I've tested almost every metadata approach over the years - manual keywording, mixed tools, AI assistants, automatic generators, cross-checking with agency data, etc.
In the end I realised something simple: no single AI tool can produce consistently accurate metadata for stock video without a human layer.

Today our workflow looks like this:

a keyword specialist reviews the footage manually

they use our internal system built on 15+ years of sales patterns (so the suggested words are based on real historical buyer behaviour, not generic AI guesses)

the specialist then writes the title, description and keyword list manually, combining:
- what's actually happening in the clip
- agency rules and limits
- long-term sales data
- client's priorities for a specific topic

after that, the metadata goes through an automatic QC inside our system
(duplicates, forbidden terms, word order issues, irrelevant concepts, compliance checks)

This hybrid approach - human expertise + data-driven suggestions + automated quality control - turned out to be far more reliable than fully manual work or fully automated tools alone.

I'm curious how others here approach metadata in 2025.
Do you rely on AI, assistants, presets, or full manual work?

If there's interest, I can also share some insights we recently found about how metadata affects bonuses and overall performance.

Very interesting, thanks.
I'm using not specific tool but chatgpt to title and keyword, and I agree with you, AI it's useless without human add and review.
It's very good to have a list of relevant keywords, but you need time to review, instruct AI about the exact focus, and generate several list before the final one.

By the way, I have to tell that in many years I haven't find a simple way to produce some useful scheme for standard keywording. It simply seems quite random when you see two similar images, or clips, with the same keywording scheme and several identical keywords, one on the top of best sellers and the other lost in the database.
If you have different experience and, as you said, "some insights about how metadata affects bonuses and overall performance", well that would be really interesting to know

I manage (trim/color correct/ship) several stock portfolios that aren't my own so I have to deal with a ton of uncategorized footage, especially drone footage...and dealing with metadata is boring AF. My goal with my workflow it make it so that I can spend 100% of my time making creative decisions only.

So I made a custom tool that can do offline title/descriptions/keywords purely from metadata or it can incorporate that info when doing rich AI descriptions and I encourage my clients to film with things like telemetry/GPS tagging enabled.

The only time I spend dealing with metadata is auditing title/descriptions/keywords but my tool makes that very fast and as it improves the need to audit is becoming moot.

Quote from: derby on January 13, 2026, 16:46
Very interesting, thanks.
I'm using not specific tool but chatgpt to title and keyword, and I agree with you, AI it's useless without human add and review.
It's very good to have a list of relevant keywords, but you need time to review, instruct AI about the exact focus, and generate several list before the final one.

By the way, I have to tell that in many years I haven't find a simple way to produce some useful scheme for standard keywording. It simply seems quite random when you see two similar images, or clips, with the same keywording scheme and several identical keywords, one on the top of best sellers and the other lost in the database.
If you have different experience and, as you said, "some insights about how metadata affects bonuses and overall performance", well that would be really interesting to know

I've seen this from the inside quite a lot. When you compare AI‑generated keyword sets with ones built by someone who actually looks at buyer behavior over time, the difference isn't in the amount of keywords  it's in the focus. Humans tend to cut things out, AI tends to add things "just in case". That small difference adds up in performance.
We've been comparing these approaches lately, especially in relation to bonuses and long‑term results. Would you be interested in that kind of comparison?

Quote from: LithG on January 13, 2026, 22:23
I manage (trim/color correct/ship) several stock portfolios that aren't my own so I have to deal with a ton of uncategorized footage, especially drone footage...and dealing with metadata is boring AF. My goal with my workflow it make it so that I can spend 100% of my time making creative decisions only.

So I made a custom tool that can do offline title/descriptions/keywords purely from metadata or it can incorporate that info when doing rich AI descriptions and I encourage my clients to film with things like telemetry/GPS tagging enabled.

The only time I spend dealing with metadata is auditing title/descriptions/keywords but my tool makes that very fast and as it improves the need to audit is becoming moot.

