Website KPIs — four ways your numbers are lying to you

A performance report is rarely empty. Far more often it is full and says something other than what it appears to say — and decisions get taken on a number nobody checked at source.

Below are four traps we hit measuring our own site. Each has a mechanism you can verify in the documentation, and each leads to a different mistake.

First, one decision that simplifies the rest: a goal gets one number, and somebody looks at it once a month. Not twenty charts on a dashboard nobody opens. A measurement changes a decision; a dashboard fills a slide.

Why you measure, and what the words mean

Measurement is not there to tell you how many people visited. It is there to answer one question: is what we are doing producing enquiries — and if not, at which stage do people drop out. Anything that does not help answer that is decoration.

There are three stages, each measured with something different. That distinction matters more than any tool: without it you cannot say where the problem is.

1. Are we visible at all. Does the site appear in search results for the questions customers ask. Search Console answers that.
2. Do people come in. How many of those who saw us clicked, and what they did next. Analytics answers that.
3. Do they leave contact details. How many sent a form, rang or wrote. Analytics again, but only if somebody deliberately set what counts as contact.

A firm with a problem at the first stage that goes looking at the third will spend its time improving a form on a page nobody sees.

Five words that mean different things in different tools

Half the misunderstandings in a results meeting come from people using the same words in different senses. Settle them once:

• Impression — your page appeared in search results. Nobody has gone anywhere yet. Search Console counts these.
• Click — somebody chose your result and landed on the site. The only moment both tools see, and the boundary between them runs through it.
• Session — one visit by one person, ended by default after thirty minutes of inactivity. The same person returning next day is two sessions, not two people.
• Event — anything somebody did: loaded a page, scrolled, clicked a button, sent a form. In GA4 everything is an event, including opening the page. That is the source of the second trap.
• Conversion, now a "key event" — an event somebody marked by hand as important. The tool does not know what is valuable to you, and by default marks too much.

The last definition is the one to remember: a conversion is not a fact, it is a decision — somebody's decision about what counts as success. If nobody made it deliberately, the tool made it for you.

Why these four traps

Each stage has its own way of misleading you. The first trap: the data never arrives complete. The second: we measure something other than what we call it. The third: we measure ourselves instead of customers. The fourth: we ask the wrong tool and get a correct answer to a different question.

We met all four on our own site inside a single week. None required an outage or a botched implementation — the default settings were enough.

Trap 1: the consent banner cuts your measurement without warning

The symptom is misleading: visits fall while scrolls stay. It looks as though the data is broken, or as though people stopped reading.

The mechanism is documented outright. When a user declines consent, Google tags still send pings without cookies — the consent mode documentation calls them cookieless pings. The events arrive but are not tied to a persistent user identifier, so some reports see them and some do not. Hence the divergence between sessions and events, which looks like a fault and is working as designed.

Google's answer is modelling: machine learning infers the behaviour of non-consenting users from those who consented. But it has thresholds — the property must collect at least 1,000 events a day without consent for 7 days and have at least 1,000 users a day with consent for 7 of the last 28 days.

And that is the crux for a small firm: you will not hit those thresholds. The modelling meant to patch the hole will not switch on for a site with a few hundred visits a month. The hole stays, and you need to know it — otherwise every dip gets explained by the market, the season or a competitor.

How to check whether this affects you. Open the events report and compare page_view with scroll or user_engagement over the same period. If there are noticeably fewer page loads than scrolls — and you cannot scroll something that did not load — you are looking at this gap. The second signal is a sudden step on the day the consent banner appeared or changed.

What to do about it: accept that the numbers are understated by an unknown but fairly constant amount, and compare periods against each other, not against reality. The trend is trustworthy; the absolute value is not. In practice: do not promise anyone "a thousand visits a month" as a target, because what you measure is traffic minus refusals. Set the target as a change against the previous quarter.

And one thing worth saying plainly, because it is a source of bad decisions: this is not a reason to redesign the consent banner. The banner has to be lawful, not convenient for the statistics.

Trap 2: "key events" are not leads

This mistake costs the most: it leads to reporting a success that did not happen.

