Changes to YouTube Algorithms in 2025: Detailed Analysis
Yeah... YouTube's algorithms are changing again. Those who have been working on the platform for the past five years remember well that this has happened more than once.
Sometimes, the changes are strongly felt, especially when everything breaks down, like it was after the introduction of Shorts in 2021. But, they go almost unnoticed.
However, today we're talking about quite significant changes that are happening in the background and aren't immediately obvious.
The main bullepoints behind these changes are:
- The functions for duplicating videos in different languages
- And the use of a new AI format within YouTube's own algorithms
Once again, we, as creators, have to adapt to new conditions for effectively promoting content on YouTube. Now, let's break down everything you need to know.
Several times a year, YouTube managers communicate with the world, dropping scattered bits of information about the platform’s operations. Extracting something truly useful from these updates can be a challenge, as we have to separate PR about new features from genuinely valuable insights.
But this time, the Creator Insider channel released a truly interesting interview about YouTube’s new algorithms:.
All the useful insights come from —Rene Ritchie and Todd Beaupré, who have been working with and on YouTube for many years.
So, let’s dive into what’s changing with YouTube’s algorithms in 2025.
How does the content search and recommendation system work?
Rene asked a very broad and complex question, so Todd’s response is somewhat fragmented. But we’ll help you piece together a more detailed picture with our commentary.
First, Todd lays out an important foundation for all creators: YouTube’s recommendation systems are designed for viewers, not for creators.
Bloggers often wonder, “Why isn’t YouTube promoting my videos?”—but the system isn’t built to do that. Instead, it works differently.
YouTube’s main goal
The platform’s primary objective is to match each viewer with the videos that will bring them the most enjoyment and align best with their interests.
This benefits YouTube because if the platform recommends the right videos, viewers will stay longer, watch more content, and in turn, see more ads—the real source of revenue.
Todd didn’t exactly say this, but we can easily foresee YouTube’s motivations when it comes to its algorithms, viewers, and content.
How do recommendations work?
Let’s say there’s a viewer who loves fishing. They watch and like specific fishing-related videos.
- YouTube notices this pattern.
- Analyzes the most effective videos from fishing creators.
- Curates a personalized selection of the best videos for the viewer’s homepage.
Important note: The most effective and best videos are not necessarily the ones with the most views—we’ll get back to this later.
If a creator evaluates their content solely based on CTR (click-through rate) or watch time, it won’t help them understand if their video is truly effective.
Todd explains that the data YouTube provides is just a fraction of what the platform actually analyzes.
For recommendations, YouTube looks at the overall combination of a video’s performance metrics, plus viewer interests—not just isolated stats.
That’s why every viewer’s homepage is unique, tailored to their specific behavior.
If it were purely about numbers, we’d all be watching MrBeast—his metrics are practically unbeatable.
The "word-of-mouth" effect
One of the most interesting points Todd makes is that YouTube’s algorithms function like word-of-mouth recommendations.
He explains it like this:
When you’re picking a movie or a show for the evening, you often ask friends what they’ve been watching lately. Their recommendations influence your decision.
YouTube does something similar:
- It groups viewers with similar interests.
- Then, it recommends videos that other users in that group have enjoyed, almost as if a friend personally suggested them to you.
What about videos that don’t get views?
This is another key topic Todd touches on. While he doesn’t give a direct answer, he advises not to write off underperforming videos too quickly.
He suggests that trending or viral videos on a similar topic could suddenly boost an older video, giving it a second life.
However, this still doesn’t clarify whether it’s better to delete and re-upload an unsuccessful video or simply wait for a future resurgence.
Todd provides an example with nostalgia-driven content.
Over time, people develop an interest in events from 5, 10, or even 15 years ago, reviving old videos from YouTube’s archives—sometimes ones we’d never expect to resurface.
What conclusions can be drawn from all this?
We believe Todd is right that any old or underperforming video can take off at any moment. Many of you have probably seen this happen firsthand. However, the decision on what to do with such a video and how to handle it is up to each creator individually.
Now, let’s move on to the next topic.
How important are factors such as time of day or device type?
We’re now diving into less obvious analytics parameters that you need to consider.
It turns out that device type and time of day play a huge role in shaping recommendations for viewers.
How does this work?
