Dress Your Palette is a clothing recommender. The visitor gives their color season and body type, and the service shows clothes from large online shops whose color and lines suit them, with a reason for each one. This is the first part of the series: the starting point and the method. There is no revenue yet, and I say so plainly.
In short (as of November 6, 2026):
- What I built: a color pipeline that measures a garment’s color from the product photo, and 983 line rules for 13 body types.
- How fast: 9 days from the first line of code to launch (October 12 → October 21, 2026).
- What it earned: €0. Of the applications, 1 affiliate program approved, 10 rejected and 19 have not answered yet.
- What is still open: accuracy compared with a human judgment has not been measured yet, and the site loads slowly on mobile.
The problem: a shop’s color name does not describe the shade
In seasonal color analysis everyone has their own palette: dark, cool and rich shades for a deep winter, for example, and greyed, cool shades for a soft summer. There is plenty of advice, but it is hard to use in an online shop. Online shop filters usually do not know color seasons, and a color name such as “blue” or “green” says very little about the shade.
The same goes for lines. Body type rules say, for example, whether a soft or a sharp shoulder line suits, but shop filters do not know them. The idea of the service is to check color and lines on the same garment and to say which rule chose it.
The method
1. The color is measured from the photo
A segmentation model separates the garment from the product photo so that skin, hair and background are left out. Shoes and bags have their own mode, because they are usually photographed without a model.
The color of the separated garment is measured in two color spaces (OKLCH and CIELAB) and compared with 12 seasonal palettes using the ΔE2000 distance, which estimates how large a person perceives the difference between two colors to be. The shop’s color name is used only as a weak hint.
Three rules keep the result honest:
- Near-black and grey garments are treated as colorless, and they are not forced into a season.
- Patterns whose color cannot be measured as one shade are left out.
- If few garments fit, the list is not padded with weak matches.
For jewelry the color is not measured from the photo. Whether a piece suits a season is decided by its metal and stones: each palette records which metals suit it, and a stone is scored by its known color.
2. The lines are checked against rules
An image model picks out the garment’s features: silhouette, waist, shoulder line, material and details. They are scored against the rules of 13 body types.
The rules are based on David Kibbe’s body types. They cover the 13 types of his 1987 book. Kibbe later dropped three of them (Natural, Classic and Gamine) and now uses ten, but all 13 are included because many people were once typed as one of the dropped types. Following Kibbe’s current approach, line, silhouette and fit weigh the most, and the book’s garment lists are only light pointers.
There are 983 rules, compiled by me because no ready dataset exists. For 857 of them the principle comes from Kibbe’s book, 119 are marked as interpretations and 7 rest on other sources. Turning a principle into garment features is always partly interpretation, which is why the origin of every rule is recorded.
The first version of the scoring did not tell the types apart: the same garment got almost the same score for every type. The second version (October 18, 2026) compares the garment with a typical garment of the same category, and the differences appeared.
3. The measurement is repeatable
On November 4, 2026 I ran the same set of 50 test photos again and compared it with the October run. The measured color was the same for all 50 photos. The season scoring had been changed in between, and it changed the top season for 18 photos (16 of them neutrals). Repeatability is not accuracy, though: whether the service agrees with a human has not been measured yet. That is why the site shows no accuracy percentages.
The technology in brief
The interface is built with Next.js and runs on Netlify. The color pipeline is written in Python and stores the products and measurements in a Postgres database. The products shown on the site are published with the site as data files. The service has an English main version and a Finnish version, and the first region is the United States.
Speed is an open problem. On November 6, 2026 PageSpeed Insights gave the home page 65/100 on mobile, and the largest content appeared only after 17.3 seconds. It is worth fixing before expecting search traffic to grow.
Timeline
- October 11, 2026 Decision to build
First a tool for my own use, later a public service that earns affiliate commissions.
- October 12, 2026 First version of the color pipeline
Separating the garment from the photo and measuring its color against 12 palettes.
- October 21, 2026 Launch
dressyourpalette.com opened 9 days after the first line of code.
- October 23, 2026 Sitemap to Google
Guides to the 12 color seasons and 13 body types were published on October 22–23, and the first impressions came on October 23.
- October 31, 2026 First affiliate program
The program approved the site on October 29, and its jewelry was added to the service.
- November 5, 2026 Free search
Visitors can describe what they want in their own words.
Starting point on November 6, 2026
| Metric | Status | Source |
|---|---|---|
| Clicks from Google | 1 | Search Console, October 1 – November 3, 2026 |
| Impressions on Google | 331 | Search Console, October 1 – November 3, 2026 |
| Average position | 45.7 | Search Console, October 1 – November 3, 2026 |
| Pages in the sitemap | 55 | sitemap, November 6, 2026 |
| Revenue | €0 | as of November 6, 2026 |
| Affiliate programs | 1 approved, 10 rejected, 19 waiting | application log, November 6, 2026 |
The impressions come from searches about body types and seasonal palettes, but the positions are still mostly between 25 and 95. The site is two weeks old, so this is the expected starting point, not yet a result.
Why affiliate programs reject a new site
Affiliate networks usually review sites by hand, and some US networks may require existing traffic. A rejection has a cost: in one network a new application can be made only after 60 days, and in another a shop’s rejection is permanent for the same site. The service needs the shops’ product feeds so that its range can grow, and a feed is only available from an approved program. That is the current bottleneck of the service, not the technology.
It also shows in the revenue. On November 6, 2026 the site shows about 2,300 products from seven shops, but only 67 of them (about 3 %) sit behind an affiliate link, and almost all of those are jewelry from the approved program. The rest are picks from the shops’ own pages, and they earn no commission.
What the service cost
| Cost | Amount |
|---|---|
| Domain dressyourpalette.com | €12.90 a year |
| Hosting (Netlify free tier) | €0 |
| Database (Supabase free tier) | €0 |
| Newsletter (free tier) | €0 |
| Product search (110 searches in October, free tier) | €0 |
| Keyword data (DataForSEO) | €0.19 |
What is still open
- Accuracy. Before the service promises anything, it will be tested with clothes whose fit is known, and the color measurement will be compared with a human judgment.
- The shops’ range skews the results. The shops are selling their autumn collections, and the seasonal wardrobes are uneven: on November 6, 2026 the fullest (Deep Autumn) showed 665 products and the thinnest (True Spring and Bright Spring) 280 and 287.
- Some body types stay thin. For the weakest types the rules accept only 3.4–5.2 % of the candidate garments (October 31, 2026), so they need more shops.
- Speed. On mobile the largest content appears only after 17.3 seconds (LCP), and that has to be fixed.
What you can learn from this for your own business
- The image model reads, the rules decide. The image model picks out the features, but the choice rests on written rules whose origin is recorded. The same principle can be used in a company’s own automation.
- Measure before you promise. A repeatable measurement is not the same as a correct result. An accuracy claim waits for the test.
- A fast launch does not bring fast revenue. The service launched in 9 days, but revenue depends on partners who approve the site at their own pace.
In the next part
In the next part I will report how the product feeds and the waiting affiliate programs worked out, what the test of the color measurement against a human judgment showed and whether the first euro came in.
If you want AI in your own company’s workflow in the same way, see AI solutions for businesses.