Visual Hierarchy and AI

Why AI defaults to centred, even layouts, and the hierarchy vocabulary that ranks them.

Ask a generator for a poster, cold. The title lands centred, the subtitle centred beneath it, the date strip centred beneath that: three elements stacked down the middle, evenly spaced, nothing ranked above anything else. Ask for a SaaS landing page, cold. Hero centred, a three-column feature row weighted the same across all three, one call to action competing with five other links. Different surface, same shape. The model fills the frame and balances it, and calls that done.

The clearest write-up of this makes the point for composition. Prashanthi Anand Rao, working with AI-generated images, observes that the model tends toward common layouts unless told otherwise, and puts it plainly: composition is not automatic; it must be written into the prompt. Hierarchy appears to follow the same pattern one layer up, though the evidence here is a composition finding, not a hierarchy-specific measurement. Centring is not ranking. A layout where everything is balanced is a layout where nothing is primary, and a reader scanning it has no idea where to look first.

Hierarchy is the deliberate ranking of elements by importance, and it has to be written into the prompt in the vocabulary designers already use for it: focal point, scale ratio, eye path, F-pattern, Z-pattern, level count. A designer who can name a hierarchy can prompt one. Everyone else gets the even, centred default, faster than before.

This is the third vocabulary set in the Fundamentals pillar, after Typography and AI and Colour Theory and AI. Same shape every time: name the vocabulary, show the uplift, apply it. One caution sits underneath, the same one that runs through the cluster: no controlled trial has tested hierarchy-vocabulary prompts against plain ones. The default is well reported and the vocabulary is well documented; the uplift in the demos below is practitioner-observed, not measured.

A reader judges a layout before reading a word of it. Where the eye lands, what it ranks as important, whether the page feels ordered or arbitrary, all of that happens on first scan, in the time it takes to glance. Get the ranking wrong and the page reads as untrustworthy, and the reader rarely knows why. Hierarchy is how a layout earns that first read.

This is also why hierarchy survives the move to working through AI rather than directly. Christopher Butler, in Craft is Untouchable, lists what does not vanish when the tool changes: “Structure still communicates before content. Visual hierarchy still guides attention. Negative space still creates rhythm.” The principle does not move. What moves is where you apply it: instead of nudging elements on a canvas, you rank them in a prompt.

The stakes are higher with AI, not lower. The centred default looks finished and reads flat on first scan: it fills the space and forgets to rank it. Because it looks complete, the missing hierarchy ships unnoticed. That is the cost of hierarchy by accident.

Brief: “Landing page for a B2B SaaS product, audience is mid-market operations directors.”

Without the vocabulary. Prompt: “Design a SaaS landing page with clear visual hierarchy.” Output: an H1 centred above the fold, a three-column feature row weighted evenly, a footer as visually loud as the call to action. The model heard “clear hierarchy” and produced “evenly distributed content,” because in the training-data middle those two read as the same thing.

With the vocabulary. Same generator, same brief. Prompt: “Landing page for a B2B SaaS product. H1 at large scale, semibold, left-aligned in the upper third as the F-pattern anchor; a single focal call to action above the fold at roughly 1.5 times the body-button scale; supporting copy at reduced weight following an F-pattern down the left edge; a scale ratio of about 3:1:0.6 across heading, sub-head, and caption; at most three or four levels of hierarchy; generous negative space on the right.” Output: scannable, ranked, and defensible to the audience it is for.

The glossary doing the work here is small. F-pattern is the Western reading default for content-heavy pages: the eye anchors top-left, sweeps right, then runs down the left edge looking for the next anchor. Scale ratio is the numeric relationship between levels, where 3:1:0.6 means the H1 is three times the body, the sub-head about one body-unit, the caption around sixty per cent. Level count is how many distinct ranks the design uses; past three or four, the eye stops ranking and starts skimming. Each term shrinks the output toward a specific layout, and each would be invisible to a prompt that never learned it.

Brief: “Film festival poster, independent programming, autumn season.”

Without the vocabulary. Prompt: “Design a film festival poster.” Output: title centred in the upper third, dates centred in the middle, festival mark centred at the bottom. Three centred elements in a column, symmetrical and forgettable, the default again.

