the server." They are asked what they saw, where, and when, which are the only facts they actually possess. A form that asks "is it the network?" gets an opinion; this form gets evidence.


Matching the complaint to the data


With the four fields in hand, you can do the thing your green graphs could never do on their own. You can look at the right place at the right moment. The procedure is the same every time.


  • Step 1: Set the window. Take the reported time and widen it generously — fifteen minutes on either side at least. That is because "around ten" means anything from 9:45 to 10:20. Confirm the clocks agree. Your endpoint scripts, your monitoring server, and your ticket system must all be synchronized to a time source. Otherwise, every correlation you do is off by an unknown amount. Fixing that is a one-time job and worth doing before anything else.
  • Step 2: Pull the device's rows. If the device reports endpoint metrics, fetch its summary rows for the window: CPU, memory, disk, Wi-Fi signal, gateway round-trip, DNS time, and application response. This is the layer your dashboard is blind to, so it goes first.
  • Step 3: Pull the local infrastructure. Find which access point or switch port the device was on. Look at its client count, retries, errors, and utilization for the same window.
  • Step 4: Pull the path and the application. Pull DNS server response times, the VPN gateway's session count and load, and the internet link. Pull any synthetic check that exercises the same application, ideally from the same site.
  • Step 5: Look for what changed, not what is bad. A number can be unusual without being wrong, and normal for that site without being good. What you are looking for is the thing that was different during the window compared with the hour before and the same time yesterday.
  • Step 6: Write down the result on the ticket, in one line, even when the answer is "nothing found." A trail of "nothing found at 10:05 on Tuesday" entries is itself evidence of a measurement gap, which is the next section's subject.

LayerWhere to look for the windowSigns to look for
DeviceEndpoint summary rows for that machineDisk busy, memory low, a scan or update running, on battery
Local networkEndpoint Wi-Fi fields; the AP or switch port the device usedWeak signal, high retries, client count spike, port errors
PathDNS server timings, VPN gateway load, internet link utilization and lossSlow lookups, a saturated link, a tunnel reconnect
ApplicationSynthetic check from the same site; server CPU, memory, and disk; the provider's status pageResponse time step-change, errors, a maintenance notice
HumanThe ticket's "what were you doing" and "anything else changed"A new task, a deadline, a comparison with a colleague's newer machine

The reported-versus-measured matrix


Every complaint you line up against the data lands in one of four squares. Each square has its own meaning and its own next step. Learning to name the square out loud is the single most useful habit in this guide.


Measured badMeasured good
Reported badClear. The data and the person agree. Fix the cause, then confirm with the person.The interesting square. Look at the device, the application, and the expectation; suspect a measurement gap.
Reported goodFix before they notice. Or ask whether this metric matters to this persona at all.Fine. Occasionally check that "reported good" is not just silence.

Measured bad, reported bad: clear

The easy one, and the reason you set up monitoring in the first place. The complaint arrives, the window shows a saturated link or a struggling server, you fix it. The only discipline required is the last step: go back to the person and confirm the fix landed. Skipping that is how "fixed" tickets turn into "it is still slow" tickets a week later.


Measured good, reported bad: the interesting square

This is where the two stories collide, and it has three honest explanations. The first is that the problem lives in a layer you are not measuring. That could be the device (a scan, a full disk, a battery-saving mode) or the local network as the device sees it. The second is that the application itself was slow in a way your synthetic check did not exercise. The check loads the login page, but the user was running a report. The third is expectation: the system performed as it always does. And the person's context changed (a deadline, a new task, a colleague's faster machine). None of these is "the user is wrong." All three are things you can investigate. The first two are measurement gaps to fix by adding a metric or moving a check closer to the user.


At Tollgate Foods a ticket reading "the intranet feels slow" was closed three times in two months. The reply was "everything looks fine on our side", which was true from the server room each time. The fourth time, a new technician ran a timed page load from the user's own desk and split it into phases. Every page spent close to two seconds in name lookup before anything else happened. The laptop's adapter listed a decommissioned DNS server first. The adapter waited for that server to time out before asking the live one. The intranet server had been answering in forty milliseconds throughout, which is what the dashboard had been saying all along. The fix was one line in the DNS list the laptops are handed at boot. The user had been right for two months, in the only units she had.


Measured bad, reported good: fix before they notice

Your data says the north-end access point has been overloaded every afternoon, and nobody has said a word. Sometimes this is the best square to be in. You get to fix a problem before it becomes a complaint, and to tell people you did. Sometimes it is a message about your thresholds instead. If the metric is "bad" and nobody feels it, the band may be too tight for this persona, or the metric may not matter to them. Decide which, because a metric that is red every afternoon and felt by nobody teaches everyone to ignore red.


Measured good, reported good: fine, probably

Enjoy it, with one caution. "Reported good" often means "nobody reported anything," which is not the same thing. Once in a while, ask a few people directly, especially in the corners of the building and the remote workers, whose silence is least informative.


A worked example: one complaint, end to end


Marcus works in a branch office connected to the head office, where the file server lives, over a site-to-site link. His ticket, captured with the template above, reads: "the shared drive is slow every afternoon, roughly 2 to 4. Opening or saving anything takes half a minute. My desktop is wired. Everyone in the office says the same." That last field alone is valuable: it is the whole site, not one device.


