Can AI Predict Virality? What a Score Really Is

No tool can reliably predict virality before you publish, and none can honestly promise it. An honest pre-publish score is a percentile against a stated reference set, not a forecast of views. In one large study, the best models explained less than half the variance in how far content spread. Prediction is a signal; measurement is the verdict.

Can AI predict whether a video will go viral?

AI cannot predict virality to a useful degree before a video is published — and the ceiling is not a temporary engineering problem waiting on a better model.

The clearest evidence comes from a study that set out to measure the limit itself. Researchers modelled information cascade size on Twitter with an unusually rich feature set covering users, content and past performance, and found their best-performing models explained less than half the variance in how large a cascade became. The implication the authors drew was the interesting part: even with unlimited data, predictive performance would likely remain bounded well below deterministic accuracy (Martin, Hofman, Sharma, Anderson & Watts, WWW 2016).

A decade earlier, a controlled experiment showed why. 14,341 participants downloaded previously unknown songs in an artificial music market, with some groups able to see what earlier participants had chosen. As the strength of that social influence increased, so did both the inequality of outcomes and their unpredictability. Quality mattered at the edges — the best songs rarely did badly and the worst rarely did well — but any other result was possible (Salganik, Dodds & Watts, Science 2006).

That is the honest shape of the thing. Craft sets a floor and a ceiling. Inside that band, the outcome is settled by who saw it, when, and against what else.

What is a virality score actually measuring?

At best, a virality score measures a percentile: this clip ranks above X% of some reference set on features that correlate with performance — opening seconds, pacing, caption legibility, retention curves of comparable clips.

Percentiles are a legitimate instrument, and YouTube publishes one itself. Its Analytics API exposes relativeRetentionPerformance, which places a video's ability to hold viewers at a given point in the video against all YouTube videos of similar length, on a 0-to-1 scale where 0.5 is the median — half of comparable videos retain better, half retain worse (YouTube Analytics API metrics).

Note what that metric does not do. It produces no view count, claims no causation, and is defined against a stated reference population. A percentile is only meaningful when you know what it is a percentile of.

Why is the outcome uncertain even when the clip is good?

A good clip's outcome stays uncertain because the clip is one input into a recommendation system the creator does not operate.

YouTube describes the Shorts Feed as personalised to what it thinks a given viewer wants to see next, and states the mechanism plainly: the system compares your viewing habits with those of viewers whose habits resemble yours, then suggests content on that basis (How YouTube recommendations work). Ranking is a statement about an audience, not about a file.

YouTube's search and discovery tips for Shorts name three external factors that influence how many people see your videos, and all three sit outside your file: topic interest, meaning how many people worldwide are interested in and watching a given subject; competition, because the system ranks your video against all the channels a viewer might watch instead; and seasonality, because traffic changes at different times of year (Search & discovery tips for Shorts).

Two of those three are about other people's videos and other people's attention. No amount of analysis of your file can observe them before you publish.

Why do 100-point virality scores keep appearing in marketing?

Virality scores out of 100 keep appearing because an unfalsifiable number is an excellent sales object.

Look at the structure. A 92 followed by a flop is explained as unrealised potential. A 92 followed by a hit is proof the score works. Without a published reference set, a threshold fixed in advance, and an error bar, the number cannot be wrong in either direction — which is exactly what makes it worthless as evidence and easy to demo.

There is a second-order failure too. Once a score becomes the thing creators optimise, it stops tracking what it was built to track. That is Charles Goodhart's 1975 observation about monetary policy, generalised well past its origin; the phrasing most people know — "When a measure becomes a target, it ceases to be a good measure" — appeared in a 1997 paper by anthropologist Marilyn Strathern, who credited it to Keith Hoskin (Goodhart's law). Clips get engineered to satisfy the scorer rather than the viewer.

None of this makes scoring dishonest by nature. It means the presentation is where the honesty lives. A percentile with a named reference set is an instrument. A three-digit "virality score" with no denominator is decoration.

So what can be predicted?

Growth can be predicted, to a degree, once publishing has already started.

A study of photo reshare cascades on Facebook found that the relative growth of a cascade becomes more predictable as more of its reshares are observed, with temporal and structural features doing the heavy lifting rather than the content itself. Early on, breadth of spread was a better indicator of a large cascade than depth (Cheng, Adamic, Dow, Kleinberg & Leskovec, WWW 2014).

This inverts the usual pitch. The genuinely predictive moment is not before you press publish; it is after, once early spread can be observed. Which makes a pre-publish number a triage device — a way to order your review queue — and makes post-publish measurement the thing worth building a workflow around.

What should you measure after publishing?

After publishing, measure your own analytics in fixed windows, pulled from the platform's authorised API.

YouTube's Analytics API gives the channel owner views, estimated minutes watched, average view duration, average percentage viewed, shares and subscribers gained, among others (metrics reference). Access runs through OAuth 2.0 with the owner's consent and read-only scopes such as yt-analytics.readonly (authorisation guide), and the owner can remove a connected app's access later from their Google Account (Google Account help).

Three practices turn that data into something you can act on:

• Fix the window. Compare every clip at the same age — 48 hours and 7 days, for instance. A clip measured at day 30 against one measured at day 2 tells you nothing.
• Compare within your own channel first. Your own median clip is the only reference set you fully control, which is why a retention baseline built from your own data is the place to start.
• Make it showable. If you cannot hand a client, a sponsor, or your future self a record they can open and check, you have an impression, not a result (how to prove clip results to sponsors).

That third one is the easiest to skip, and it is where Everpop puts its weight. Everpop is an AI clipping tool that turns your own uploaded videos into Shorts and Reels and publishes them, review-first, to the channels you connect. Its pre-publish number is a percentile with a stated reference set: once your channel has 8 measured clips, it predicts how a new clip will rank among your own clips and signs that prediction before the clip goes out, so anyone can later check it has not been rewritten. Eligible YouTube Shorts then come back with 48-hour and 7-day numbers from the YouTube Analytics API, behind a receipt link a client can open without an account. The signature covers the prediction only; the measured numbers come straight from YouTube Analytics and are not signed (how receipt signatures work). The receipt grades the prediction. It never promises an outcome.

The same logic argues for keeping your files portable. If measurement is the point, you should be able to walk away with the work — Everpop's FCPXML, EDL and SRT exports into Final Cut Pro, Premiere Pro and DaVinci Resolve mean the edit stays yours regardless of which numbers you end up trusting (plans and what's included).

What should you check before trusting any score?

Check what the score is a percentile of, whether its definition stays stable, and whether the tool shows measured results a third party can verify. Six questions cover it:

• Can the vendor say what population the score is a percentile of?
• Is the score's definition stable, or does it shift silently between releases?
• Does the tool report what actually happened after publishing, or only what it guessed before?
• Are post-publish numbers pulled from the platform's official analytics API under your own authorisation?
• Can a third party — a client, a sponsor, an agency lead — verify a result without taking your word for it?
• Does the marketing imply a growth or revenue outcome? Nothing in the research cited here supports that.

A tool that answers the first five well is worth evaluating. A tool that leans on the sixth is selling the one thing it cannot deliver.