Todd explains it like this:
- In the morning, while getting ready for work, you might have news playing in the background on your phone.
- In the evening, when you finally settle in, you might switch to your TV and watch a deep-dive documentary about history’s deadliest epidemics.
This means that even your own recommendations as a viewer can vary between morning and evening.
What does this mean for creators?
The time your audience watches your video is something you can’t fully control, but you can help yourself by publishing content when more of your viewers are active on YouTube.
Where to find this?
You can check this data in YouTube Analytics → Audience tab to determine peak activity times.
Device type matters
This is even easier to analyze—just open detailed analytics to see which devices your audience is using.
Why is this important?
- If most viewers watch on smartphones, make sure your text overlays and subtitles are large and easy to read on small screens.
- If most watch on TV or desktop, focus on video quality—low-resolution visuals or tiny details won’t look great on a big screen.
Optimizing for these factors can make a real difference in engagement and retention.
Authors often look for a single analytical parameter to rely on. How can you actually optimize your content?
As Todd correctly points out, we would all love it if just improving one parameter, like retention, would automatically increase our views on YouTube. But there is no "one" parameter. The problem is that the algorithm is too smart and knows that different parameters and their combinations need to be analyzed at different times. The importance of these parameters will vary depending on the context.
For example, on TV, watch time will be more important, while on mobile devices, audience reactions like likes, dislikes, and comments will play a bigger role.
Or, for podcasts, watch time will be more crucial, while for music videos, audience engagement will matter more.
And if you think about it this way, it becomes clear that no single analytical parameter will give you a clear answer to the question of where to focus to optimize your video.
Here’s a recommendation from us: you will fill in these gaps with experience and analysis of your target audience. Because only through practice and detailed analysis will you discover all of these things.
Many people think that algorithms are about keeping viewers engaged at any cost, but YouTube is not focused on maximizing engagement.
So, set aside your 100% viewer retention, nobody needs it anymore.
Alright, let's leave the excessive sarcasm behind and explain what's really going on here…
If your video is watched from beginning to end—that's good, but if a video on a similar topic by another creator is watched from beginning to end AND receives a like, that will be more important to YouTube.
This is where the thesis emerged that not all viewer time spent watching videos is equally valuable.
We've come full circle again and arrived at the conclusion that YouTube doesn't analyze video parameters separately, only collectively, so your 100% retention might indeed be useless.
The platform needs to understand which videos the viewer watched and ultimately enjoyed.
Therefore, YouTube introduced the concept of "viewer satisfaction" precisely to not just count watch time and retention, but to understand viewer behavior during the viewing and how they feel about the time spent on the video.
This is done through surveys that sometimes appear after watching videos, as well as through likes, dislikes, the "Share" button, rewatches, and video skipping.
This data is fed to algorithms so they can understand what to evaluate in creators' videos to make better recommendations for viewers.
Through this method of trial and error, YouTube staff realized that if they consider many parameters at once, viewers return to the platform more often.
What's more important: looking at analytics metrics numbers or observing how they work for a specific creator and thinking about how to improve these values?
And this is again a very strange question on the same topic: bloggers are trying to find specific criteria by which they can evaluate their success, and how good they are compared to other bloggers, but in the end you still won't succeed, you fools.
According to Todd himself, success depends on WHO watches your video, that is, who YouTube shows the video to, what reach it has, and how many views it ultimately gets.
So at this point we make ourselves a note: sometimes creators have no control over what audience YouTube sends their videos to.
And further, Todd emphasizes once again the fact that objectively evaluating your video and a competitor's video won't work, so even if you place their watch time and yours side by side, it won't be objective.
Even if you put two of your videos side by side—it still won't give you any clear answer.
And yes, YouTube shows us these metrics in analytics, but Todd advises us to focus on our goals, not on metric numbers.
If you want to get views to sell your products and services, then think about what's more important to you: a 20% CTR with 10 views or a 5% CTR with 100 views?
That is, where will the conversion from view to purchase be higher?
The point here is: not all analytics metrics are objective, and not all need such strong focus, because ultimately only reach and number of views matter, and everything else is just details.
But here we have a comment: yes, this is certainly the most important thing, but the respected YouTube representatives don't tell us which metrics can help solve the problem with reach and number of views.
