Reading a dip without panicking

The line goes down and you start looking for a reason. Within a minute you have one — the post that underperformed, the change you shipped Thursday, the competitor who has been busy lately. Notice what just happened: you did not test a theory, you generated one, and the search for a cause has a perfect success rate because there is always something that happened recently. That reliability is the trap. A method that produces an answer every single time is not diagnosing anything; it is pattern-matching, and it will happily hand you a confident explanation for a dip that has no cause at all.

Start with what ordinary looks like, because almost nobody writes it down and it is the only baseline that makes a dip readable. Suppose you average five hundred visits a day. Days like that do not arrive as five hundred, five hundred, five hundred — they arrive as four-thirty, five-sixty, four-eighty, five-twenty, and a Tuesday at four-ten is inside the ordinary spread rather than outside it. The smaller your numbers, the wider that spread gets in percentage terms. At fifty visits a day, a swing from forty to sixty is a forty percent move that means nothing whatsoever. If you have never worked out what a normal week's wobble looks like for your own product, every wobble is going to read as an event.

Then check that you compared like with like, which is where most alarming percentages are manufactured. Last seven days against the previous seven is the default in every tool and it quietly compares different things: a week containing a public holiday against one that does not, or a week with five weekdays against one with four. Weekday shape alone can account for the whole story — a product used at work has a Monday that looks nothing like its Sunday, and a stretch that happens to be weekend-heavy will show a drop that is really just a calendar. Compare the same weekday to the same weekday, and compare periods of equal shape, before you believe the number in the corner.

Seasonality is the next layer, and it is the one small products are least equipped to see, because seeing it normally requires years of your own history and you have months. So borrow the calendar instead of the chart. If your buyers work in offices, mid-August and the last two weeks of December are quiet everywhere, and a dip that lines up with a holiday most of your audience shares is a fact about their year rather than a failure of your marketing. If your product tracks school terms, or tax dates, or a weather-dependent activity, the same applies. This is the cheapest diagnosis available: before you open a single report, look at what week of the year it is and ask who is not at their desk.

every dip has a story available within a minute. the true ones are boring, and the exciting ones are usually just a tuesday.

None of which means dips are never real, and the real ones have a shape you can learn to recognize. A genuine break is usually sharp rather than gradual, and it is usually narrow rather than everywhere. Traffic from one source falling to nothing overnight while everything else holds steady is not seasonality — that is a link, a page, or a recording problem, and it deserves an immediate look. A gentle slide across every source over three weeks is a different animal entirely and almost never has a single cause. Sharp and narrow means go and check something specific. Broad and gradual means wait, and keep shipping.

So the first move on any dip worth investigating is to split it rather than stare at it. Break the number down by source, by landing page, by device, by country. A drop that appears in every row is telling you about the world or about your measurement. A drop that lives in exactly one row is telling you where to go and look, and that is a five-minute job instead of an afternoon. The single most common finding, and the one people check last, is that nothing dropped at all — a link stopped carrying its tag, so the same visits are now sitting in a row called direct, which is the same failure mode as a campaign quietly splitting into four rows with different spellings.

Be honest about the cost of getting this wrong in the panicked direction, because it is larger than it looks. A dip read as a verdict produces action: a page gets rewritten, a campaign gets killed, a cadence gets doubled. Each of those resets something that needed stable time to prove itself — and now the following weeks are not readable either, because you changed three things at once during a period you did not understand. This is the same trap as killing a campaign on a number too small to mean anything, and the same as reading a two-place move in your average position as a judgment on your pages. Reacting to noise does not just waste the afternoon. It destroys the baseline you would have needed to read the next month.

There is a version of this that is worth building into how you look at things, and it takes about ten minutes once. Write down, for your two or three real numbers, what a normal range actually is and what a genuinely bad day would be — an actual figure, decided while nothing is wrong. Write down which dates you already expect to be quiet. Then, when something drops, the question stops being "what could explain this" and becomes "is this outside the line I drew when I was calm," which is a question with a real answer. Roughly the same thing is true of what we put on one screen in SiteOps: traffic, revenue, positions and spend sitting together, with panels that say plainly when a provider is not connected instead of drawing a zero you might read as a collapse. A number missing and a number falling should never look alike.

The rule of thumb, then: a dip earns an investigation when it is sharp, narrow, and outside a range you set in advance. It earns a note and nothing else when it is broad, gradual, or lines up with a week when half your audience is on holiday. And when you cannot tell which one you are looking at, the correct action is almost always to wait a week and keep doing the work — because the cost of a week of patience is one week, and the cost of a panicked rewrite is every reading you take after it.