In GA4 any event can be marked as key. In practice too many are: after migrations and with default settings, scrolls, user engagement and subpage views all end up marked as key.

Measured on our own property in September 2026, on traffic from a single channel: 576 events, of which 536 were marked as key. Almost all of them. A report reading the "key events" column would have shown hundreds of conversions against zero forms sent.

Here is the conversation that follows. The agency brings a report with five hundred conversions, everyone is pleased, and the inbox holds not one new enquiry. Nobody lied — the conversion was a scroll.

What to do about it: do not read the aggregate column. Write out the names of the events that genuinely are contact — form submitted, phone tapped, email revealed — and count only those. If you cannot point at the name, lead measurement does not exist, whatever the chart shows.

And one check worth running quarterly, whatever your tooling. Send an enquiry through your own form and count how long it takes anyone to reply. An event says the form was sent. It does not say whether the message arrived, whether it went to a spam filter, or whether the mailbox has an owner. We have seen forms counting conversions faultlessly for months from which no message reached anybody.

The number in the reporting and the number in the inbox should agree. If they do not, that is not a measurement problem — it is a sales problem visible in the measurement.

Trap 3: your own traffic stays in the data for good

Visits by the team, post-launch testing, automated checks — all recorded exactly like customer traffic.

At our end, a single automated test run against production added 98 false hits before anyone noticed. At a few hundred visits a month that is not noise but distortion large enough to produce the wrong conclusion about which pages matter.

The warning signal is simple: the ratio of events to users. If a handful of people generate an unnatural number, you are most likely looking at your own team or at a tool.

What to do about it: define internal traffic in the data stream settings — GA4 then adds a traffic_type parameter with the value internal — and exclude it with a data filter.

And the thing that settles the order: a data filter does not work retroactively and needs 24–36 hours to take effect. Traffic already in the data stays there permanently. So excluding your own traffic is something you do before the first test, not after the first odd report.

Your own team is not the only source of dirty data. Three more worth checking:

• Uptime monitoring. A service checking every minute whether the site responds can generate more visits than customers do. It usually shows as traffic from one location, at an even rhythm, with zero time on page.
• Contractor and agency tooling. SEO crawlers, audit bots and preview plugins enter the site exactly as a person does.
• Referrals from sites you do not recognise. Domains with no connection to your trade, with zero time on page, are not visitors — that is noise.

None of these would skew a large shop's numbers. All of them can overturn the statistics of a firm with a few hundred visits a month — most of the firms this article was written for.

Trap 4: Search Console and Analytics answer different questions

This is the commonest diagnostic mistake we see, and the one that sends a firm in the wrong direction for longest.

Our own case is the best illustration, because for a long time we drew the wrong conclusion from it ourselves. Analytics showed that almost nobody read the blog. The natural reading: the writing is weak.

Search Console said otherwise. Of 182 published articles, 124 had not had a single impression in search results — not "a poor position", complete absence. Over 90 days the whole blog collected one click against 4,747 impressions.

The writing could not have been weak in a reader's eyes, since it was never shown to a reader. That is a different problem and takes different work: Analytics diagnoses what happens to a person who already arrived; Search Console, whether they ever got the chance.

The rule that settles it in a minute: open Search Console, set the range to three months, look at two numbers.

• Impressions are there, clicks are few → the problem is in the title and description in the results. The page is in the game, nobody is choosing it.
• Almost no impressions at all → the problem is indexing. Improving the writing changes nothing while the search engine is not showing the page.
• Visits are there, but not from search → the problem is the site itself, and only then does the work in Analytics start.

And what to do in each of the three cases, cheapest first:

With impressions but no clicks, rewrite the title and description of the pages with the most impressions and fewest clicks. It is the only change of the three visible in weeks, and the only one that does not require touching the site.

With no impressions, check first whether the site is submitted and indexed at all — Search Console has a separate indexing report and a tool for checking a single address. Only then does thinking about content make sense. Writing more articles for a site the search engine does not show is the most expensive way of not solving the problem.

