Avatar Algorithm Guide

An avatar algorithm is a named discovery mechanism: define a customer, scroll TikTok and Instagram as that person on a real device, and take the creators the feed keeps serving. This page is the how-to. Inputs, session, recurrence, output, where the mechanism stops.

The one-page entity is Avatar algorithm. The first-person method essay is How we find creators by scrolling a real TikTok avatar. Keep this guide if you need the mechanism written like a spec.

What the mechanism claims

Hashtag indexes and creator databases answer who exists (or who tagged a string). A For You / Reels session answers who is being distributed to this person right now.

Those are different objects. Completeness is a database virtue. Recurrence in a live session is a session virtue. The avatar algorithm bets that matching quality for an indie app (or a single client avatar) is a session problem.

It does not claim the algorithm “understands” your user. It claims that watching the feed as them is a better research method than watching a tag index.

Inputs (what you actually specify)

A scrollable avatar is narrower than a media plan.

Problem in the user’s mouth, not your App Store subtitle
Adjacent content they already watch (specific videos or formats)
What they already quit
Language and region that match the store listing
One ICP per session

“Women 18-34” is not an input the mechanism can use without lying. Two phones that both liked three climbing videos last month will still diverge. That personalization is the signal, not a bug.

The session, not the profile

TikTok and Instagram do not show “fitness” as a channel. They show a sequence of videos to a device that has a history. That sequence is what you watch.

A practical loop:

Warm the device toward the niche (watch, linger, skip the off-niche).
Record who appears.
Keep going until faces repeat.
Those repeaters are the list.

Headless crawls of public metadata are not a For You page. The For You page is a device, a login, a watch-time pattern, a language, a region. JumpTrend’s public description is specific: AI on real phones, in the avatar’s niche, on TikTok and Instagram. Slower than pinging a giant catalog. That slowness is the point.

ICP feed scraping, in this cluster, is slang for that observation loop on operated devices (public content, public business contacts). Not unauthorized access and not a pirate mirror of TikTok.

Recurrence as the filter

One appearance is noise. A creator who returns across a warmed session is being pushed to that persona. That is closer to “my customer already watches them” than “they used #habitapp.”

Weird niches help. A narrow session produces repeating faces. A wide session produces the same celebrities everyone else is pitching.

You still will not learn who will go viral. Viral is a tail. Recurrence tells you who is already in the room.

Output (what the mechanism is for)

A short list. Eight to a few dozen. Handle, one video, why they belong to this feed, public contact if published.

The list is not a booked campaign. It is not an email sequence. It is not a Spark plan. Matching is the controllable part. You have other names when one person is a no.

Artifact: a real outreach constraint

The mechanism stops at the feed. You cannot email a For You page.

Creators who dominate a specialized session often do not want to be found by brands. They may have no email in the bio, a closed form, or a manager who only answers after a rate you cannot pay. The algorithm can still be right about fit and the inbox can still be empty.

Write that constraint into the job before you buy a list:

Outreach constraint (do not skip):
- Contact source: public bio / public form / nothing published
- If nothing published: we do not scrape private addresses
- We do not invent a reply-rate
- Several in-feed names beat one macro with a booking agent
- First line must name a video we actually saw
- Two or three follow-ups, then stop

If the only people in the feed are uncontactable macros, the avatar may be too famous, or the product may need a narrower session (the educator, not the athlete). That is a spec change, not a CRM feature.

What the mechanism is not

Not a giant creator index
Not a hashtag export
Not a promise that AI picked a winner
Not a tutorial for evading platform terms or automating logins you do not own
Not measurement, not Spark, not ASO

JumpTrend is this method as a product. Scout is $1,500/month for the list. Outreach is $2,500 if you want contacting and follow-up. Dealmaker is $5,000 if you want paper you approve. Talent extra. No demo. No download guarantee. The method would still be the method if JumpTrend did not exist. You would just contaminate your personal algorithm doing it.

Failure modes

Five ICPs on one phone, then complaining the list is messy
Warming for twenty minutes and calling it a census
Treating the first celebrity who appears as the roster
Exporting the list into a 1,000-email drip (you will poison a niche)
Asking the mechanism for CPI

If you already run 200 ambassadors in a CRM, you needed operations more than another discovery method. If you needed to know who a specific customer is being shown this week, you needed a session.

See For You page, ICP feed scraping, Creator marketplace vs feed discovery.