Saying: "Just make interesting videos" is not enough. A huge number of creators do everything right but don't get impressions from YouTube, so what then?
Todd's advice here is very vague: take a step back and evaluate metrics comprehensively, as the platform's algorithms themselves do.
What about multilingual audio tracks: how will algorithms work with videos in different languages and with different target audiences?
Well, finally we've reached one of the most important and exciting topics!
And Todd starts right away with the most interesting part: along with launching the feature to upload audio tracks in different languages, they expanded the capabilities of recommendation systems on the platform.
This is done so that algorithms automatically understand that a video is available to an audience with certain interests, but in different languages.
Moreover, they set up special feedback to evaluate the effectiveness of each audio track separately, so that it doesn't happen that your video was watched well in any specific language, and then impressions were cut off in both cases.
Recommendations for working with multilingual videos are as follows:
- Translate video titles and descriptions manually in different languages
We'll add on our own: try to make video thumbnails that are clickable mainly due to the visuals, since no one will translate the text on the thumbnail for you.
And no, the A/B testing feature won't particularly help you here, it's not that smart. Although in the future, anything is possible.
Manually, you can translate the title and description of a video in the settings of each individual video in the "Subtitles" section. There you can add a translation in the required language and edit the "Title and description" field.
- Translate as many videos on your channel as possible
If a viewer watched a video they liked, but came to the creator's channel and couldn't find any more videos in their language, then there's almost no point in that single view.
This thesis from Todd once again emphasizes the main promotion method on YouTube in 2025: you need to sell the viewer not just one of your videos, but all videos on the channel. The more time a viewer spends on your channel, the more the platform itself will like you.
And the third piece of advice:
- It's better to translate 80% of your content into multiple languages than 20% of your videos into 100 languages.
I think this advice simply continues the thought above, and we can also assume that although they've adjusted the algorithms for this purpose, there are definitely still some imperfections remaining.
Oh yes, and also keep track of global trends, because in any case your success will be tied to current and in-demand topics.
Views declines and increases: what to do if the number of views starts to decrease?
One way or another, the question is quite painful: every creator has experienced ups and downs in views, but not everyone understood what actually happened.
And especially what to do about it.
If your views have decreased, it doesn't mean at all that your channel is finished. In general, declines and rises are normal for any channel, everyone experiences them, and it's impossible to keep the bar at the highest mark anyway.
There are concepts such as seasonality and cyclical viewer interest.
We came up with the second term ourselves, but this is exactly what Todd is talking about.
Seasonality is how viewers' interest in content changes depending on the time of year, holidays, vacation periods, or when most of your viewers take time off.
But cyclical viewer interest is more about human psychology.
You've probably noticed that at first, when you're really into something, you're ready to spend day and night on your favorite activity. But then interest turns into routine, and eventually you abandon your hobby altogether. But sooner or later for many people, that day comes when this cycle begins again.
It's the same with YouTube channels: we get hooked on a creator's content, binge-watch all their videos, get tired, forget, and then return again.
And if a creator and their content took off quickly, the audience that became fascinated with their videos will begin to tire of the blogger around the same time.
In this case, Todd advises first not to worry about it and instead think about how you can fix the situation, for example, by changing the format or topics you cover.
To objectively assess what caused the decline, look at analytics for at least 90 days, or better yet a year, then you can easily track seasonality, for example.
If your graphs are wave-like and the rises coincide with some regular events, then that's definitely it.
If over a long period you had a hill-shaped graph with one rise to the top and a sharp descent down, followed by a flat line—it's time to think about why you might have bored your audience.
Another problem that Todd and Rene noted turned out to be not so obvious, but very interesting: supply and demand in the market.
If viewer demand for a topic is high, but there are few video offerings from creators, then all views will go only to the lucky ones who still made videos on the topic.
But as soon as other bloggers realize where the gold mine is and start bombing content on this topic, the views will scatter across a large number of offerings, and then demand will wither altogether. This is essentially how trends work.
This is the hardest thing to track: you'll have to constantly keep an eye on what's happening with competitors and in general in your niche. But if you have to choose which analytics to spend your time on—this one, or the one in Creator Studio, then choose competitor analysis.
