The bar a16z published
Asked what the firm screened for when hiring its new media team, Ben Horowitz gave an answer with no marketing vocabulary in it at all. The best signal, he said, is whether you have done it: whether you have built some brand on something, built an audience, and can deliver it end to end.
He was explicit that heavy experience in traditional media counts against a candidate rather than for them, because the two disciplines run on opposite rules.
Marc Andreessen added the part that matters operationally. The skill is storytelling, the elite level of it is far above the average level, and the people who cannot do it never learn. Most people from a conventional marketing background never had to, because they always had a principal to supply the story.
That bar produced a fellowship that accepted 65 people out of 2,000 applications. A little over three per cent, at a firm that can pay whatever it wants and interview whoever it likes.
We screen creators against effectively the same criterion, for the same reason. Proof of work over credentials, demonstrated audience over stated reach, and the ability to write something a stranger will read to the end. The buy-side guide sets out what that means in practice from a creator's side, including the parts that disqualify people.
Three ways to own distribution
The new media conversation collapses three different businesses into one phrase, and they have almost nothing in common operationally.
| Model | Example | What it reaches | Structural weakness |
|---|---|---|---|
| In-house media arm | a16z, roughly 70 people | One firm's own audience | Costs a balance sheet almost nobody else has |
| Creator-owned show | TBPN, around $5M on ten staff | One audience, which the hosts are | The hosts cannot be replaced or duplicated |
| Network operator | This model | Many audiences, none of them ours | Depends on being genuinely useful to both sides |
The first two are excellent businesses and the numbers say so. They share one property worth naming, which is that the audience and the individual are the same asset.
When a host leaves, the audience leaves. Analysts writing about creator-led media list key-person dependency as the material risk in the category, and they are right to.
A network is built the other way around. No single account is load-bearing, because the roster is designed so that any one creator can be absent from any one launch without changing the outcome. That property is unglamorous and it is the entire reason the model scales past the founders.
What intake filters for
Applications arrive continuously. Four things are assessed, in this order.
Median, not peak
Every account has a best post. The best post tells you almost nothing, because it usually reflects one moment of luck or one borrowed audience. The middle of the distribution over a recent window is what predicts how a briefed post will land, and it is the number a proposal should be built on.
Audience composition against a buyer profile
A 200,000-follower account whose audience is other creators is worth less to an infrastructure company than a 20,000-follower account read by platform engineers. Overlap with the specific buyer is the measurement that matters, and it is why follower count sits near the bottom of the list.
Reliability inside a window
Launches are timed. A creator posting four hours late is outside the window that made the campaign work at all, so consistency about time is weighted heavily and is the most common reason we stop working with someone.
Whether they can write
Every creator writes their own copy from a brief. Identical posts across five feeds are recognised as an ad campaign within seconds, so an account that can only run supplied text is not usable at any follower count.
Why a roster has to keep moving
A roster held static becomes a list of people who used to be right. Three things degrade it, all of them gradual enough to miss without measurement.
Reach decays. Accounts have seasons. Median performance can fall for a year while the follower count keeps climbing, which is exactly the pattern that makes follower count misleading as a hiring signal.
Without continuous measurement you find out when a launch underperforms.
Audiences drift. Creators change subject. Somebody who was the right account for developer tooling in early 2025 may now be writing about something else entirely, and the audience that followed them for the first thing has already gone.
Categories change shape. New buyer segments appear faster than rosters refresh. Agent infrastructure barely existed as a category two years ago, and covering it now means creators who were not on anybody's list then.
Most movement on a roster is reassignment rather than departure. A creator who stops fitting one category often fits another better than the people currently in it, and the useful question is which launches somebody belongs on rather than whether they belong at all. The bench model covers what this looks like when a roster is assembled around one client.
What is automated, and what stays human
The measurement layer should be mechanical, and ours is. Post-level performance across the roster, median scoring against a rolling window, decay flags, audience overlap against a supplied buyer profile, and reach modelling for a proposed creator set are all arithmetic.
Doing them by hand is how agencies end up quoting reach from a spreadsheet somebody updated last quarter.
Four things are deliberately kept off that layer.
- The brief. Written per creator, per launch, because the same release reads differently to a working editor than to an AI tooling account.
- The angle. Whether a claim is worth defending in public at 9pm is a judgment, and it is the one that decides whether a launch produces conversation or embarrassment.
- The relationship. Creators decline launches. That right is the reason their audiences still trust them, and it is covered in how slots are allocated.
- The final read. Somebody looks at every set before it goes out. Automated matching produces plausible sets that are wrong in ways only a person notices.
A network that automates the second list produces scripted posts that read like a media buy, which defeats the purpose of using people the audience chose to follow.
The honest limit on all of it
The sharpest criticism of everything above came from a16z's own team, and it is worth quoting because it is correct.
Founders overindex on distribution and tactics before getting the message right, and distribution is a multiplier on the message. Amplify the wrong claim and you have told everybody you care about that you are not interesting.
That is the real constraint on a network, and no amount of roster quality removes it. A perfectly matched creator set carrying a claim nobody cares about produces reach and nothing else.
It is why every launch we run starts with the claim rather than the creator list, and why we ask clients for a brief rather than finished copy. The sequencing is set out in launch narrative design.
The corollary is the useful part. Once the message is right, the multiplier is the whole game, and the size of the multiplier is decided by how well the network has been maintained in the months before anybody called us.
Questions we get asked
Why does a creator roster need to rotate?
Because reach decays and audiences drift. An account that matched a buyer profile eighteen months ago may have shifted subject matter, changed posting rhythm, or seen its median performance fall while its follower count kept rising.
A roster held static becomes a list of people who used to be right. Rotation is a quality standard rather than churn, and most movement is a creator moving between categories rather than leaving.
What part of running a creator network can be automated?
Measurement, matching and monitoring. Pulling performance on every post, scoring accounts on median rather than peak, flagging decay, checking audience overlap against a buyer profile and modelling reach for a proposed set are all mechanical and should be. Briefing, negotiation, judgment about whether an angle is worth defending in public, and the relationship itself are not, and a network that automates those produces posts that read like an ad campaign.
Is a creator network different from an influencer agency?
The difference is whether the roster is standing or assembled per campaign. An agency that sources creators when a brief arrives is negotiating with strangers on a deadline. A standing network has already vetted, worked with and measured the same accounts across many launches, which is what makes reach predictable in advance rather than discovered afterwards.