Leaderboard › Methodology › Technical blueprint
Technical blueprint · version 1.3 · 16 Nov 2026
The blueprint: everything needed to rebuild this test
This document sets out what has to be established about each search site every week, the rule for each check, and the arithmetic that turns the checks into a score. It describes outcomes, not tooling: how you load a page or look up a profile is your choice, so long as you end up with the same facts. Someone who can program and can collect the pages has what they need here to run the test and compare their numbers with mine.
1Which sites are tested
- The list
- The 25 OnlyFans search sites with the most visitors. Nothing else decides who is in.
- Source of the figure
- Similarweb’s estimate of monthly visits for each domain. The figure is used only to choose the sites. It is never part of a score.
- What counts as a search site
- A site where a visitor can look for OnlyFans creators by search, place or tag and is shown a list of them. Lists of links, forums and single-creator pages do not count.
- Changes
- Estimates are refreshed monthly. A site that moves into the top 25 replaces the smallest site in the test from the following weekly run. The sites waiting are listed on Sites I don’t test yet.
2The week’s searches
One run a week, starting on Monday. Every site gets the same searches in the same run.
- Home page
- The page a visitor lands on, as served.
- A country
- Drawn at random from a pool of 12 countries.
- A large city
- In that country. The pool holds 20 country, city and town sets.
- A small town
- In the same country, with roughly 5,000 to 30,000 residents, and never a suburb of the week’s city.
- Two tags
- Drawn from 20 tags that can be judged from a photograph: hair colours, ethnicities, body types, and visible things such as tattoos, glasses or cosplay.
- The draw
- Made when the run starts. No country, city, town or tag used in the previous 28 days can be drawn again. The picks are published with that week’s results and not before, so no site can prepare for them.
- Combined filter search
- The same every week: Sydney, blonde, young, female, cosplay, applied together.
- Image search
- One creator’s profile photo, drawn at random each week (section 12).
- Where the visitor is
- Location searches are run as a home-broadband visitor in the country being searched. The home page and tag searches are run from the UK. Where a site refuses that country, the country actually used is recorded with the result.
The first run, on 14 November 2026, used fixed searches (England, London, Sedona in Arizona, Asian, red hair). Random weekly searches apply from the run of 16 November 2026.
3Finding each page on a site
Use the page a normal visitor would reach by clicking through the site. Work down each list and stop at the first that exists.
- Country and city
- The site’s own page for the place. Failing that, the site’s search for the place name. Failing that, the search is recorded as not available on this site.
- Small town
- The site’s own page for the town. Failing that, its page for the nearest real city or town in the same country within about 200 km. Failing that, its search for the town’s name. A page for a whole state, region or country is never accepted in place of a town.
- Tag
- The site’s own tag or category with the same meaning. Failing that, its search for the word.
- What is not a search
- An address that returns the same list whatever word is put into it. Test it with a nonsense word: if the list does not change, the site has no search for that term.
- Not available
- A place the site cannot show counts against it (section 7). A page whose address I could not establish myself is “not tested” and never counts against the site.
4Recording a results page
- What to record
- Every listing, in the order the page shows it: position, platform, username, the address the listing links to, whether it is a paid placement, and whether it is visibly labelled as one.
- How far down
- Keep going through “load more”, scrolling or next-page links until about 230 listings are recorded or the site has no more to show. The aim is 200 unpaid listings per search.
- Usernames
- Where a listing links to the site’s own page for the creator, open that page to read the OnlyFans username. A listing with no username that can be established is recorded but cannot be checked.
- Ad links
- Never follow a paid link as a visitor would; that would charge the advertiser for a click. Read the username from the listing’s own data.
- The two samples
- The first 200 unpaid listings with a username are the sample for the creator, location and tag checks. The first 50 listings of any kind are the sample for ad density.
5Paid placements
- What is a paid placement
- A listing whose link passes through a tracking, campaign or redirect address in place of going straight to the creator’s profile (OnlyFans campaign links of the
/c15kind, advertising-network redirects, a site’s own/out/or/goroutes), or a listing the site’s own page data marks as paid. - What is a label
- Visible words on the listing itself: Ad, Sponsored, Promoted or Featured. A label that only appears in the page’s code, or on a different part of the page, does not count.
