You looked straight at it

A lesson from Why Good Design Dies. Play it above, or read it through below.

• A quarter of the content area. One fixation inside it.

  The common read: A quarter of the page is a quarter of what there is to look at, so a slot that size gets a share of the attention in proportion to it.

  A quarter of the area is not a quarter of the looking. A page position has an identity, and a reader learns it before they read anything in it. In the 2018 eyetracking work, 26 participants worked one task on a real page carrying ads in the top banner and the right rail. The page took 148 fixations. 132 of those landed inside the content area, and exactly one of the 132 landed in the right rail, which is a quarter of that area by size.

  That is 0.8 percent of the attention inside 25 percent of the content area. The same readers read the body text beside the rail, so the attention was there and the rail did not get it. What the rail held was the site's ads, and what the readers brought was a verdict on the slot, learned somewhere else and spent here before anything inside it was resolved. On your own site that verdict resurfaces as a heatmap filed against the visual design.

  This is not one study and it is not new. NN/g first documented the behaviour in 1997 through usability testing, replicated it with eyetracking in 2007, and found it again in 2018, which they state as three decades. So price the slot before you price the restyle. Moving a module out of that rail costs one layout argument and one sprint. Leaving it there spends the module on a quarter of the page that took one fixation in 132.

• Nothing about the module was read. The thing beside it was.

  The common read: An element nobody looks at is being skipped for something about that element, so the fix is somewhere in the element.

  Proximity did it, and the module was never read on its own terms at all. The factor you already have, where the smallest interval decides what pairs with what, is the one NN/g reaches for here, in its own words: content placed in the same part of the screen as an advertisement is often considered to be an ad and is also ignored, as a simple consequence of the Gestalt law of proximity.

  Which makes this a different failure from the rail, with a different fix. The rail is a slot the reader learned somewhere else. This is a group the reader forms on this page, in the moment, out of two things that happen to sit together. Nothing in the module caused it, so nothing you do to the module ends it. Separation does: space, a rule, a band of its own, anything that stops the two reading as one thing.

  The ticket history on one of these reads: "Sidebar pick gets no clicks, give it a border." Then: "Sidebar pick still gets no clicks, try the brand colour." Then, eleven months later: "Sidebar pick gets no clicks."

  And one bad fixation is not local. NN/g reports that such an experience may change not only local behaviour on that specific page but can also influence general web behaviour, so a reader may not look at the same area on other pages or even other sites. That hedge is the source's own and it is worth keeping intact. Treat a region's first impression as the expensive one, because from that person you do not get a second measurement of it.

• The phone does not make it easier. Size takes the choice away.

  The common read: A phone screen is small and readers scroll it fast, so an inline ad is the easiest thing on the web to skip past.

  A phone screen is smaller and scrolls faster, so the honest version of what eyetracking finds there is that a full-width inline ad is

  Size takes the choice away. NN/g's wording for the mobile case is that a mobile inline ad is so large in proportion to the screen size that users cannot avert their eyes from it as they swipe down, and that holds for readers who know exactly what it is and do not want to look at it.

  So the mechanism is not one rule with a phone variant. It reverses. On desktop, a learned slot buys avoidance, and a module sharing that slot inherits the avoidance. On a phone, an element large enough relative to the viewport gets fixated whatever the reader thinks of it, which is good news for the thing you want read and bad news for whatever is sitting next to it.

  That arrives as a mobile heatmap contradicting the desktop one, filed as an instrumentation problem, closed by nobody. The number it usually produces is an engagement figure for the inline unit, which a room reads as interest. A fixation is not interest. So decide a full-width mobile slot by what you are willing to have seen. Deciding costs you one argument about what that slot is for. Not deciding costs you whatever sits beside it, because the gaze arrives regardless.

• Open eyes are not attention. The conversation is spending it.

  The common read: Their eyes were pointed at it and nothing was covering it, so they saw it and chose not to act.

  Both word-generation conditions impaired detection, equally, and the radio, the book on tape and repeating words back did not impair it at all. Strayer and Johnston ran a simulated drive with a pursuit-tracking task and intermittent red and green lights, against radio, a book on tape, an easy shadowing task on a handheld phone, and a harder word-generation task on handheld and on hands-free phones. Handheld and hands-free came out level. The handset was never the variable.

  The paper could not be read this year. The publisher returns a 403, the index page returns a consent wall, and no participant count appears in anything reachable, so no number from this study ships anywhere in this lesson. The boundaries ship, and the boundaries are the half that changes a decision: passive listening cost nothing, repeating words back on a handheld phone cost nothing either, generating them cost detection, and the deficit grew as the driving task got harder.

  Seeing is not a property of where the eyes point. It is bought out of a budget, and a conversation has already spent most of it. An element in plain view, in a slot with no history, goes unregistered while the person looks straight at it: a second cause of "did not notice it", with nothing to do with position. When a screen matters and somebody will be talking, buy that visibility with a second channel, a sound and a line of copy, not layout alone.

• One symptom, two causes. The fix for one does nothing for the other.

  The common read: Not noticed is one finding, so one fix covers it.

  Two participants both walk past the same module in the same session, and the finding you can honestly write down is that

  They separate on what they depend on. A learned slot depends on the position: it survives a reader with all the attention in the world, it does not care what the element looks like, and it moves when the element moves. A spent attention budget depends on the load: it strikes a slot with no history at all, it does not care where the element sits, and it moves when the competing task moves.

  So the run that tells them apart moves one variable and holds the other. Put the identical element in a slot with no advertising history, keep the concurrent task exactly as it was, and watch. Still missed means attention was the constraint. Found means the slot was. Restyling answers neither question, and a larger sample repeats whichever confound you started with, more expensively.

  And the fixes do not overlap. A learned slot is answered by moving out of it, or by separating from whatever taught it. A spent budget is answered by a second channel, by holding the decision until the task spending the budget has ended, or by not putting the decision there at all. Apply either fix to the other cause and you have spent a release to change a variable that cause does not turn on.

  That arrives as a redesign in a new visual language, shipped against a finding that said users miss the alert, with the alert missed at the same rate afterwards. The readout never separated the two, because the template has a column called issue and no column called cause. So write the cause into the readout before anybody in the room prices a fix.

• The report was accurate. It was not evidence about the screen.

  The common read: The session recording shows the banner on screen for eleven seconds, so the user telling me it never appeared is misremembering.

  The recording shows the banner on screen for eleven seconds with the cursor crossing it twice, and the user says it never appeared, so the accurate reading is that

  Both records are accurate and they measure different things. A session recording is a record of what the screen did. A person's report is a record of what reached them. When the two disagree nothing is being misremembered, and treating the report as a memory problem is how a real finding gets closed as cannot reproduce.

  Which is a problem for the room the test runs in. A moderated session is one participant, one task, a quiet room and somebody watching, and that protocol removes the competing load on purpose so the task can be observed at all. It is the right protocol and it has a blind spot with a name: inattentional blindness from divided attention, and every failure with that cause has been engineered out before the session starts.

  So run one condition that keeps the load. A scripted phone call the participant has to hold up their end of, a second task in another window, an interruption on a timer. It costs a recruiting line and one extra session block. Skipping it costs you the failure class your product actually produces, and you meet it later in a support queue where nobody can reproduce it.

  And take the sentence literally when it arrives. "It was never shown" is a true statement about somebody's experience and no statement at all about the screen, and the two get confused because they share a verb. Ask what else they were doing, and ask where the thing sat. Those two answers are the diagnosis, and they cost one question each.