X Will Now Show You Your Visibility Labels.

On 13 August 2026, X published a much larger slice of the code behind its For You ranking, and started testing a tool that shows an account the visibility labels applied to it.

The announcement came through the @XOpenSource account, with VP of Product Keith Coleman describing the scope of the release.

For anyone running a brand account, the second part is the one that changes your week. Reach drops on X have always been explained the same way, by a suspected shadowban nobody could confirm or rule out.

Now a small number of accounts can actually look. That changes the order of work when the numbers fall, and the order is where most teams go wrong.

What X actually shipped

Two separate things, on the same day.

The first is an expanded publication of the For You ranking code on GitHub under the Apache 2.0 licence. The 13 August update adds the configuration parameters that weight predicted actions into a post score, the systems that filter a post out of the feed, the models that produce labels, Phoenix training code and SimClusters retrieval.

The second is a tool at x.com/i/under_the_hood, which lets an eligible account download information on the labels applied to its account and posts over the past month.

X has not published everything. The Grok prompts used to predict rule violations and some botmaker rules are held back, on the stated reasoning that people could use them to game the system.

The labels tool is a test, not a feature you have

This is the sentence most coverage skips, and it is the one that decides what you do on Monday.

The tool is in testing with a randomised group of eligible accounts. Eligibility means an account at least one year old, and X has described the threshold as ten or more posts in the past month.

What comes back is a JSON file of aggregate statistics for the past calendar month. It reports which visibility-affecting labels touched your account or your posts. It does not break that down post by post, and it does not cover ad systems or timelines other than For You.

So if you open that page and find nothing there, you have not learned anything about your account. You have learned that you are not in the test group.

What the published code says a label can do

Worth understanding before anyone opens a JSON file and panics.

Visibility filtering returns one of three answers for a given post and viewer: show it, show it behind a warning the viewer can dismiss, or drop it entirely.

The inputs are not only the post. Labels come from account-scoring models as well as content classifiers, and the code applies different rules to recommendations than to posts served to people who already follow you.

That distinction has a commercial shape. An account can be held back in For You recommendations while its own followers keep seeing everything, which reads as a collapse in reach alongside a perfectly healthy engagement rate.

Account-level signals in the published systems include patterns the platform reads as fake activity, the ratio of blocks and reports to favourites, and a credibility score derived from the follow graph. Post-level labels come from classifiers reading text, images and video for things like spam, adult content and violent imagery.

Why “shadowban” became the default explanation

Reach falls, someone says the word, and the room relaxes. A penalty is external, unearned and nobody’s fault.

It has been the most comfortable explanation available and, until now, the least checkable. A theory that cannot be tested is a poor basis for a decision that costs a quarter.

The real cost is not the theory, it is what happens next. Formats get abandoned, calendars get torn up and a strategy gets rewritten to solve a punishment that may never have existed. Most of the reasons social accounts stop growing are structural rather than punitive, and structural problems do not respond to a panic rebuild.

1. Check the label

If you have access, start here, because it is the only step that can close the question in one move.

A label found is a specific problem with a specific fix. A classifier misreading a piece of media, a spam signal thrown by a link pattern, a run of posts tripping the same rule.

No label found, or no access to the tool, and you move on. What you do not do is keep the theory anyway.

2. Check the content

Put the fortnight that underperformed next to the fortnight before it, post by post.

You are looking for the things a team changes without deciding to: a new format, a weaker opening, a shift from conversation to broadcast, a run of link posts, an editorial voice that drifted while someone was on leave.

Then ask the blunter question. Was the work simply less good than it was? Nobody volunteers that answer, and it is correct far more often than a penalty is. An outside audit and strategy review earns its fee mostly here, because the person who made the work is the worst placed person to judge it.

3. Check the rhythm

Publishing pattern is the quietest cause and the last one anyone looks at.

• A gap that broke a daily cadence, even a short one
• A day of the week that quietly stopped being covered
• Volume that jumped, so each post took a smaller share of the same audience
• Approvals that slowed, so timely posts went out a day late and landed cold
• A platform-native format traded for a cross-posted one

None of these makes a single post look wrong, which is exactly why a content review misses them.

The same order works where there is no tool

X is currently the only platform handing an account its labels, and only to a test group. Everywhere else, step one is unavailable.

That makes the discipline more important, not less. You cannot check a label on Instagram or TikTok, but you can refuse to invent one. Treat a suspected penalty as the last explanation you reach for, after content and rhythm have been ruled out, never the first.

There is a second gain here that has nothing to do with the tool. The ranking weights and filtering rules are now a document you can read, so the mechanics of the feed become something a team studies rather than something it repeats as folklore. That is ordinary platform work, and it beats guessing every time.

What good looks like

A team that knows the order before it needs it. Written down, three checks, in sequence, and nobody touches the strategy until all three have run.

• The drop is described with a number and a date range, not adjectives
• Label data is pulled where it exists, and its absence is recorded as unknown rather than as proof
• The content of the period is compared honestly against the period before it
• Cadence, volume and approval speed are checked as causes in their own right
• Only then does anyone open the strategy document

This is not scepticism for its own sake. It is the difference between fixing the actual constraint and rebuilding a system that was working fine.

Next step

If your social output feels busy but not effective, start with a diagnosis rather than a theory. Our free social media audit checklist walks the same order, and it works on the platforms that will never show you a label.