How many KPI tiles a page can carry

The page had twenty-two tiles. I counted them twice because it seemed unlikely. Four rows of five, plus a two-tile strip along the bottom that had been added when somebody wanted a sponsorship number and there was no room left.

The show director who used it every Monday could name three of them from memory. Not because she was inattentive. She had learned where the two numbers she needed were, plus the one her chief executive asked about, and she had stopped seeing the rest, in the way you stop seeing a poster you walk past daily.

How many KPI tiles a page can carry is therefore a question about the reader rather than about the screen. The screen will hold twenty-two comfortably. The reader will not.

What happens when you add the eighth tile?

Nothing visible, which is the problem. The page still loads, the tile still renders, and the person who requested it can see their number, so the change looks free.

What actually happens is that the attention arriving at the page gets divided one more way. A reader who gives the page ninety seconds and takes in six tiles is taking in six. Add a seventh and they still take in six, and one of the original six drops out. Which one drops is decided by position and size rather than by importance, so the tile lost is likely to be the one in the least prominent slot, which in most layouts is the one added most recently before this.

The consequence worth sitting with: a page past its capacity does not degrade evenly. It converts from a monitoring surface, which the reader scans to find out what has changed, into a lookup surface, which the reader searches to find a number they already knew they wanted. Everything a monitoring surface is for is lost at that transition, and the transition is silent.

Miller's seven, and what he actually wrote

George Miller published The Magical Number Seven, Plus or Minus Two in Psychological Review in 1956, volume 63, pages 81 to 97, and every tile-count argument since has quoted it. It opens with the line "My problem is that I have been persecuted by an integer", which should be a warning about how the paper gets used.

Miller measured the span of absolute judgement, the number of categories along a single dimension a person can reliably distinguish, and put it "usually somewhere in the neighborhood of seven". He also measured the span of immediate memory, which is a different quantity, and was explicit that "the span of absolute judgment and the span of immediate memory are quite different kinds of limitations". On why the same figure turns up in both, his own view was that "I suspect that it is only a pernicious, Pythagorean coincidence".

So the paper does not license a seven tile dashboard. It licenses caution about seven of anything.

Nelson Cowan revisited the question in Behavioral and Brain Sciences in 2001, in a target article arguing that the real capacity limit for chunks in working memory sits nearer three to five once chunking and rehearsal are prevented. That figure is the more relevant one for a dashboard, because a reader scanning a page cannot rehearse and is holding several tiles in mind at once to compare them.

My own reading of the two together is that the tile budget should be five or six. Seven is defensible when the tiles are genuinely simple, meaning a number and nothing else. It is too many when each tile carries a value, a comparison and a label, which is three things to hold rather than one.

Microsoft's Power BI implementation planning guidance, published on Microsoft Learn and dated December 2024, arrives at the same place from an operational direction, advising teams to "Measure a limited number of KPIs or key results" because "It's better to have a few strategic, valuable indicators than many which you can't regularly monitor or follow up". The test there is follow-up rather than perception, and it lands on a similar number for a different reason.

An attention budget in seconds

The version of this argument that survives contact with a stakeholder is arithmetic rather than psychology, because a stakeholder can argue with a citation and cannot argue with their own behaviour.

Time how long the page is actually open. Most weekly report visits by a busy operating lead run somewhere between sixty and a hundred and twenty seconds. Take ninety as a working figure and check it against your own usage logs, which is a measurement your workspace already collects even if nobody looks at it.

Now cost a tile. Reading one properly means taking in the label, the value, the comparison and forming a verdict. Call it six seconds, which is generous towards the tile. Ninety seconds divided by six is fifteen tile-reads, and that assumes the reader does nothing else, applies no filter, and reads no chart.

Fifteen sounds like room for fifteen tiles. It is not, because the reader is not reading each tile once. A comparison across two tiles costs two reads. Noticing that the pacing tile and the space tile disagree costs three, since the third read is the check. On a page whose tiles relate to one another, which is the only kind worth building, actual reads run at roughly twice the tile count.

