[donatj]

I think this thing is just trained to like large hero images. Tried some of the most beautiful sites I know and got scores in the 10s and 20s, presumably from the presence of text.

Then I tried a number of sites with worthless huge hero images and got scores in the 80s.

I mean like all ai, take this with a grain of salt, but I think there is an obvious bias in this to the 'giant hero image that goes all the way to the fold' design that a lot of low effort start-up websites use.

[rikroots]

Agreed. I tested two of my personal websites[1][2]. The one with the text-heavy poetry scored 38% with a "meh" comment. The one with the gratuitously oversized banner image got 72% with a "oh yeah!" comment.

It does have an AI running the analysis. I like my poetry website. Maybe the AI was trained on good poetry and reacted poorly to my stuff?

[1] - https://rikverse2020.rikweb.org.uk/

[2] - https://scrawl-v8.rikweb.org.uk/

[etimberg]

Tested my own site as well that has no hero image and it got a 26. https://myraah.io/index.php/visualmind/report/aHR0cDovL2V2ZX...

[csande17]

I got 72% on my largely-text-only website, so either it likes white-on-black text or it's heavily overweighting the little photo of me in the top-left corner. https://myraah.io/index.php/visualmind/report/aHR0cDovL2Nhcn...

[hansvm]

If it likes hero images, the black background might look suspiciously like an image to the algorithm if they haven't explicitly done something to ensure similar results on color inversion.

[jszymborski]

This sounds about right. I got dinged a lot on my website for "Visual Clarity" although imho it's a pretty standard blog page with sidenotes, and the "similar scores" all had a decent amount of text.

Looking at the "similar scores" carousel for higher scores was primarily huge hero images with little text.

https://myraah.io/index.php/visualmind/report/aHR0cDovL2pzen...

[rasso]

My site [1] got a pretty good score, although it only consists of text if first visited. Large text, though...

[1] https://rassohilber.com

[Aachen]

I don't find your site's links clear at all by the way. It turns out that emojis are supposed to indicate actions/links, but until that's common you might want to give link text a slightly different color or maybe underline it.

[rasso]

Thanks for the feedback. On desktop, the links become orange if hovered, but on mobile, it's quite unobvious, you are right.

[Aachen]

Not sure I hover over what seems to be regular text on a typical website. On desktop you need to put your cursor on it and on mobile your finger, not sure it's that different.

[kroltan]

My website's homepage which consists of a 128x128 picture of me, text and a table got 54. I guess the mere presence of a image made the AI like it a bit more.

[myraahio]

Hi, Thanks for your feedback you are right. Currently algo evaluate how a user will feel...when they look at your site above the fold...it's like tinder you just look at the pic and decide to swap right or left.. this is based on : Users make lasting judgments about a website’s appeal within a split second. This first impression is influential enough to later affect their opinions of a site’s usability and trustworthiness...of course lot of areas we can improve...hope u enjoyed..

[Aeolun]

But do users really enjoy full size hero images?

[myraahio]

That's the reason why program says it's good..

[lionkor]

I think your project will not be taken seriously by the target audience if you imply that the AI isnt extremely biased, as any "pretty or not" judgement.

Providing a "what people might like better statistically" service is interesting, claiming you have a "the AI knows how the user will feel" algorithm is not a good strategy

[KaoruAoiShiho]

Is it an actual AI though, trained through objective data or some kind of human programmed analysis?

[myraahio]

It's a deep learning model..hence sometimes illogical as well :). Thanks for trying...it will learn and improve over time...

[aavshr]

By improving over time do you mean training the model for more epochs or having some kind of feedback loop that tells the algorithm whether the score was accurate or not based on human input?

I think the later makes more sense. Some kind of input from the users like "I actually liked the website visually" when the algorithm gives a low score and vice-versa would help the model be more accurate?

[KaoruAoiShiho]

What kind of data are you inputting and is it objective? You know the truism in ML... garbage in garbage out.