That's a familiar situation. When metadata is treated as a technical task, automation works great. But what I've noticed is that the moment priorities come in - what to emphasize, what to downplay, what not to include - that's where human judgment still beats automation. The most efficient setups I've seen use tools to remove friction, not to replace decisions.
Do you feel that balance shifting as your tool improves?

Quote from: videostock.system on January 16, 2026, 16:36

Quote from: LithG on January 13, 2026, 22:23
I manage (trim/color correct/ship) several stock portfolios that aren't my own so I have to deal with a ton of uncategorized footage, especially drone footage...and dealing with metadata is boring AF. My goal with my workflow it make it so that I can spend 100% of my time making creative decisions only.

So I made a custom tool that can do offline title/descriptions/keywords purely from metadata or it can incorporate that info when doing rich AI descriptions and I encourage my clients to film with things like telemetry/GPS tagging enabled.

The only time I spend dealing with metadata is auditing title/descriptions/keywords but my tool makes that very fast and as it improves the need to audit is becoming moot.

That's a familiar situation. When metadata is treated as a technical task, automation works great. But what I've noticed is that the moment priorities come in - what to emphasize, what to downplay, what not to include - that's where human judgment still beats automation. The most efficient setups I've seen use tools to remove friction, not to replace decisions.
Do you feel that balance shifting as your tool improves?

In general, my tool generates better descriptions/keywords than I can do manually.

It will let you add you add any context you want before using AI, so for stuff like character names you can just provide them. After using AI, very rarely will I have to drag a different keyword to the top, but my tool is pretty good at identifying the main subjects and has refinement steps built into to re-write descriptions and re-order keywords if they are too far off or don't include necessary context. It's also very fast for ripping through descriptions and modifying descriptions/keywords manually along the way.

That being said, I don't build keywords and descriptions based on market analysis so I don't know how well they sell. My prompt system emphasizes literal context and only starts to get flowery when you try to generate large descriptions or long lists of keywords and it has to get creative.

Quote from: LithG on January 16, 2026, 19:33

Quote from: videostock.system on January 16, 2026, 16:36

Quote from: LithG on January 13, 2026, 22:23
I manage (trim/color correct/ship) several stock portfolios that aren't my own so I have to deal with a ton of uncategorized footage, especially drone footage...and dealing with metadata is boring AF. My goal with my workflow it make it so that I can spend 100% of my time making creative decisions only.

So I made a custom tool that can do offline title/descriptions/keywords purely from metadata or it can incorporate that info when doing rich AI descriptions and I encourage my clients to film with things like telemetry/GPS tagging enabled.

The only time I spend dealing with metadata is auditing title/descriptions/keywords but my tool makes that very fast and as it improves the need to audit is becoming moot.

That's a familiar situation. When metadata is treated as a technical task, automation works great. But what I've noticed is that the moment priorities come in - what to emphasize, what to downplay, what not to include - that's where human judgment still beats automation. The most efficient setups I've seen use tools to remove friction, not to replace decisions.
Do you feel that balance shifting as your tool improves?

In general, my tool generates better descriptions/keywords than I can do manually.

It will let you add you add any context you want before using AI, so for stuff like character names you can just provide them. After using AI, very rarely will I have to drag a different keyword to the top, but my tool is pretty good at identifying the main subjects and has refinement steps built into to re-write descriptions and re-order keywords if they are too far off or don't include necessary context. It's also very fast for ripping through descriptions and modifying descriptions/keywords manually along the way.

That being said, I don't build keywords and descriptions based on market analysis so I don't know how well they sell. My prompt system emphasizes literal context and only starts to get flowery when you try to generate large descriptions or long lists of keywords and it has to get creative.

That makes sense. If the tool is fast and accurate, that already solves a big part of the problem. From what I've seen, AI is usually very good at figuring out what's in the shot.
Where it starts to get tricky is when the goal is sales, not just correct descriptions. Metadata can look perfectly fine, but still miss how buyers actually search or what agencies tend to push over time. That difference is small, but across a big portfolio it adds up.
Have you ever checked how this stuff performs long‑term, not just how "right" it looks at upload time?