Fifteen reads divided by two is between seven and eight tiles as an absolute ceiling with no charts on the page at all. Put one chart on it, which costs perhaps twenty seconds of the budget on its own, and you are down to about eleven reads, which is five or six tiles.

That is the number, and it is derived rather than asserted, which makes it much easier to defend on a Tuesday afternoon.

Twenty-two tiles down to six

Back to the page. The twenty-two tiles were, in the order they appeared: total registrations, registrations this week, registration pacing index, verified attendance last edition, no-show rate last edition, registrations by type, registrations by country, registrations by job function, corporate domain share, exhibitor count, new exhibitor count, net square feet sold, net square feet against target, average stand size, sell-through percentage, rebooking rate, revenue booked, revenue against budget, sponsorship revenue, sponsorship inventory remaining, website sessions, and email click rate.

Six survived: registration pacing index, weekly registration movement, net square feet against target, sell-through percentage, revenue against budget, and a watch list of accounts whose renewal signal had moved.

Sixteen went to two detail tabs, split by the follow-up question that leads to them. An audience tab took registrations by type, country and job function, corporate domain share, verified attendance, no-show rate and total registrations, which is seven. A commercial tab took exhibitor count, new exhibitor count, net square feet sold, average stand size, rebooking rate, revenue booked, sponsorship revenue and sponsorship inventory remaining, which is eight. Website sessions and email click rate went to the marketing report where they had come from, since neither had a decision attached on this page.

Seven and eight on the detail tabs breaks the same budget the front page respects, and deliberately. A detail tab is opened by somebody who has already decided what they are looking for, so it is a lookup surface by design and the budget does not apply. The budget applies to pages people scan.

How do you decide which tiles survive?

Rank them by how often a tile has changed somebody's behaviour in the last year, which is a question you can ask rather than a number you can compute.

Go to the named reader with the full list and ask, tile by tile, when they last did something differently because of it. The answers sort quickly into three groups: tiles with a recent, specific story attached, tiles with a story from two years ago, and tiles where the honest answer is that it is nice to know. The first group is your front page. The third group is your detail tab, and nobody has to lose anything, which is what makes the conversation survivable.

Where this collides with the reader's instinct is the number they check for reassurance. Total registrations is usually one of these: nothing is ever decided from it, and its absence is noticed immediately. Keeping one reassurance tile is a reasonable use of a slot, and the important thing is to know that is what you are spending it on rather than pretending it is a performance indicator. Deciding which questions the page exists to answer at all is the prior step, and doing it first makes this ranking much shorter.

Where this stops

The budget assumes a page that gets scanned in a minute or two by a person with other things to do. Not every page is that.

An analyst's exploration page is a different artefact. Somebody who sits with a report for forty minutes, filtering and comparing, has an attention budget an order of magnitude larger and is not harmed by twenty tiles. Applying the front-page budget to an analysis surface produces a page that is too thin to work with, and the resulting complaint is legitimate.

The show week board is the other exception in the opposite direction. A page on a screen in an operations room, glanced at by people walking past, supports fewer tiles rather than more, and it needs a different refresh cadence and a different layout. That variant has its own constraints and the numbers here do not transfer to it cleanly.

The honest limit on the arithmetic itself is that six seconds per tile-read is my estimate rather than a measured constant. It comes from watching people use these pages and it will be wrong by a factor of maybe 1.5 either way depending on how familiar the reader is with the layout. The conclusion survives that error, since even at four seconds a tile the ceiling lands around eight, and even at nine seconds nobody is defending twenty-two.

Pull the usage log for your main report this week and find the median session duration. Divide it by six, halve the result, and you have a defensible tile ceiling for that specific page and that specific audience. Then count the tiles. The gap between the two numbers is the work, and deciding where the survivors sit comes next. Both are cheaper than any new report you were planning to build for event business intelligence this quarter.