Features
Reading fame rankings with a sceptical eye
Fame rankings order a proxy, never fame: what each signal literally counts, the distortions common to all of them, and why most of a list is noise.
Every fame ranking is a ranking of a proxy. There is no fame meter. Someone picked a countable signal, sorted by it, and published the order. The interesting question about any such list is not who came first but which signal was counted and what that signal is made of.
This page is about that question, and about being straight on where these datasets are useful and where they are not.
What to take away
- Importance, merit, achievement, or influence.
- The obvious next move with a fame dataset is to group by country, region, or community and count.
- Fame data is good evidence about attention and about the reference works that record it, and bad evidence about everything people want to use it for.
The proxies, and what each one actually counts
| Signal | What it literally measures | Built-in distortion |
|---|---|---|
| Encyclopedia pageviews | Visits to one page, in one language, over a window | Spikes on news, deaths, anniversaries, and unrelated same-name events |
| Search volume | Queries typed into one engine, by that engine's users | Ambiguous names absorb unrelated demand; engine market share varies by country |
| Number of language editions | How many language communities wrote an article | Depends on editor numbers per language, not on the subject |
| Article length or edit count | Effort spent by volunteers | Rewards contested topics and active fandoms |
| Inbound links | How many pages point at this one | Reflects site structure and templates as much as interest |
| News mentions | Occurrences in an indexed press corpus | Corpus coverage is skewed by language, country, and paywalls |
| Book-corpus frequency | Occurrences in a scanned book collection | Reflects what was scanned and what was published |
| Follower counts | Accounts subscribed on one platform | Platform demographics, account age, purchased and inactive followers |
| Recognition surveys | Share of a sample who recognize a name | Sample frame, and whether recall was prompted or unprompted |
Each of these is a real measurement of a real thing. None of them is a measurement of fame, and none of them is a measurement of importance. When a list is published without naming its signal, you cannot evaluate it at all, and the correct response is to treat the ordering as decorative.
Fame proxies and their distortions
Signal
- Encyclopedia pageviews
- Visits to one page
- Search volume
- Queries on one engine
- Language editions
- Communities with an article
- Article length
- Volunteer effort spent
- Inbound links
- Pages pointing here
- Follower counts
- Subscribed accounts
What it counts
- Encyclopedia pageviews
- News, deaths, anniversaries
- Search volume
- Ambiguous names, market share
- Language editions
- Editor numbers per language
- Article length
- Rewards contested topics
- Inbound links
- Site structure and templates
- Follower counts
- Purchased and inactive followers
Built-in distortion
- Encyclopedia pageviews
- Search volume
- Language editions
- Article length
- Inbound links
- Follower counts
Distortions that hit all of them at once
Language. Almost every one of these signals is language-scoped or language-weighted. A ranking derived from English-language material ranks visibility in English, and a name that crossed a script splits across several forms of romanization besides. The gap between that and global prominence is enormous and is never a constant offset.
Recency and the living. Digital signals began when the platforms did. Anyone whose peak visibility predates the measurement window is systematically undercounted, and the living generate ongoing traffic that the dead do not. Any list mixing living and historical figures on a digital signal is comparing two different things.
Digitization. Book, newspaper, and archive corpora contain what somebody paid to scan. Coverage varies by country, period, language, script, and copyright status, and the gaps are not random. The custody side of that is set out on the US state pillar, and the period side on the century pillar.
Name ambiguity. A person sharing a name with a company, a place, a film, or a more prominent namesake inherits their traffic. Conversely, a person known under several forms of a name has their signal split across all of them. Neither error is small, and neither is visible in the output.
Events. Deaths, trials, awards, and anniversaries produce spikes that dominate any short measurement window. A ranking is always partly a report on what happened during its measurement period.
Automated traffic. Some proportion of counted requests is not people. Filtering methods vary between publishers and are rarely described.
Structural artifacts. Redirects, disambiguation pages, template links, and how a platform canonicalizes a name all move counts around for reasons that have nothing to do with the subject.
The feedback loop
Fame metrics are reflexive. Appearing on a ranking generates coverage; coverage generates searches and pageviews; those feed the next measurement. A list published annually is partly measuring the effect of its own previous edition, which is the Matthew effect with a publication schedule.
The same loop runs through the reference works these lists draw on. A person becomes prominent enough to get a well-maintained article; that article ranks better and is easier to cite; the citations reinforce the prominence.
