Content Clusters: Finding What Actually Works for a Creator

You brief a creator whose engagement rate sits comfortably above the median for their niche, and the post underperforms almost everything else on their feed. Nothing went wrong in the execution. You commissioned a product review from someone whose audience shows up for training footage and quietly leaves when a box appears on camera.

The engagement rate did not lie. It was an average across a year of posts, and an average is the one number guaranteed to conceal the variation you were about to spend money on.

An average describes a creator; a distribution describes their content

Every creator has content that beats their own baseline and content that loses to it, usually by a wide margin. That spread is not noise. It is the most actionable thing in the account, and it is invisible in any single figure.

The useful question is not "how does this creator perform?" It is "which of the things this creator makes performs, relative to everything else they make?" Those are different questions, and only the second one produces a brief.

The critical move is the comparison group. You are not ranking this creator's gym content against other creators' gym content: that comparison is confounded by audience size, market, posting frequency and platform. You are ranking it against their own median, which holds all of that constant automatically. Their audience, their algorithm treatment, their posting cadence, their year.

Why "look at their top posts" gives you the wrong answer

Scrolling a creator's best-performing posts is the standard shortcut, and it is close to the least informative thing you can do.

• Top posts are the extreme tail, and the tail is mostly luck. One post caught a recommendation surge. That tells you the ceiling, not what to expect. You are buying the median, not the maximum.
• They skew to whenever the account was biggest. Ranking by absolute engagement favours recent posts on a growing account and old posts on a declining one. Either way you are reading account history, not content quality.
• They skew to whatever format the platform is currently pushing. A top-posts list can be entirely one format the creator adopted eight months ago, which tells you about distribution rather than about subject matter.
• They cannot show you what fails. This is the decisive objection. The theme you should not brief never appears in a top-posts list, and it is the one that would have cost you the campaign.

The method

Six steps. It is tedious by hand for one creator and impossible by hand for forty, which is the honest case for tooling.

1. Fix a window. Enough posts to have a distribution, recent enough to describe the account as it is now. Forty to sixty posts, or six to twelve months, whichever is tighter.
2. Compute that creator's own median. Median, not mean: one viral post drags a mean upward and leaves the median where it belongs.
3. Group posts into themes by subject and intent, not by format. "Reels" is not a theme. "Training footage", "recipes", "gear reviews", "personal and family" are themes. If your groups turn out to be formats, you have measured the platform, not the creator.
4. Set a minimum count per theme. Three posts is an anecdote. Below about five, record the theme but hold the conclusion loosely.
5. Index each theme against the account median. Theme median divided by account median. 1.0 is baseline; anything meaningfully above or below is the finding.
6. Read the spread, not only the centre. A theme with a high median and enormous variance is a gamble. A high median with a tight spread is a reliable thing to ask for.

Illustrative output, with invented numbers, for a hypothetical fitness creator:

That last row is the entire value of the exercise. This creator is not bad. This creator is telling you, in public, exactly which format of sponsorship their audience tolerates and which one it ignores.

Step 3 is also where the sub-niche question stops being abstract. If a creator's themes do not sit inside one coherent subject, you are looking at an account whose label is doing more work than its content: which is its own problem.

This changes the brief, not the shortlist

Cluster analysis rarely reorders a shortlist much. Two creators with similar audiences and similar authority both stay on it. What changes is what you ask each of them to make.

That turns a generic brief into a specific one: your gym-floor content runs well above your own baseline and your standalone reviews run below it, so we want the product inside a training session rather than as its own post.

Three things follow from that. It is an easier conversation with the creator, because you are asking them to do the thing that already works for them. It is a better negotiating position, because you can say why. And it removes what is probably the most common quiet cause of an underperforming sponsored post: a brief that asked for below-median content from an above-median creator.

In the wider process this sits in step four, scoring what survives the hard filters, after the mechanical cuts, before you compose the mix.

Where the method breaks

This is inference from public counts, and it should be held with the confidence that deserves.

• Clustering is interpretive. Where you draw a theme boundary changes the answer. Two analysts will produce two groupings of the same account, and neither is wrong. Be consistent within a brief rather than pretending to a canonical taxonomy.
• Most creators have too few posts per theme. Differences of ten or twenty per cent between themes with single-digit post counts are noise. Act on wide gaps, not narrow ones.
• Sponsored posts confound everything. Paid content generally underperforms organic across the board. If a creator's "product reviews" theme is entirely sponsored, you may have measured the paid-post penalty rather than the theme. That is still useful, but it is a different finding, and worth separating where you can identify disclosures.
• Other variables ride along. Posting time, format, caption length, whether a post was boosted. A theme index is a correlation with a lot inside it.
• The past is not a limit on the brief. This measures what has been tested, not what is possible. An audience can absorb something new, and a creator who has never done a review well may simply never have been given a good one.
• Engagement is not the outcome you are buying. A theme with high saves and shares and unremarkable likes may be the commercially better one. Public metrics rarely expose saves, and no third party sees reach; only the creator does.

Where Lyren fits

Lyren groups a creator's posts into themes and ranks each theme against that creator's own median, for exactly the reason above: the comparison that matters is internal, and it is the one nobody has time to do by hand across a long list.

The grouping is an inference, not a fact about the creator's intentions, and it is shown as a ranking with the post counts visible so you can see when a theme is resting on five posts. Reading the result and deciding what to ask for stays a judgement call, and it stays yours. Access is invite-only via the waitlist, starting with Romania.