How can you tell a website was designed by AI?

What are the most reliable surface-level tells?

The em-dash is the clearest single signal. Human writers use it occasionally for emphasis or interruption. AI language models reach for it constantly — often several times per paragraph — because it bridges two ideas without requiring a proper connective. In body copy, a high density of em-dashes is a mechanical pattern, not a stylistic choice.

Filler marketing verbs are the second tell: words like 'streamline', 'empower', 'leverage', 'elevate' and 'transform' that promise motion without describing any. They appear because they are statistically common in the training data for marketing copy, not because they communicate anything specific. A page heavy with them tends to say a great deal about aspiration and nothing about mechanism.

Placeholder names that survived the build are a third category: team members called 'Alex Johnson' or 'Sarah Williams', testimonials attributed to 'CEO, Tech Company', or pricing tiers labelled 'Starter / Pro / Enterprise' with no actual prices filled in. These are scaffolding left standing after the scaffolding was supposed to come down.

Do visual patterns give the same kind of evidence?

They do, though they require a slightly different eye. The stock purple gradient — a violet-to-indigo sweep across a hero section or card background — became the default aesthetic for AI-generated SaaS pages because it was the dominant look in the design examples those tools were trained on. It is not that purple gradients are wrong; it is that this specific hue range appears with a frequency that outstrips any organic design trend.

Accompanying it are a set of layout habits: hero sections with a large centred headline, a subheading in grey, and a pair of buttons (one filled, one outlined) sitting above a device mockup. The mockup itself often shows a dashboard with circular progress indicators and a sidebar, regardless of whether the product has a dashboard. These are not aesthetic failures — they are statistical averages of what 'a SaaS website' looked like in the training data.

Icon sets also cluster. AI-built pages tend to use the same small library of line icons — a lightning bolt for speed, a shield for security, a chart for growth — arranged in three or four feature cards below the fold. The icons are fine individually; the combination and placement are a fingerprint.

Why does counting these patterns beat asking a model for its opinion?

When you ask an AI model whether a piece of writing or a design looks AI-generated, you are asking it to assess the statistical properties of its own output against a distribution it cannot fully observe from the inside. The model has no reliable ground truth for what 'human-made' looks like versus what it tends to produce. Its answer is itself a piece of generated text, subject to the same biases.

A count of specific, pre-defined signals is different. It does not require a judgement call. Either the em-dash appears more than a threshold number of times per thousand words, or it does not. Either the phrase 'streamline your workflow' is present, or it is not. Either the colour value of the hero background falls within the flagged gradient range, or it does not. These are facts about the artefact, not opinions about it.

This matters practically because a count can be audited, disputed and improved. If a signal turns out to be a poor predictor — perhaps a human copywriter also loves em-dashes — it can be removed from the set. An opinion cannot be audited in the same way. GROX's site check applies this principle: it counts the measurable tells of AI-generated design as counted facts alongside standard SEO checks, rather than asking a model to form a view.

Counted fact

A signal whose presence or absence can be verified by anyone inspecting the same artefact — em-dash frequency, a specific colour value, a placeholder string.

Model opinion

A generated assessment of whether something 'feels' AI-made, produced by the same class of system being assessed and not independently verifiable.

False positive

A counted tell that appears in genuinely human-made work — a reason to weight signals carefully and update the set when evidence warrants it.

Scaffolding artefact

Content or structure inserted during generation as a placeholder, intended to be replaced, that survived into the published page.

When is a simpler check good enough?

If you are reviewing your own site before launch, a manual pass looking for the patterns above is often sufficient. Read the copy aloud and count the em-dashes. Search the page source for 'lorem ipsum', 'placeholder' and common placeholder names. Load the hero section in a colour picker and note the hex values. These steps take minutes and catch the most glaring tells without any tooling.

A structured audit becomes worth the effort when the site is large, when it was built by a contractor whose process you cannot inspect, or when the stakes of being publicly identified as an AI-generated property are meaningful — a recruitment site, a financial services page, a publication that claims editorial standards. In those cases, the consistency of a defined signal set matters more than the speed of a manual scan.

It is also worth noting that none of these tells are fatal on their own. A single em-dash is not evidence of anything. A purple gradient can be a deliberate brand choice. The signal is in the combination and density, not in any individual element. Treat the count as a prompt to look more carefully, not as a verdict.