Nothing improper is happening. The metric is not independent of the thing being ranked, so it cannot be used as evidence about it.
Rank numbers imply a precision that is not there
Three specific problems, all avoidable and all routine:
Rank gaps are not equal
- Top two placesgap can exceed places 20 to 200
- Below the topadjacent entries differ by less than metric variation
- Re-run a month laterlarge stretches reorder
The intervals are not equal. Most fame signals are extremely skewed: a small number of subjects account for most of the total, and the rest sit close together. The gap between the top two places can exceed the gap between places twenty and two hundred. Presenting them as evenly spaced ranks hides the entire shape of the data.
Most of the ordering is noise. Below the top of such a distribution, adjacent entries differ by less than the metric's own variation between measurement windows. Re-run the count a month later and large stretches reorder. Publishing a specific rank for such an entry states a precision the data does not contain.
Nobody publishes uncertainty. These lists appear as integers with no error bars, no measurement window stated, and no note on how ties were broken. All three are knowable and all three are usually omitted.
If you build a list, present bands rather than positions, state the window, state the tie rule, and say how much of the ordering is stable across windows. If you are reading someone else's list and it does none of that, treat the top few entries as informative and the rest as approximately unordered.
What this data can support
Being frank about the useful half:
- Attention over time for one subject.A time series of one person's pageviews or mentions, with events annotated, is genuine evidence about when interest rose and fell.
- Comparisons within a single signal, language, and window.Like-for-like comparison is defensible if you say it is like-for-like.
- Detecting events.Spikes reliably indicate that something happened, which is useful even before you know what.
- Coverage analysis of a reference work.These signals are excellent evidence about the reference work itself: what it covers well, what it neglects, how its attention is distributed.
- Disambiguation warnings.Unexpected traffic patterns are a good indicator that a name is being confused with something else.
What it cannot support
- Importance, merit, achievement, or influence. Attention is not a measure of any of them, and the influence page covers why that claim needs different evidence entirely.
- Cross-language or cross-era comparison on a digital signal.
- Any statement about a person's character, standing among peers, or historical significance.
- Any per-population or per-group tally. This is the important one, and it gets its own section.
Why there are no ranked lists by origin here
The obvious next move with a fame dataset is to group by country, region, or community and count. This site does not do it, for two separate reasons.
The first is technical and, on its own, decisive. Grouping requires an origin field for every person, and that field is one of the least reliable in any biographical dataset: as the country and heritage pages set out at length.
Crossing an unreliable grouping variable with a metric dominated by language and digitization coverage makes output order depend almost entirely on which archives were scanned. It also depends on the corpus language. It measures infrastructure, though it looks like it measures people.
The same infrastructure problem runs through notability maps and data, where every step from a birthplace field to a dot on a map leaks.
Second, the framing is wrong even when the arithmetic works. "Which country produced the most" scores populations against each other using the achievements of individuals, and they did not consent to being tallied that way.
We are not interested in supplying that answer, and we do not think a better dataset would make the question a good one.
Reading someone else's ranking
Ask these before quoting it:
- What is the signal, exactly, and is it named in the methodology?
- What language, platform, and corpus does it cover?
- What is the measurement window, and does it include a major event?
- Are living and historical figures mixed on a digital signal?
- How were names disambiguated?
- Was automated traffic filtered, and how?
- Who chose the candidate pool, and was the pool itself a prior ranking?
- Does the publisher have an interest in the resultan audience to please, a client, an awards program?
- Would the order be stable if the window shifted by a month?
If the methodology is not published, the answer to all of these is unknown, and the list is entertainment. That is a legitimate thing for it to be. It is just not a source.
Bottom line
Fame data is good evidence about attention and about the reference works that record it, and bad evidence about everything people want to use it for. Name the signal, state the window, publish bands instead of positions, and refuse the group tallies. What is left is smaller, less quotable, and defensible.
Common questions
Which fame signal is the least misleading?
None of them in isolation. A single subject's attention over time, from one named signal with events annotated, is the one shape these numbers genuinely support.
Why do rankings reorder so much between editions?
Because most of the ordering sits inside the metric's own variation between measurement windows. Only the top of a skewed distribution is stable.
Can a ranking be fixed by publishing the method?
It can be made evaluable, which is a real improvement. It cannot make a proxy into a measurement of fame.
What should be cited from a published ranking?
The claim, the publisher, and the date. Never the position as a fact about the person, because that sentence stops being true at the next edition.
