Algeria AgencyDigital marketing
The day the measurement broke without saying so
Broken measurement does not look like an outage: it looks like a bad month. How to tell the difference, in ten minutes.
There is a sentence heard in every business that measures anything: “the figures dropped a lot in March”. In a notable share of cases activity did not drop at all — the measurement stopped, and nobody was told.
It is the most expensive defect in this whole field, because it is indistinguishable from the problem it imitates. A server outage announces itself; broken measurement carries on producing reports, with filled columns, plausible variations and totals nobody can dispute.
This article is about the plumbing: the six ways measurement breaks, the traffic you generate yourselves, the payment page that erases where the customer came from, tagged links, bots, and the ten-minute check that tells a drop from an outage. The companion article covers the report itself — the five numbers, their definitions, attribution, the monthly meeting — and none of that is repeated here.
It contains no chart. Every quantity it discusses is a property of your installation rather than of the market, and the companion article already devotes a full section to that impossibility.
Broken measurement looks like a bad season
When tracking stops, reports signal nothing: they simply show less. Fewer visitors, fewer conversions, less of everything, with the same formatting and the same apparent confidence as last month.
The natural reaction is to look for a commercial explanation, and there is always one available: the season, competition, a competitor who cut prices, the economy. Those explanations are plausible, impossible to disprove quickly, and they lead to decisions — cutting a budget, stopping a campaign — taken on a measurement that no longer exists.
Two signs distinguish an outage from a drop, and they are reliable. A commercial drop is gradual and uneven: it hits some channels more than others, some days more than others. A measurement outage is sharp, dated, and applies uniformly to everything the tool was counting.
The second sign is more decisive still and nobody looks at it: comparison with a figure from a different source. The number of orders in your management software, the calls received, the money taken. If one of those has not moved while the measurement collapsed, the question is settled in thirty seconds.
The six causes, by frequency
The first, by a long way, is the site rebuild. The tracking code is reinstalled by somebody who did not know there was one, or is not reinstalled at all, or is placed on the new page template but not on the checkout one. The rebuild is also the moment nobody thinks to check, because everybody is looking at the new design.
The second is a change of consent banner or setting. A tool legitimately blocked pending agreement does not measure; if the mechanism is misconfigured, it never measures, and the screen displays the consent message perfectly while nothing is being counted behind it.
The third is a change made in the tag manager by somebody else, often to add another tool. The fourth is a change of payment or messaging platform, which breaks conversion tracking without touching traffic tracking — hence reports where visits are normal and sales have vanished.
The fifth is an access expiring: an account tied to the person who left, a property nobody administers any more. The sixth is a filter added one day to exclude something and never removed. None of those six produces an alert, and that is the characteristic they share.
The traffic you generate yourselves
Part of what your tool counts is not customers: it is you, your team, your provider, and whoever checks the site three times a day. On a low-traffic site that share is far from negligible and distorts every proportion.
The most misleading effect is not volume but behaviour. You visit pages customers do not, you linger on pages being corrected, and you never convert. The result is a depressed conversion rate, popular pages that are not, and reading times that mean nothing.
Exclusion is done by network address when the business has a fixed connection, which is the simplest case. It is trickier when the team works from mobile connections, which is common here: the address changes, and an address filter does not hold.
The solution that works then is a marker set once on the team’s devices, through a dedicated page each person visits once. It takes a quarter of an hour, it survives connection changes, and it is nearly always the first correction that makes a small site’s figures usable.
The payment page that erases the origin
One particular defect deserves a whole section because it hits exactly what matters most. When a customer leaves to pay on an external service then returns to your confirmation page, they often return as a fresh visitor, coming from that payment service.
The consequence is precise and destructive: the sale is counted, but it is attributed to the payment provider rather than to the channel that actually brought the customer. Your reports then show a major conversion source that is in reality your own payment page, while advertising, search and social appear to return nothing.
That is what produces a whole category of absurd decisions: stopping the channel that works because it shows no sales, when it is bringing them and the credit goes elsewhere. The mechanism is known, it has a technical name, and it is almost never checked in the installations we take over.
The fix is to declare payment domains as not being sources — an exclusion list provided for exactly this in every serious tool. It takes five minutes and is easy to verify: place a test order and look at which source it is attributed to. That test is the highest-return one on this page.
Tagged links, and the discipline they require
To know where a visitor came from, you add tags to the links you publish yourself: in a campaign, in a post, in a message to a list. It is simple, free, and it nearly always fails for the same reason.