With traffic from outside search, look at where people actually come from and whether the landing page answers what they came for. A campaign pointing at the home page instead of a specific offer loses most of what it paid for.

What to actually measure

The indicator is chosen to fit the goal, not what the tool offers. The right-hand column matters as much as the middle one — those numbers grow independently of sales, which is why they end up in reports most eagerly.

Every indicator needs a baseline recorded before the change. Without one, after a quarter nobody can say whether anything improved — and the discussion becomes an exchange of impressions, won by whoever holds their view more firmly.

One example of an indicator pointing at the work: Baymard Institute research, pooling 50 separate studies, puts average cart abandonment at 70.22%. If your shop sits around there, abandonment is not an anomaly to put out but a norm — and the work lies elsewhere. The number alone says nothing until there is something to hold it against.

The minimum set to start with

Three tools, all free, in a specific order. Firms usually have the first and not the other two — and it is those two that answer the questions asked most often.

Google Analytics 4 — what people who already arrived do. Three things suffice at the start: internal traffic excluded, a correctly named contact event, and one report somebody opens monthly. The rest can wait a year.

Google Search Console — whether the search engine shows you at all. It needs no configuration beyond verifying the domain, and answers a question Analytics cannot pose. It is the tool whose absence we most often find at firms arriving with "why is there no traffic".

Microsoft Clarity — why people do what they do. Session recordings and click maps.

Deliberately not on the list: A/B testing tools and paid measurement platforms. At a few hundred visits a month an A/B test will not reach statistical significance in a reasonable time — you will be making decisions on chance and calling it data. It starts making sense at traffic where the difference shows in weeks.

How to read an agency report

Five questions that settle whether a report describes work or merely illustrates it. Each follows from the traps above.

1. "Which specific events do you count as conversions?" The answer has to be a list of names, not a number. If it is "GA4 conversions", we are back at the second trap.
2. "How many of those conversions reached our inbox?" The numbers should agree. The divergence matters more than the value.
3. "Did you exclude internal traffic, and from when?" The filter is not retroactive, so the date matters.
4. "How many impressions did we have in Search Console?" A report without it describes only the bottom half of the path and cannot say whether the problem is the site or the visibility.
5. "What changed against the previous period, and why?" A report giving a state instead of a change is not for taking decisions — it is for closing the month.

These are not trick questions and a good agency answers all five without preparation. Difficulty with any tells you more than the whole set of charts.

Analytics tells you WHAT, Clarity tells you WHY

This distinction saves the most time and costs one installation.

Analytics answers "what happened" — that eighty per cent of people leave the page with the contact form, say. You know where you are losing people and not why.

Microsoft Clarity answers "why" — it records sessions and shows click maps. You watch a recording and see the "send" button not responding in one browser, with people clicking it several times in a row. The tool calls that rage clicks, a signal no chart will show.

We use it on this site and it is free. The order matters: first the indicator that says where the problem is, then the recording that says what causes it. The other way round ends in watching recordings with no hypothesis.

Where to get a baseline if nobody recorded one

The commonest situation: you want to change something and nobody knows what you are starting from. Three ways, in order of reliability.

From the data you already have. Analytics and Search Console keep history — even unlooked at, the data was collecting. Take three full months before the planned change and write the numbers into a document with a date.

From the inbox and the phone. If there was no measurement at all, count last quarter's enquiries by hand. It sounds primitive and is more accurate than most measurement setups, because it counts things that actually happened.

By waiting a month. The least attractive option and sometimes the only honest one. A month of measurement before a change saves a quarter of argument after.

What not to do: do not take industry averages off the internet as a reference point. You do not know how they were counted, on what sample, or whether they describe your market — and you are comparing yourself against yourself anyway, not against an industry.

The cycle that actually works

Three steps and one rule.

1. Measure the baseline and write it down. Not in your head — in a document with a date.
2. Change one thing. One, not five.
3. Check after a sensible interval — at low traffic that is weeks, not days.

The rule: one change at a time. After five simultaneous changes the indicator may rise and you still will not know which helped and which did less harm than the others did good. The same rule applies to diagnosing a drop in rankings, for the same reason.