- Ads labelled
- Across every page recorded for the site (home, the five searches and the filter search): labelled paid placements ÷ paid placements whose label could be checked × 100.
- Ad density
- For each of the five test searches, the percentage of the first 50 listings that are paid. Average those, and subtract from 100.
6Live and active creators
Each creator in the 200-listing sample is looked up on OnlyFans after the results page has been recorded, never from an older copy.
- Live
- The profile exists.
- Dead
- The profile is gone (deleted, banned or renamed), or no profile was ever found under that username.
- Active
- Live, and either the profile’s “last seen” is within the past 14 days, or its count of posts, photos or videos has grown within the past 14 days.
- Cannot be judged
- Live, but last seen is hidden and there is no earlier count to compare with. From the run of 16 November 2026 these creators are left out of the active percentage. The first run counted them as not active.
- Unchecked
- A lookup that has not succeeded within 72 hours. Left out of every percentage. Creators on platforms other than OnlyFans are recorded but not judged.
- Per search
- Live % = live ÷ checked. Dead % = dead ÷ checked. Active % = active ÷ creators that could be judged. The date of each lookup is kept.
- Per site
- The average of each percentage over the test searches that had checked results.
The count comparison needs history. A rebuild starting from nothing can only use “last seen” until it holds counts at least 14 days old, and will judge fewer creators than I do in its first two weeks.
7Location accuracy
The creator’s location is the one on their own OnlyFans profile, matched to a real place. What a search site says about the creator plays no part.
- Three outcomes
- In the place searched; somewhere else; or not disclosed.
- Not disclosed
- No location, text that names no place, or only a wider area than the one searched (a country when a city was searched). Left out for every site.
- Country fit, city fit
- In the place ÷ (in the place + somewhere else) × 100, over the 200-listing sample.
- Small town, precision
- Take the first 20 results that can be placed. In the town: 100. Within about 30 miles (50 km): 75. Elsewhere in the same state or region: 40. Further away: 0. Average them.
- Small town, coverage
- Distinct creators in the town that this site showed ÷ distinct creators in the town shown by any tested site that week × 100.
- Small-town score
- The average of precision and coverage.
- Location accuracy
- The average of country fit, city fit and the small-town score. A place the site cannot show at all enters as 0. A town page with nothing placeable enters as 0 when another site did find creators in that town, and is left out when none did.
8Tag accuracy
- What is judged
- The profile photo of each creator in the sample on the two tag searches, and nothing else: not the bio, not the site’s own tags.
- Who judges
- An image model, asked one fixed yes-or-no question per tag with randomness switched off. I use Qwen2.5-VL (7B). The week’s questions are published with the results.
- Answers
- Yes, no or unsure. Photos showing no person are left out. For hair and body tags, a photo where the hair is covered or the body is out of frame is left out too.
- Tag fit
- Yes ÷ (yes + no + unsure) × 100 for each tag search. A search needs at least 10 judged photos to count. The site’s figure is the average of its tag searches.
9Combined filters
- The search
- Sydney, blonde, young, female, cosplay, all applied in one search through the site’s own filter controls. Typing the words into a search box does not count as filtering.
- Per criterion
- 1 if a visitor can apply it and the results change when it is applied. ½ if only a weaker version exists, such as a country where a city was wanted. 0 if the control is missing or changes nothing.
- Score
- 20 × the total, so five working criteria score 100.
- Also recorded
- The number of results the site claims, and the listings actually shown, which go through the same paid-placement and creator checks as any other page.