Set the window. The complaint is recurring, so instead of one window you look at the same two hours across several days.


Device. Marcus's endpoint rows show a quiet machine: low disk activity, plenty of memory, gateway round-trip under a millisecond on the wired connection. The device is cleared in a minute.


Local network. The branch switch shows nothing unusual on Marcus's port or the uplink. Cleared. Two layers down, and the head-office dashboard is still green, which is starting to feel personal.


Path. The branch's link to head office tells the story. From about 1:50 each afternoon, utilization climbs to the ceiling and stays there until nearly 4:00, with packet loss appearing on top. The top talker across the link during that time is a backup job on a branch server. That job was recently repointed at storage in head office. Marcus's file requests are queuing behind a backup.


Application. The file server itself is idle, and a synthetic file-open check from head office is fast all afternoon. This is why the head-office dashboard was green: from where it sits, nothing was wrong.


The square. Measured bad, reported bad, once you looked from the branch's side: clear. The fix is to move the backup window to the evening or throttle it, both of which are five-minute changes. The part that closes the gap is what happens next. You tell the branch what you found and what you changed. A week later you ask Marcus whether afternoons are better. His answer goes on the ticket next to the link graph, so that the fix has both a measurement and a feeling attached to it. Most closed tickets have neither.


Building feedback loops


A feedback loop is just a habit of asking people whether things got better, and recording the answer where you can see it later. It is the cheapest piece of experience monitoring there is and the one most often skipped. After a fix the natural instinct is to move on to the next fire.


The loop has three parts. Tell people what changed, in one plain sentence, when you make a change that should affect them. For example: "we moved the backup that was clogging the branch link to the evening." Ask after a short delay — a week is about right. Use one question and a scale, not an essay: "Are afternoons on the shared drive better, worse, or the same?" Record the answer next to the measurement. That way, a graph showing the link at half utilization sits beside a human saying "yes, much better," or "no, still slow." The latter sends you straight back to the matrix.


The loop works in the other direction too. When you fix something from the "measured bad, reported good" square, tell people anyway. For example: "the north-end access point was overloaded in the afternoons. We added a second one on Monday." Some of them will reply that yes, actually, calls had been dropping and they had assumed it was normal. That is a silent sufferer becoming a data point. It is also a person deciding that reporting things to IT is worth the effort.


Tip: keep the asking light. One question, a week after a change that affects that person, and never a recurring survey. People will answer a colleague who asks about a specific fix; they will ignore a form that arrives every Friday.


Over a few months, the answers become a second experience dataset, one made of feelings rather than milliseconds. You can lay it next to your scores. When the two agree, trust the score more. When they disagree, you have found the next measurement gap. Feelings, it turns out, are a perfectly respectable unit.


Reporting improvements in user terms


The final piece of the gap is language. When you report what you fixed, the temptation is to report the measurement, because the measurement is what you worked on. But "90th-percentile DNS lookup time fell from 800 ms to 40 ms" means nothing to the people who felt the problem. They did not experience a percentile. They experienced every website taking almost a second to begin loading, and then not.


Translate every improvement into the journey it changed, for the people it changed it for. The rule of thumb is that a user should be able to read the sentence and think "yes, that is the thing I noticed." That means naming the place, the action, and the before-and-after in human units: seconds, dropped calls, retries. The engineering version still goes in the ticket and the monthly report for IT; the user version goes in the announcement.


What you measuredWhat you tell users
P90 DNS lookup time from 800 ms to 40 ms at the branchEvery website in the branch office now starts loading almost a second sooner.
Second AP installed; north-end retries down, signal strongThe four desks at the north end should no longer drop off calls in the afternoon.
Backup window moved; branch link utilization off the ceiling from 2–4 PMOpening and saving on the shared drive in the afternoon is back to normal speed.
Endpoint disk active time down after scan rescheduledYour laptop will stop grinding to a halt at ten every morning; the security scan now runs at lunch.
Site 10th-percentile experience score from 48 to 72The people who were having the worst days in the sales office are now having ordinary ones.

Reporting this way has a side effect worth wanting. It makes the next complaint easier to capture. People learn that IT talks about their journeys, not about links and percentiles. And they start describing problems the same way. No ticket has ever been opened about a percentile.


Where to go next


The gap between "it feels slow" and "everything looks fine" closes the same way every time. Capture the complaint with time, place, app, and device. Widen the window and walk the layers. Name the square. Fix. And ask. Do that consistently and the two stories start to converge. With that consistent process, your measurements grow toward where the users sit. And the users learn to give you facts. You will still be asked, in a meeting, whether everything looks fine on your side. The difference is that you will have a fifteen-minute window to point at instead of a sentence. The tools for the measuring side are endpoint metrics, experience scores, and a budget-sized rollout plan. They are the rest of Digital Experience Monitoring: From Device Health to User Happiness. For the deeper question of whether a slow journey is the application or the network, the app-versus-network slowness series picks up where this one stops.


Put this into practice with SysPeep Monitor

Lining a complaint up against the data only works if the data is there for the window you need. That means the branch link at 2 PM, the access point's client count on Tuesday, the DNS server's response time last week. SysPeep Monitor keeps that history. It discovers your devices and collects their health over SNMP and WMI. It gives you the graphs and reports to check any fifteen-minute window from one web console. The free trial installs on any Windows machine.

Download Free TrialLearn About SysPeep