Because it's quite an art—to be able to spot some in-demand topic in its infancy and grab all those millions of views before other creators notice it.
And then the conversation between the two specialists turns to a completely different topic:
The 'Subscriptions' tab is the only way to objectively assess how your audience reacts to your content without the involvement of algorithms.
In the "Subscriptions" tab, there are no recommendation systems, and videos are simply arranged in the order they appear on the platform. Therefore, the views that come to you from this tab will tell you more about your audience than any data from the home page.
Additionally, this tab is the only area that creators can control 100%.
Todd advises us to compare the CTR of different videos from the "Subscriptions" tab, and then objectively compare metrics. If the CTR has dropped somewhere, you can think: is it about the thumbnail, title, intro, or video topic?
This kind of analysis helps understand whether you're creating quality content that viewers like, which is more important in the long run as it helps retain viewers on your channel, or if you're just being pushed by algorithms.
And all this sounds wonderful, logical, and great, except that YouTube doesn't have a parameter in analytics like "views from the 'Subscriptions' tab."
When analyzing traffic sources, we have an item called "Notifications," which might somewhat resemble this tab, but it's still not the same thing.
And in the "Audience" tab, there's an item "watch time by subscribers," which doesn't mean that your subscriber watched the video from the "Subscriptions" tab. It simply indicates that they're subscribed to the channel.
And the last thing one might consider is returning viewers, but again this doesn't equal "subscribers" or "views from the 'Subscriptions' tab." These are just viewers who visited your channel during the selected period and visited it again.
So the idea is great, of course, but in reality, it doesn't exist.
At the end, Rene makes an obvious but correct point: some topics interest a narrow audience, and others interest a broader one, and this affects reach.
YouTube is incorporating more Large Language Models (LLM) into its algorithms
Now we've come to a complex but important topic that could significantly change how recommendation systems work.
But first, let's understand the basic concepts.
There are Small and Large Language Models used for Artificial Intelligence work—abbreviated as SLM and LLM.
A Language Model is a software algorithm that analyzes and generates text based on studied material. So for AI to give you answers, it needs to learn and analyze a huge number of parameters.
Accordingly, a Small Language Model works with fewer parameters, and a Large one—with more.
And if the Small one processes significantly fewer parameters, the Large one is capable of working with an almost infinite amount of materials and digging much deeper into the essence.
To make it clearer, Todd tells us that if you've worked with Gemini, Chat GPT, and similar chatbots, you're already familiar with a Large Language Model.
Now Large Language Models are being widely implemented in YouTube's recommendation systems to make selections for viewers even more relevant.
It works as follows: LLM analyzes not just the video topic, but all its content, format, context, and even mood.
For example: imagine a viewer who is interested in cars, or more specifically—car reviews.
If a Small Language Model could recommend videos from all bloggers who do car reviews, a Large Language Model can detect that this particular viewer doesn't like serious reviews, but only humorous ones.
Therefore, in recommendations for them, it will show only these kinds of bloggers, not all reviewers indiscriminately.
For both viewers and content creators, this is excellent news, as YouTube managers promise us that content and its target audience will be able to find each other faster and more easily.
However, neither Todd nor Rene gave any recommendations on how creators should better work with their channels now, when these rather global algorithm changes and new AI capabilities are being implemented.
From their entire dialogue, we would draw the following conclusions:
- Recommendation systems work for viewers, not creators, so we, as bloggers, need to study our target audience, not how algorithms work
- Algorithms show the most effective videos, not those that got the most views
- Recommendation systems work on the principle of "word of mouth": viewers with similar interests will be recommended similar content
- Analytics of individual parameters won't help us understand which videos are most successful, we need to look at analytics comprehensively
- Analyze and evaluate your own content depending on your goals
- We cannot 100% influence which audience YouTube will show our videos to
- New algorithms will analyze the use of audio tracks in different languages and build recommendations taking this into account
- Ups and downs in views are a natural process, and we need to analyze why this happens in order to influence the situation
- The implementation of Large Language Models will make content recommendation systems more accurate, but how creators should act in this case is still unclear
This is quite an interesting and ambiguous interview. Not all of its points can be applied in practice right now. But knowing the direction YouTube is setting for 2025, it will be much easier for us in the long term to detect algorithm changes and adapt to these changes when possible.