10Features
A feature earns its point only if a visitor can actually use it. One that is present but limited or partly broken earns half. For each feature the address and what was seen are written down.
| Feature | What earns it | Points |
|---|---|---|
| Near-me search | Finds creators close to the visitor, by detected location or a chosen distance, and returns local results | 1 |
| Image search | A photo can be uploaded and results come back | 1 |
| Location search | Creators can be searched or browsed by country, region or city | 1 |
| Map search | Creators can be browsed on a map and a place picked from it | 1 |
| Tag search | Creators can be searched or browsed by tag or category | 1 |
| Filters that combine | Two or more filters can be applied to the same search | 1 |
| Sentence search | A plain sentence such as “blonde cosplayers in Sydney” returns results that match its meaning, not only its words | 1 |
| Free-trial filter | Results can be limited to free accounts or free trials | 1 |
| Gender search | One point for each that can be searched: women, men, trans creators, couples | up to 4 |
| Trans women and trans men | The two can be searched separately | 1 |
| Couples search | Separate couple types, such as woman and man, two women, two men. A single “couples” option earns half | 1 or ½ |
| Sort options | Three or more ways to order the results. One or two earn half | 1 or ½ |
| Result layouts | Two or more ways to view the results, such as grid and list | 1 |
| Bookmarks | Creators can be saved to come back to | 1 |
| Detail levels | More than one level of information about a creator, from the card up to fuller statistics | 1 |
| Recent changes | Shows what changed lately for a creator: new photos or videos, a price change, last seen | 1 |
| History over time | Shows a creator’s numbers by month, such as likes, posts or price | 1 |
| AI chat | An assistant that can be asked for recommendations | 1 |
| Own creator pages | The site has its own page for each creator and does not only link out | 1 |
| Platforms | Two or more creator platforms are indexed | 1 |
| Languages | The site is available in more than one language | 1 |
| No account needed | Search and creator details work without signing up | 1 |
| Dark mode | A dark background can be chosen, or the site follows the device setting | 1 |
| Works on a phone | Usable on a screen 390 pixels wide | 1 |
| Pop-ups | Intrusive pop-ups, pop-unders, full-screen interruptions or redirects on arrival. An age check is never counted: every adult site should have one | minus 1 |
27 points are available. Features score = points ÷ 27 × 100. Couples count twice on purpose: once as a gender that can be searched, and once for how precisely couples can be searched.
11Speed
- Search speed, what to time
- Each test search, each of the five filters on its own, and the combined filter search. Three runs each, in a fresh desktop browser with nothing cached.
- Search speed, what to record
- Two times per run: from the visitor’s action to the first result card being visible, and the response time of the request that carries the results. Take the median of each.
- Which one is scored
- The response time, because it reflects the site and not the length of my connection. The time to the first result card is the figure shown on each site’s page.
- Search speed score
- 100 − (milliseconds − 200) ÷ 18, held between 0 and 100.
- Page speed
- Google Lighthouse performance score for each test page, once as a phone and once as a desktop, nothing cached. Take the median per device. Runs stopped by a bot check are discarded, and at least three phone results are needed.
- Page speed score
- 0.6 × phone + 0.4 × desktop.
12Image search
- The photo
- The current public profile photo of one creator, drawn at random each week from creators with a live profile, activity in the last 14 days, between 20,000 and 1.5 million likes, and a photo showing exactly one face at least 15% as wide as the photo’s shorter side.
- The test
- Upload that one file to every tested site that offers search by photo. Record the first 20 results in order, with paid placements marked.
- Found
- A result counts as the creator only when its username is the creator’s own. A similar face under another name does not count.
- Image search works
- 100 if the site has an image search and it returned results. 0 if it has none or it failed.
- Image search results
- Result 1: 100. Results 2 to 3: 75. Results 4 to 10: 50. Results 11 to 20: 25. Not in the first 20: 0. Sites without an image search are not given this score.
- How it is built
- Reported, never scored, and judged only from what any visitor’s browser can see during the upload: the address the photo is sent to, whether the answer arrives in one reply or through a job that the page keeps checking, the time from upload to reply, and what the reply contains (similarity figures, face or attribute details, any processing time the site itself states). No information from inside a site is used, for any site.
- Time to results
- From submitting the photo to the first result card being visible, in the same browser and on the same connection for every site.
- Timing
- Both parts enter the score from the run of 16 November 2026. An image test more than six days from the run is not used.
The tools that examine the photos, and the separate lookalike test that is reported but not scored, are described on the image search test page.
