How to Audit Your Own Site for the AI Look

Score it, do not squint at it

Self-assessment fails here for a reason that has nothing to do with taste. You chose every color on that page, so none of them surprise you. You wrote the headline, so you read the meaning you intended rather than the words sitting there. By the fourth pass your site has stopped being a page and become a memory of a page.

So the audit is built to take judgment out of it. Each check is a thing you count, a string you search the stylesheet for, or a test with a stated pass condition. You could hand the list to a stranger and get roughly the same number back, which is the only property that makes a checklist worth anything.

One note on scope. This measures how default your site reads, not whether building it with AI was a mistake. I built this site with Claude and I score it against the same ten checks. If you want the fixes written up rather than the diagnosis, the six tells and what to do about each one is the other half of this.

What to have open before you start

Three things: your site in a desktop browser with developer tools open, your site on your own phone, and a blank note with ten numbered rows.

Scoring is coarse on purpose. Two points if a check passes cleanly, one if it half passes, zero if it fails. Coarse bands stop you negotiating with yourself over half a point, which is how self-audits turn into self-congratulation.

Finish the whole pass before you change anything. The urge to stop at the first check and fix the font is strong, and it produces an audit you never complete attached to a site with a new font and all the original problems.

Run it on your homepage and on one inner page. Inner pages almost always score worse, because attention goes to the front door and the about page gets whatever the model produced.

1 to 3: the visual defaults

1. The type. In developer tools, select your main headline and read the computed font-family in the styles pane, which Chrome documents in its CSS reference. Full marks if headline and body are set in two different families and you can explain why one of them is there. One point if you deliberately chose a single family and separated levels with weight and size. Zero if the answer is a system sans stack you never touched.

2. The color. Search your stylesheet for gradient( and count the hits, then count your distinct accent colors. Full marks if no gradient is doing structural work and you have one accent plus neutrals. Two accents are fine when the second one has a job. Zero if the hero is a gradient and the accent sits somewhere between violet and cyan.

3. The rhythm. Count your sections, then count how many are a heading followed by three equal boxes. Full marks if there is at most one group of three and at least one section whose shape nothing else on the page repeats: a long passage with a single image, a table, one pulled quote, a bare list. Zero if every band is the same card grid with different icons in it.

Fifteen minutes for all three. These are also the ones a stranger processes before reading a single word, so a low score here colors everything that follows.

4 to 6: the copy

4. The swap test. Paste your first fifty words into a note and replace your business name with a competitor's. Read it back. You pass if the paragraph is now false or faintly ridiculous. Zero if it still reads as true, because copy that fits any company in your category is describing none of them.

5. The proper nouns. In the first screen of text, count the proper nouns, prices, place names, dates, and named tools. Three or more passes. Mine are the tool names and the price. Zero is the usual result on a generated page, and it explains why clean sentences can still feel airless.

6. The breath test. Read one paragraph out loud and notice where you run out of air. You pass if that paragraph contains at least one sentence under eight words. Generated prose settles at a uniform medium length, and uniformity is audible well before anyone can describe it.

Nearly everyone runs the copy checks second, me included, and they matter more than the three above them. Looking at design is easier than reading your own writing. I went at the sentence-level work on its own in website copy that converts.

7 and 8: images and evidence

7. Provenance. List every image on the page and write where each one came from. You pass if you can name a source for all of them and at least one is a photo you took, a screenshot of your own work, or something you drew. Zero if the list is three stock photos and an abstract shape.

Screenshots are the underrated option here. Nobody can fake a capture of your actual dashboard, your actual spreadsheet, your actual bench, and a real one carries specificity no illustration will.

8. Evidence. Count the claims on the page, then count how many link to anything outside your own domain. You pass if every figure and every assertion about the wider world points somewhere a reader can go and check. Zero for a page of confident statements with no exits, which is the characteristic shape of generated marketing text.

9 and 10: the parts nobody prompts for

9. The edges. Hover every button. Tab through the page using only the keyboard. Visit a URL on your domain that does not exist. Then look at the browser tab. You pass if buttons answer the cursor, keyboard focus is visible, the missing-page view is yours, and the tab carries your favicon beside a title that is not the template default.

Those four never make it into a prompt, so they rarely make it into the output. That is exactly what makes them a reliable signal, and all four are an afternoon of work to fix.

10. The real device. Open the site on your own phone, outside if you can, and try to finish the one action the page exists for. You pass if nothing slides sideways, the tap targets suit a thumb, and you complete the action without zooming. The browser's device simulator does not count toward this one. It is wrong about font rendering and it knows nothing about how far your thumb reaches.

Adding it up

Total the ten rows out of twenty.

Sixteen and over: the site reads as the work of someone who was paying attention. Keep the list for the next page you add, since new pages drift toward the default.

Thirteen to fifteen: nobody calls it generated, and nothing about it stays with them either. The missing points are usually in copy and images rather than layout.

Nine to twelve: it reads as a template with your name on it, and some of your bounce rate is a design problem. Two or three changes move this band a long way.

Eight or under: the page is a default. Worth saying plainly that this is an afternoon of work, not a verdict on you or on the tool you used.

Fix in this order

Ranked by signal gained per hour spent, which is not the order most people pick.

Headline and first fifty words, rewritten until they fail the swap test. One hour, and the biggest single move available to you.

One real image in place of one stock image. Twenty minutes.

Hover states, focus ring, favicon, missing-page view, page title. An afternoon, and it clears four of the cheapest signals at once.

Type pairing next, including a reading pass on a phone. Half a day.

Layout rhythm last. It is the most work and the least noticed of the five, and it is the one people reach for first.

When the score does not matter

If your site sees a handful of visitors a week, this audit is not your constraint. A clean twenty on a page nobody loads changes nothing, and I have watched people redesign their way out of doing distribution for months. Traffic first, then this. I covered that separately in getting traffic to a new website.

The other exemption is a page that already converts. If people arrive and buy, a low score is a curiosity rather than a problem, and you should spend the afternoon on the next page instead. What visitors conclude about a site is a separate question from what the checks measure, and I went into that in whether people can tell.

So: does your site look AI generated? Go and score it. Ten checks, twenty points, an hour of your morning, and at the end you have a number and a short ordered list instead of a vague worry. Most first passes land between nine and fifteen. Mine did. The gap closes quicker than people expect, because the cheapest checks on the list are precisely the ones no model was ever asked to handle.

Related: How to Write AI Copy That Doesn't Sound Like AI · How to Make Claude Check Its Own Work on a Website

Common questions

Does my website look AI generated?

Run the ten checks in this post instead of guessing, because you have seen your own site too often to judge it by eye. Score type, color, layout rhythm, copy specificity, image provenance, evidence links, interface details, and phone behaviour out of twenty. Twelve or lower reads as generated to a first-time visitor.

How can I tell if my website looks like a template?

Count how many sections are a heading followed by three equal boxes, then look for one section with a shape nothing else on the page repeats. A page where every band is the same card grid reads as a template even when all the colors are yours.

What is the single fastest fix if my site looks AI built?

Rewrite the first fifty words until they fail the swap test, meaning they stop being true the moment you paste a competitor's name in place of yours. That is about an hour of work and it shifts more than any visual change you could make in the same time.

Do AI detectors work on websites?

Not usefully for design, and the person deciding whether to trust you is not running one. The practical test is the audit itself: specific copy, images you can source, interface states that respond, and a page that works on a real phone rather than in a simulator.

Should I run this audit on every page?

Start with the homepage and one inner page, since inner pages usually score lower. After that, run the ten checks on each new page before publishing, because new pages slide back toward the default faster than old pages get repaired.