The cause of failure is not technical, it is spelling. “facebook”, “Facebook” and “fb” create three different sources in the reports; so do “promo-ramadan” and “promo_ramadan”. After six months, with ten people each improvising, reports contain forty sources half of which mean the same thing.
The fix is a written convention of three lines, decided once: all lowercase, one separator, and a short list of permitted source values. That document does not need to be elegant; it needs to exist where the links are made.
You also have to know what must never be tagged, and that is the missing half. Do not tag internal links on your own site: it resets the visitor’s session and reassigns them a new origin, destroying precisely the information you were after. It is a frequent error and it is silent.
Two domains, two visitors
Many businesses have more than one domain without thinking about it: the main site, a shop on a subdomain, a booking page at a provider, a form hosted elsewhere. To the measurement tool, those are distinct places.
By default, a visitor moving from one to the other is counted twice: they end their visit on the first and start a new one on the second, with no link between them. The real journey — one person, one intention — becomes two visitors neither of whom made the full trip.
That produces two simultaneous errors running in opposite directions, which makes them hard to spot. The visitor count is inflated, and the first domain’s conversion rate is collapsed, since the conversion happened on the second. Both figures are wrong and they do not visibly contradict each other.
The setting that resolves this exists and consists of declaring that those domains form one set. It has to be put in place when the second domain appears, which is rarely done, because that day a shop is being installed and nobody is thinking about measurement.
Bots
Part of the recorded traffic is not human: search engines, monitoring tools, content scrapers, uptime probes you may have installed yourself. Serious tools filter the best-known ones, and they do not filter the rest.
The characteristic sign is recognisable: a sharp spike, on a single page, with zero duration and no interaction, often in the middle of the night and sometimes from a country you do not sell to. It repeats at regular intervals, which is the signature of a program rather than an audience.
On a high-traffic site that noise dilutes and does not matter. On a small Algerian site — a few hundred visits a month — a bot passing every hour can represent half the figures and completely shift the reading of which pages are popular.
The fix is modest and sufficient in most cases: check that known-bot filtering is switched on, which is not always the default, and explicitly exclude the monitoring tools you set up yourself. That last point is the one that gets forgotten, because nobody thinks of their own probe as traffic.
Double counting
The opposite defect to an outage also exists and is more flattering, therefore less challenged: the same tracking code installed twice. Every visit is counted twice, every conversion too, and the reports describe a business twice its size.
The usual cause is historical stacking. The code was placed directly in the page template by a developer, then again by the tag manager installed later, then a third time by a module added for an integration. Each of those additions was reasonable and nobody removed the previous one.
The clearest sign is an abnormally low bounce rate — a few per cent — because the second firing is interpreted as a second interaction. A site whose bounce rate looks too good should be suspected before it is celebrated.
The check is to look at the page source in a browser and count occurrences of the measurement identifier. It is a thirty-second move nobody makes, and it settles a question that would otherwise be discussed for months — because flattering figures prompt no investigation.
The ten-minute check
Everything above comes down to a short monthly check, and its value is that it catches an outage while it still costs little. Broken measurement found after a month costs a month; found after a year, it costs the ability to compare anything with the previous year.
Four moves, in this order. Open the real-time report and load your site from a phone: you should see yourself appear. Place a test order or fill in the test form and check the conversion is recorded, with the right source. Compare the tool’s order count with your management software’s. And look at the source list to see whether a new absurdity has appeared in it.
The third move is the most important and the most neglected. Two systems counting the same thing and diverging is the most reliable signal you have, and it is precisely the check nobody does because it requires opening two tools instead of one.
As with any periodic check, what decides is not the method but the anchor. Attach it to a moment that already exists — the month-end close, the monthly meeting the companion article describes — or it will be done twice and then forgotten, and this page will have changed nothing.
Rebuild day
The site rebuild is the leading cause of lost measurement, and it is also the only one whose date is known in advance. That makes it the one case where the outage is entirely avoidable, provided three things are written down beforehand.
Before going live: the list of measurement identifiers in place, the list of tracked conversions with how each is triggered, and a capture of the previous month’s figures. That last looks like nothing and is what will make it possible to say, two weeks later, whether the drop is real.
On the day, a full check before considering the launch finished: the code present on every family of page — home, category, product, basket, confirmation — and not only on the home page, which is the one everybody tests and the only one whose presence proves nothing about the others.
The following week, a quick daily comparison against the earlier capture. A successful rebuild moves the figures, sometimes a lot, and that is normal; what is not normal is a zero, or a source that has vanished entirely, or conversions that have stopped while sales continue in your management software.
