Google, TikTok, Instagram, X: four algorithms became the same machine

For twenty years the big distribution systems were different machines. Google ranked pages by links. Twitter showed you who you followed, newest first. Instagram was a photo feed of your friends. TikTok was the weird one, guessing what strangers might watch. Then, in about eighteen months, all four quietly rebuilt themselves into the same machine, and almost nobody said it out loud.

The machine works like this. Take everything that could be shown. Have a model read it, the actual content, not the metadata. Predict, per person, the probability that this specific human will give it deep attention: finish the video, reply to the post, send it to a friend, not click away from the answer. Rank by that prediction. Repeat billions of times a day.

The receipts, platform by platform

X made the loudest move: in January 2026 it discarded its legacy ranking stack entirely for a Grok-based transformer that reads every post and watches every video, making around 5 billion ranking decisions daily. The reported weights tell you what the model is for. A reply counts roughly 27 times a like. A genuine back-and-forth conversation, roughly 150 times. Likes, the currency of the 2010s, are now the copper coin. Even the Following feed is algorithmically re-sorted, which is a quiet way of saying the follow relationship no longer decides much.

Instagram said it with a metric. Adam Mosseri spent 2025 telling creators that watch time ranks Reels and that sends per reach, the share of viewers who DM your content to someone, is weighted three to five times a like. Then the app consolidated everything into one number: views. Not followers. Views. When a platform renames its primary metric, it is telling you what its model optimizes.

TikTok, which invented the machine, spent the period proving how valuable it is: the January 2026 US joint venture led by Oracle, Silver Lake, and MGX exists substantially because the algorithm could not simply be handed over, so it is being retrained on American data as a separate fork. Meanwhile the bar rose. The completion rate that used to trigger wide distribution, around half, is now reported near 70 percent. The machine got pickier as everyone learned to feed it.

And Google, the biggest publisher-facing change of all: AI Overviews went from about 6.5 percent of queries in January 2025 to roughly 48 percent of searches by early 2026, and somewhere between58 and 68 percent of searches now end with no click to any website. In the fully conversational AI Mode, the no-click figure reportedly reaches 93 percent. Google still reads the web. Increasingly, it reads it so you do not have to.

Why they all converged

Not conspiracy. Economics plus capability. Every one of these companies sells attention to advertisers, so every one of them is paid in retention, and retention is best predicted by deep engagement signals, not declared relationships. That pressure always existed. What changed is capability: models got good enough to read the content itself, cheaply, at feed scale. Once you can score a video by watching it, links and follows and keywords are just noisy proxies you no longer need. Each platform reached the same conclusion because each was solving the same equation with the same new tool.

What it means if you publish anything

First, audiences are rented by the piece now. A follower count is a mailing list the platform charges you to use; every post starts nearly from zero and earns distribution on its own predicted depth. Second, the openings are everything: the three-second rule on video, the first sentence of an answer, because the model samples before it commits. Third, depth beats breadth everywhere at once: one piece that 70 percent of viewers finish outranks five pieces they skim, on every platform, simultaneously, because it is the same machine. And fourth, for the written web specifically: being the source a model cites has replaced being the link a person clicks, which favors pages with verifiable, unusual substance over pages with volume.

Where this leaves you

The convergence is bad news for tactics and good news for material. Every trick tuned to one platform's quirks depreciates, because the quirks are being replaced by models that read like careful humans. What survives is what would survive a careful human: things worth finishing, worth replying to, worth sending to a friend, worth citing. The four machines disagree about formats and durations. About substance, for the first time, they all agree.