Put a creator scorecard in front of a client and someone will point at the number. "Why is she an 87 and he's a 74?" It is the right question, and most tools cannot answer it. The score is a black box with a colour gradient, and the honest reply is that nobody in the room knows.
A composite score compresses several measurements into one, and compression is lossy by definition. That is acceptable (a shortlist of eighty has to be ordered somehow) but it is only defensible if two things hold: you can decompose the number back into its parts, and the parts were measured against the right comparison set.
The two questions that decide whether a score is worth anything
Can you see inside it? A score you cannot decompose is a rank you have to take on faith. When a client challenges a placement, "the model says so" ends the conversation badly. Being able to say "high follower quality, engagement in the top quartile for this niche, but posting has been irregular for two months" turns a defence into a discussion.
Was it scored within a niche or across everything? A global score compares a B2B accountant to a dance creator and declares one of them better. Since the underlying metrics (engagement above all) vary far more between niches than between strong and weak creators, a globally scored composite mostly ranks niches and only incidentally ranks people. Anything that sorts an entire index into one leaderboard is doing this, whatever the marketing says.
The components, and how each one bends
Lyren's authority score is a 0–100 composite of six inputs, ranked within the creator's niche. They are worth walking through not because the recipe is proprietary (it should not be) but because each is a proxy for something it does not directly measure, and each has a known failure mode.
| Component | Proxy for | How it gets gamed |
|---|---|---|
| Audience size | How many people could see a post | Slow drip purchases from aged accounts with real photos |
| Engagement quality for that size | Whether this audience reacts more than audiences of the same size usually do | Pods, comment-for-comment groups, engagement-bait formats |
| Engagement volume per post | How many people actually react to a typical post | Chasing whichever format the algorithm currently favours |
| Consistency | Whether delivery is predictable and current rather than lucky | Volume posting, which smooths the median without raising it |
| Growth velocity | Whether the account is ascending or fading | One viral post, or a purchase, inside the measurement window |
| Authenticity | Whether behaviour resembles a person rather than an operation | Careful, expensive fraud that mimics organic patterns |
Two of these deserve more than a table row.
Engagement quality is the heaviest input in most composites and the one most sensitive to what it is compared against. Engagement rates fall as follower-based audiences grow and rise as view-based ones do, so a comparison that ignores size mostly ranks size: every small account with a loyal audience outscores every large one. Comparing against creators of the same size on the same platform removes that, at the cost of a new question: what counts as typical. Lyren measures it across its own index, per platform and follower decade, and a thin decade is a thin comparison. The other trap is the comparison group moving under you: an early version scored engagement against whichever peers happened to be in the database that day, and a 12,000-follower account judged against twelve micro accounts came out on top of a niche. A score that depends on the order creators were discovered in is not a score.
Engagement volume asks a different question from the quality comparison: not whether an audience reacts unusually well, but how many people react at all. It separates a 300,000-subscriber channel whose videos draw five thousand views from one that draws four hundred thousand, which the follower count alone cannot. It is also the input most distorted by format cycles. A creator who happened to be making the format the platform is pushing this quarter looks stronger than their craft warrants, and will look weaker in six months having done nothing differently.
What a single number destroys
Compression is the point of a score and also its main cost.
It hides disagreement between components. Two creators can arrive at 78 by opposite routes: one with excellent audience quality and mediocre reach, another with strong reach and a questionable audience. Averaged, they are identical. Commercially they are not remotely the same booking, and for most briefs one of them is disqualified outright.
It flattens confidence. A score built from three months of dense data and one built from six posts on a small account render as the same width on the same bar. The interface makes them look equally solid. They are not.
It invites false precision. An 87 and an 84 are not meaningfully different; any composite carries enough estimation error that the gap is noise. Sorted lists encourage people to read rank order as a ranking, when the honest reading is bands: strong, plausible, worth a look, no.
It becomes the argument. Once a number is on a slide it stops being evidence and starts being the claim. The reasoning that produced it, which is the part the client is actually paying for, quietly drops out of the deck.
Using one honestly
A composite score is a sorting aid. It is not a decision, and a scorecard that presents it as one does the client a disservice.
- Sort with it, then stop using it. Its job is to move eighty candidates into a reviewable order. Once you are down to a shortlist, look at components, not totals.
- Read the components for disagreement. A wide spread between inputs is more informative than their average. Any creator whose score is carried by a single component deserves a second look.
- Check it against a filter it does not contain. Audience market is the obvious one. A score says nothing about where the followers are, and that mismatch is a more common and more expensive failure than a mediocre score.
- Put the reason on the slide, not the number. "Top-decile engagement in her niche, consistent output, audience mostly in-market" is defensible in a meeting. "Authority 87" is not.
- Re-score before you buy. These inputs move. A score computed six weeks ago describes a creator who has posted since.
What this score cannot do, ours included
- It cannot detect fraud. It can flag patterns consistent with fraud. From outside the platform nobody sees the real data, and a careful seller can produce a clean profile on every public surface.
- It cannot measure brand fit. Tone, values, aesthetic, prior partnerships and whether a founder will be comfortable in the same frame are not in the model, and should not be.
- It cannot predict conversion. Nothing publicly observable correlates reliably with sales, and any tool implying otherwise is selling a correlation it has not earned.
- The weights are a judgement call. Ours are chosen, not derived from an outcome dataset, because that dataset would require campaign results we do not have and would not be ours to use. That is a real limitation, and it applies equally to every competitor asserting the same thing more confidently.
- Small niches produce shaky scores. In a narrow category in a small market the peer group can be a few dozen creators. The score is still computed. It is simply less trustworthy, and that is a property of the sample rather than of the creator.
Where Lyren stands on this
Lyren scores creators 0–100 within their niche rather than globally, and exposes the components rather than only the total, because a number you cannot take apart is not evidence. Exported reports carry the reasoning for the same reason: what survives a client meeting is the explanation, not the digit.
We would rather you disagreed with a component than trusted a total. Lyren is invite-only via the waitlist while coverage is built out, starting with Romania.