13Results and access
- Test searches with results
- Count the test searches that produced at least 20 checked unpaid results. Divide by the number of test searches I was able to run on the site, and multiply by 100.
- UK access
- Open the site ten times as a UK home-broadband visitor. If most of those visits are redirected away or refused, the site scores 0. Otherwise 100.
14The score
Each component is a number from 0 to 100. The score is their weighted average: multiply each component by its weight, add them up, and divide by the sum of the weights used. It is rounded to one decimal place, and sites are ranked from highest to lowest.
| Component | Weight | What goes in | How it becomes 0 to 100 |
|---|---|---|---|
| Fresh results | 20 | Share of checked unpaid results whose creator was active in the last 14 days | The percentage itself. 0 if the site returned no unpaid results on any test search. |
| Location accuracy | 15 | Country fit, city fit and the small-town score | The average of the three. A place the site cannot show at all counts as 0. |
| Tag accuracy | 15 | Share of profile photos on the two tag searches that match the tag | The average of the two tag percentages. |
| Combined filters | 15 | How many of the five filter criteria can be applied and really narrow the results | 20 points per working criterion, 10 for a partial one. |
| Ads labelled | 15 | Share of paid placements that carry a visible label | The percentage itself. 100 if the site shows results and I found no paid placements. |
| Features | 15 | Points earned on the feature list, out of 27 | Points ÷ 27 × 100, never below 0. |
| Live profiles | 10 | Share of checked unpaid results whose profile no longer exists | 100 minus the dead percentage. |
| Ad density | 10 | Share of the first 50 listings that are paid, averaged over the five test searches | 100 minus that average. |
| Test searches with results | 10 | Test searches that returned at least 20 checked unpaid results | Searches with 20 or more ÷ searches I was able to run × 100. |
| UK access | 10 | Whether the site opens for a visitor on UK home broadband | 100 if it opens, 0 if most UK visitors are turned away. |
| Page speed | 10 | Lighthouse performance score, phone and desktop, on the test pages | 0.6 × phone median + 0.4 × desktop median. |
| Search speed | 5 | Median response time of the request that returns the search results | 200 ms or faster scores 100, 2,000 ms or slower scores 0, a straight line between. |
| Image search resultsfrom 16 Nov 2026 | 10 | Position of the right creator when their own profile photo is uploaded | Result 1: 100. Top 3: 75. Top 10: 50. Top 20: 25. Not found: 0. Only for sites with an image search. |
| Image search worksfrom 16 Nov 2026 | 5 | Whether the site has an image search that returned results that week | 100 if yes, 0 if there is none or it failed. |
The weights add up to 150 for the run of 14 November 2026 and to 165 from the run of 16 November 2026, when the two image components join. A component that could not be measured for a site that week is left out, and that site’s score is the weighted average of the rest. Weights are fixed before a run and never changed for a run already published.
15When my own check fails
- A page that comes back short
- If a recorded page has fewer than half the listings the same search returned before (and that earlier count was 20 or more), record it again, up to twice. If it is still short, mark it “capture incomplete” and leave it out of that week’s score.
- A page I could not locate
- Marked “not tested”. It is not counted as a search without results.
- Corrections
- Each tested site has a right of reply on its own page, and a result shown to be wrong is corrected with the date of the correction.
16Rebuilding it yourself
Run on the same day, with the same searches, a rebuild that follows this document should arrive at my numbers. Four things have to match.
- The week’s picks
- Take the country, city, town, tags and image-search creator from my published results for that week. They are random, so you cannot derive them.
- The day
- Search sites reorder their results constantly and creators come and go. Record the pages on my run date and look up the creators straight afterwards.
- Creator history
- The 14-day activity rule needs earlier post, photo and video counts (section 6).
- The image model
- Tag accuracy depends on the model named in section 8. A different model will give slightly different answers on borderline photos.
If you rebuild it and your numbers differ from mine, tell me. Either I have described something badly or one of us has made a mistake, and I would rather know. Changes to this document are versioned: this is version 1.3, issued 16 Nov 2026.