I don't see an easy way to test how they perform without being able to compare them to other peoples sales data. The only metric I have is doing my own searches and seeing where my shots land on in the results. The libraries I work with are very niche, dominate the search results for their given subjects and sell very well. It's hard to say whether those sales are generated by good keywording or just having the best or only shots of their respective subjects.  For AI stuff, it's a numbers game and most shots don't sell at all, but the ones that do sell, sell a lot.

How do you test your results? Do you just have enough clients that provide you with sales data that you can measure? Is there a paid service that provides market insights?

Quote from: LithG on January 17, 2026, 21:21
I don't see an easy way to test how they perform without being able to compare them to other peoples sales data. The only metric I have is doing my own searches and seeing where my shots land on in the results. The libraries I work with are very niche, dominate the search results for their given subjects and sell very well. It's hard to say whether those sales are generated by good keywording or just having the best or only shots of their respective subjects.  For AI stuff, it's a numbers game and most shots don't sell at all, but the ones that do sell, sell a lot.

How do you test your results? Do you just have enough clients that provide you with sales data that you can measure? Is there a paid service that provides market insights?

Good question. We've been working with stock contributors since around 2010, and with clients directly since 2014 - some of them have been with us for many years. Over time, some tried switching to AI‑only workflows or paused stock because of other projects.
Since we still handle uploads for many clients, we continue to see real portfolio performance and actual sales data. What we've noticed is fairly consistent: files keyworded with human review usually perform better than fully AI‑generated metadata. Some clients saw that difference themselves and came back, others didn't - that's part of the reality too.
At the moment we don't offer market forecasting or "this will sell" analytics, mainly because there are many nuances and the same data can be interpreted very differently. If someone wants to discuss that side in more detail though, we're always open to a conversation.

I have been training a custom GPT for my keywords. So far it is doing great. Takes a lot of work to get right

Quote from: jjneff on January 19, 2026, 18:47
I have been training a custom GPT for my keywords. So far it is doing great. Takes a lot of work to get right

Yeah building prompts for keyword and descriptions is an interesting process. You would think that clean and simple would work but in my experience, lot's of redundancy and re-enforcement works better. My prompts looks like a mess, with some rules described in slightly different ways in multiple times in different sections. But, if I try to eliminate redundancies, important stuff gets left out or misinterpreted. I think a lot of people who give up on AI just think it's all about making the perfect prompt and getting the correct results on the first try, when it doesn't they just say AI sucks and give up. It's about feedback/refinement/iteration. It's about finding the models that give good results without costing $$$. You got to put in some work and testing to get things right and it will save you massive amounts of time doing busy work on the back end.

Yeah, I agree with you here, AI workflows are rarely about a "perfect prompt". Iteration, feedback and refinement matter way more, and most people give up too early.
Where I see a difference in practice is what the iteration is based on. Prompt tuning and visual checks help you get cleaner and more accurate metadata, but they don't always answer whether that metadata matches real buyer behavior or agency incentives over time.
In our case, iteration usually comes from comparing how similar files actually perform across different portfolios, not just how they rank in search or how "correct" they look. Both approaches work,  they just optimize for slightly different things.
Out of curiosity, have you ever tried comparing AI‑optimized sets against market‑adjusted ones on the same type of content?

Up until last year, i manually described and keyword videos/images. It took up a ridiculous amount of time and I'll never do it again. My manual keyworded shots often shared base templates (based on subject matter) rather than having every shot completely keyworded from scratch. This created some commonality that I'm sure hurt me when it came to search results. My AI described shots consistently do better than those, but it could be simply because my AI described shots are most recent.

One of my clients literally hired me because a company hired to do manual keywording was doing an awful job. I manually keyworded those shots for two years before switching to my custom AI solution and allowing me to focus all of my time on creative decisions which increased my output and income substantially. I have absolutely no desire to pay a human being to write my descriptions.