Taking over inherited measurement
Inheriting an installation is the most frequent case, and it arrives in a recognisable state: several tools nobody knows the purpose of, properties created by successive providers, filters whose object has been forgotten, and access depending on a personal account.
The first thing to do is not to clean up, it is to establish ownership. Who is an administrator, on which account, and does the business hold that level of access anywhere. While the answer is no, everything else is work done on an asset you do not control, and which can disappear with a person.
Then comes the inventory, read-only: which codes are actually loaded by the pages, which conversions still fire, which filters are active. That inventory nearly always reveals at least one of section two’s six causes, often installed long ago.
The rule of caution is the same as for any takeover: delete nothing while you are still looking. An incomprehensible filter may be excluding real noise; an unused property may hold the only history of two years. What gets deleted in measurement is never rebuilt, because past data exists nowhere else.
What we do, and what we refuse
What we take on is bounded: the inventory of what is actually installed, correcting the six causes where present, the link-tagging convention, the payment-domain exclusion list, and the first monthly check done with you.
We do not install a second measurement tool beside a first one that is not working. It is the easiest proposal to sell — a new tool, a new dashboard — and it leaves the outage in place while adding a source of disagreement: two tools counting differently produce meetings about which one is right, which is time taken from the business.
We do not reconstruct data for a period that was not measured. It is the request every discovered outage produces, and it has no honest answer: what was not recorded exists nowhere, and an estimate presented in a report becomes, within six months, a figure somebody cites. We mark the period as unmeasured and leave it visible.
Finally, we do not promise the measurement will be accurate. It never is: ad blocking, refused consent and browsers that limit tracking mean some visitors will not be counted, and that is a property of today’s web rather than an installation fault. What we can achieve is measurement that is consistent over time — comparable with itself — and that is enough to decide on, which absolute accuracy would not add to.
Frequently asked questions
How do we know whether our figures dropped or the measurement broke?
Two reliable signs. A commercial drop is gradual and uneven — some channels, some days; an outage is sharp, dated and uniform across everything the tool counted. Above all, compare with a figure from another source: orders in your management software, calls received, money taken. If one has not moved while the measurement collapsed, the question is settled in thirty seconds.
Why do our sales appear to come from the payment provider?
Because a customer who left to pay on an external service returns to your confirmation page as a fresh visitor from that service. The sale is counted but credited to the wrong channel, and advertising or search appear to return nothing — which gets the working channel stopped. Declare payment domains in the source exclusion list, then place a test order and check its attribution.
Should we exclude our own team from the statistics?
Yes, and on a small site it is often the correction that makes the figures usable. The misleading effect is not volume but behaviour: you visit pages customers do not and never convert. Exclusion by network address is enough with a fixed connection; when the team works on mobile, which is common here, set a marker once on each device through a dedicated page instead.
Our reports contain forty different sources. Why?
Because link tagging is a spelling question before it is a technical one: “facebook”, “Facebook” and “fb” create three sources. Decide a three-line convention — all lowercase, one separator, a short list of permitted values — and put it where the links are made. And never tag internal links on your own site: it reassigns the visitor a new origin and destroys the information you wanted.
Our bounce rate is very low. Is that good?
It is grounds for suspicion before it is grounds for satisfaction. A bounce rate of a few per cent is the characteristic sign of a tracking code installed twice — typically placed in the page template by a developer and then again by the tag manager. Look at the page source in a browser and count occurrences of the identifier: thirty seconds, and a question that would otherwise be discussed for months.
What should be done before a site rebuild?
Write three things down: the list of measurement identifiers in place, the list of tracked conversions with their triggers, and a capture of the previous month’s figures. On launch day, check the code is present on every family of page — home, category, product, basket, confirmation — not just the home page, which is the only one everybody tests and whose presence proves nothing about the others.
Where we come in
Two systems counting the same thing and disagreeing is the most reliable signal you have. What they do not say is which of the six causes has you.
- We list the tools still active, including the ones nobody ever removed.
- We isolate the cause in one session, working back from the test order.
- We put an alert on the gap between your two counts.
We do not reconstruct an unmeasured period, however it is asked for: that month stays a hole, and it is better to know it.
Read next
Analytics: five numbers, defined, every month
A useful report fits on one page and carries the same five numbers every month. The rest describes the agency’s activity, not yours.A dashboard that says no
A dashboard earns its place through what it refuses to assert. Three refusals to insist on, and the hole never to fill in.A dashboard is read or it dies: the ten-minute review
Most dashboards are not bad, they are abandoned. What keeps one alive is a habit rather than a feature.
Let us talk about your project
A free audit, no commitment: we look at your online presence and tell you what is holding it back.