With the vocabulary. Same generator, same brief. Prompt: “Film festival poster. Title dominant in the upper-left focal-point quadrant; festival dates as a supporting element at roughly 30 per cent scale, lower right; a Z-pattern eye path from the title across to the programming list, then down to the dates; asymmetric balance anchored by negative space in the upper right; no centred elements.” Output: ranked, directional, and built around a clear first read.

A different surface calls for a different pattern. Z-pattern is the eye path for image-heavy or low-text layouts: top-left to top-right, diagonally down, then across to the bottom-right. Focal-point quadrant names where the first-read element sits, and a corner placement breaks the centred default where a centred title would surrender to it. F-pattern suits a content page scanned top to bottom; Z-pattern suits a poster read at a glance. Neither always wins. The surface and the reading task choose, not a rule.

Five words cover most of what a generator can act on, and they belong on a cheat sheet:

• F-pattern. Eye anchors top-left, sweeps right, descends the left edge. The default for content-heavy pages.
• Z-pattern. Top-left to top-right, diagonal down, across to bottom-right. The default for image-heavy or low-text layouts.
• Focal-point quadrant. Say where the first-read element sits. Corner and off-axis placements break the centred default; centring surrenders to it.
• Scale ratio. The numeric relationship between levels. A model honours “3:1:0.6” more reliably than “make the heading bigger.”
• Level count. Keep it to three or four ranks in practice; the layout literature allows up to five. Past that, the eye gives up sorting and skims.

Some voices argue the models are getting better at hierarchy on their own, so the vocabulary will stop mattering. The strongest counter comes from the tools themselves. Recraft, whose own guide praises the model for treating typography as a structural component of the composition and keeping the layout consistent, still counsels describing the hierarchy explicitly. In its own words, it handles a layout better when the layout logic is described rather than left to the model to guess. The most layout-aware generators on the market are not telling you to stop specifying. They are telling you to specify well.

Hierarchy mistakes are among the ones AI catches on itself. Asked to review its own output, a model reliably flags surface problems, inconsistent spacing, mismatched sizes, a sub-head competing with its heading, and fixes them without being pointed at each one. In JM’s practice with the Zoom-In Method, that self-review pass catches roughly seven in ten of the surface mistakes in a draft. It is a working observation, not a measured benchmark, and the pattern matters more than the figure: a second pass asking the model to check its own ranking is close to free and usually worth it.

The other three in ten are the calls a designer keeps. Whether this element deserves to be primary for this audience is strategy, not surface. Whether the ranking matches the question the reader is actually asking is judgement. Whether an F-pattern even serves this reader, or whether the page wants a different scan entirely, is a decision the model cannot make from inside the file. Self-review tidies the hierarchy. It does not decide what the hierarchy should be.

Hierarchy ties to the other three fundamentals. Type is hierarchy enacted: every choice of scale, weight, and contrast sits inside a ranking, which is why Typography and AI and this article keep pointing at each other. Colour is the most legible hierarchy signal after scale, marking what leads and what recedes. Composition is the sibling question: hierarchy answers what is primary, composition answers where it sits, which is why the poster demo borrows from Composition and AI.

It maps onto the phases in Build Taste, Generate, Refine too. In Phase 1 you choose references for the hierarchy moves in them, not just the surface look. In Phase 2 the vocabulary becomes prompt constraints. In Phase 3 the refine pass catches hierarchy drift, while you keep the strategic calls. AI multiplies the skill, including this one: with hierarchy vocabulary it produces ranked, scannable work; without it, the even, centred default, faster than you could by hand.

AI does not rank a layout. It centres it, balances it, fills it, and stops, because the layouts it learned from skew symmetrical and even. The fix is not a better model. It is the vocabulary: the patterns that name an eye path, the quadrant that names a focal point, the ratio that names the levels, and the judgement to know which one a surface wants. Name the hierarchy and the model has something to build toward. Ask for “clear hierarchy” and you get evenly distributed content, ranked by nothing.

AI doesn’t rank the page. Designers who know hierarchy rank it with AI.

Cert IV in Design is where this vocabulary gets built, taught as the foundation before any AI integration, so a student can name a hierarchy long before asking a model to produce one. For working designers folding hierarchy and AI prompting into a daily practice, Design@Work carries it further.

Next: Composition and AI → (where what’s-primary meets where-it-sits). Also useful: Build Taste, Generate, Refine (this vocabulary inside a working AI workflow), Design Critique with AI (the structured self-review pass in full), and Typography and AI (the cluster mate, hierarchy enacted in type).