The standard pitch for X marketing services rests on an information gap: the algorithm is opaque, it changes constantly, and you need someone who understands it. That pitch stopped being true in January 2026, when X published the code.
What is left is a narrower and more useful question. If the ranking system is public and the growth shortcuts are prohibited, what should a brand on X actually be paying for?
Key Takeaways
- X open-sourced its recommendation algorithm on GitHub in January 2026.
- Ranking now runs on a Grok transformer that reads posts rather than counting interactions.
- Buying, selling or trading X accounts and usernames is explicitly prohibited.
- So is compensating anyone to inflate likes, reposts, views or follows.
- Promoting third-party services that do either is itself a policy violation.
The Algorithm Is Not a Secret Anymore
On 10 January 2026, Elon Musk said X would publish the code determining which organic and advertising posts get recommended, and would repeat the release every four weeks with developer notes. The repository went live later that month under an Apache 2.0 licence at github.com/xai-org/x-algorithm.
The system it describes is a departure from the 2023 release. Ranking is handled by a transformer built on the same architecture as xAI’s Grok model, replacing the hand-coded heuristics that ran the previous For You feed. Rather than counting interaction signals against fixed weights, the model reads the post and scores relevance to each individual user.
Two consequences follow, and they point in opposite directions for the services industry.
The first is that engagement-velocity tactics lost their mechanism. Reply pods, retweet chains and first-hour interaction thresholds worked because they produced the exact signals a counting system counted. A model that reads content semantically is not counting those signals in the same way, which means the manufactured version of them buys less than it did.
The second is a caveat worth keeping. X released portions of its ranking logic in 2023 and that repository went largely unmaintained afterwards, so the four-week cadence is a commitment rather than a track record. Published code also does not guarantee that the published version matches what is running in production. Treat it as the best available evidence, not as ground truth.
What X’s Rules Actually Prohibit
A large share of X growth advice describes conduct the platform bans by name. X’s Authenticity policy lists it directly.
Prohibited: trading, buying, selling or soliciting access to X accounts, whether through money or other compensation, and including the temporary or permanent transfer or sale of accounts and usernames. Prohibited: coordinating to exchange engagement across Likes, Polls, Replies, Reposts, Lists, Views or Follows. Prohibited: coordinating with or compensating others to inflate account metrics across any of those features.
And one clause that catches more people than the others: using or promoting third-party services that perform any of those transactions is itself covered by the policy. Recommending the vendor is treated the same way as using it.
The practical reading for anyone buying services: an agency offering aged accounts, purchased usernames, or engagement delivered on your posts is not offering a grey-area tactic. It is offering the specific conduct in the policy, and the enforcement risk sits on your account rather than theirs.
What Social Proof Research Actually Found
The argument for buying followers usually leans on social proof, and the underlying research is real. It also says something less convenient than the pitch.
The canonical study is Salganik, Dodds and Watts, published in Science in 2006, which built an artificial music market and showed participants either download counts or nothing. Popularity signals did influence choices, substantially. The finding that gets left out is the second half: social influence made outcomes both more unequal and less predictable. Which items became hits was close to arbitrary, and early signals compounded into large gaps regardless of quality.
That is an argument for the importance of early traction. It is not an argument that manufactured traction produces the same effect, because the mechanism in the study is real people responding to real aggregate behaviour, not a counter with a number on it. Nothing in that literature suggests you can buy the cascade.
What Consistency Actually Does
Posting consistently is genuinely useful and routinely oversold. It does not guarantee growth, and any service promising that it does is describing an outcome nobody controls.
What consistency does is produce enough output for a relevance model to characterise your account, and enough surface area for the occasional post that travels. An account posting twice a month gives the system almost nothing to work with. An account posting daily on an incoherent mix of subjects gives it something contradictory.
Under a model that reads content, subject coherence matters more than it did under a counting system. Pick a territory narrow enough that a stranger could describe what your account is about after scrolling for ten seconds.
Replies Are the Underrated Half
The original point about conversation holds up, and it is the part of X marketing that agencies genuinely can help with.
Responding to mentions, answering questions, and joining live conversations produces the thing a relevance model is looking for and the thing a customer remembers. It is also the part that scales badly, which is a legitimate reason to pay someone. Note the distinction that matters: an agency writing considered replies on your behalf is doing marketing work. An agency generating replies from accounts it controls is doing something the Authenticity policy prohibits.
So What Is Worth Paying For
Time and judgement, mostly. Not algorithmic access, which is now public, and not engagement, which is prohibited and decreasingly effective.
The defensible services are production capacity, a consistent posting cadence you would otherwise abandon in month two, someone monitoring and responding during hours you are working, and analysis that connects what you publish to what your business actually gets. Before signing anything, ask directly whether any part of the engagement is generated rather than earned, and ask whether accounts will be created, purchased or transferred. The same evaluation logic applies as in any other marketing agency assessment, and the answers should be in writing.
Conclusion
X marketing got simpler to reason about in 2026, if less comfortable to sell. The ranking system is published, the shortcuts are named in policy, and the model reading your posts is harder to game than the counter it replaced.
What remains is unglamorous: say something worth reading, say it often enough to be characterised, and answer people when they reply. Anyone selling more than that on X is selling either something public or something prohibited.
Read Next
- How the Instagram Algorithm Works in 2026: Signals, Surfaces, and What Actually Changed
- How to Evaluate an SEO Agency: What to Check Before You Sign
- Personal Branding on Social Media: What Actually Builds Authority