I don't understand how a human would be better at interpreting sales statistics and applying that information to keywords/descriptions. This is an algorithmic/database problem that doesn't even need AI. Seems like a human would just introduce human errors.

I'm curious how you determine the sales of your clients. Do they give you account access? Do they send you their sales results?

Quote from: LithG on January 21, 2026, 19:19
Up until last year, i manually described and keyword videos/images. It took up a ridiculous amount of time and I'll never do it again. My manual keyworded shots often shared base templates (based on subject matter) rather than having every shot completely keyworded from scratch. This created some commonality that I'm sure hurt me when it came to search results. My AI described shots consistently do better than those, but it could be simply because my AI described shots are most recent.

One of my clients literally hired me because a company hired to do manual keywording was doing an awful job. I manually keyworded those shots for two years before switching to my custom AI solution and allowing me to focus all of my time on creative decisions which increased my output and income substantially. I have absolutely no desire to pay a human being to write my descriptions.

I don't understand how a human would be better at interpreting sales statistics and applying that information to keywords/descriptions. This is an algorithmic/database problem that doesn't even need AI. Seems like a human would just introduce human errors.

I'm curious how you determine the sales of your clients. Do they give you account access? Do they send you their sales results?

Since we not only key but also manage many clients on a turnkey basis (cut and prepare videos and also upload), we see their income.
Of course, not all clients are 100% supported by us - for some we only key or cut videos and do not have access to their accounts. However, we periodically conduct surveys, ask about our work. Some clients themselves complain about how some mistake was made by AI and pass it on to us, 100% artificial intelligence can do something and sometimes correctly sees the effect on the video and object. However, the main thing that artificial intelligence does not have is that it does not have sales data, and does not see as a buyer. Many will disagree with this information and say - I am satisfied with how it does. This is the same as food prepared without attention. automatically, and food that is monitored by a chef. Both are food, and you can eat your fill, but in the second option it would make sense... But turn it not into food, but into something else. Because food from a chef is already a luxury, and here it needs to be presented as the correct algorithm for increasing sales

Quote from: videostock.system on January 26, 2026, 12:59
However, the main thing that artificial intelligence does not have is that it does not have sales data, and does not see as a buyer.

What's stopping it from having access to sales data or building meta from a sales perspective? How do your human beings apply sales data when writing manual keywords? How do you avoid human errors? How do you balance keywords applied to multiple competing portfolios? This would all be easy to automate with an AI.

Quote from: LithG on January 26, 2026, 19:23

Quote from: videostock.system on January 26, 2026, 12:59
However, the main thing that artificial intelligence does not have is that it does not have sales data, and does not see as a buyer.

What's stopping it from having access to sales data or building meta from a sales perspective? How do your human beings apply sales data when writing manual keywords? How do you avoid human errors? How do you balance keywords applied to multiple competing portfolios? This would all be easy to automate with an AI.

Just to clarify how it works on our side.

First - our AI does not have direct access to client dashboards or live sales accounts. It only works with historical data we've collected earlier. From our perspective, that's the right and ethical approach.

Second - sales data is still used during manual keywording, but not in a "human guessing numbers" way. A request is made to our internal database, relevant keywords are returned, and then the specialist compares those keywords with what's actually happening in the footage. After 10+ years, this process becomes more about pattern recognition than intuition.

Third - human error is real, but interestingly, many actual sales happen on misspellings or imperfect keyword variations. Everything we write also goes through internal automated checks, so the human layer is not working alone.

Regarding competing portfolios, we're a bit lucky here. Most of our clients shoot different subjects and different niches, so internal competition is minimal.

And yes, all of this could theoretically be automated. The real question is cost vs return. In many cases, building and maintaining such systems costs more than the value they generate. Large agencies like Shutterstock almost certainly already run this kind of AI internally. We're a smaller company, but we can still work with sales data in a way that makes sense both for our business and for our clients.