AI keywording for stock footage: what it gets right and wrong

How AI keywording reads a clip

The models behind keywording tools are image models. OpenAI's vision guide lists PNG, JPEG, WebP and non-animated GIF as accepted inputs. Anthropic's guide for Claude lists JPEG, PNG, GIF and WebP, and says that for an animation only the first frame is used. Neither guide lists a video format.

So every tool that keywords video does roughly the same thing:

1. it pulls a few still frames out of the clip (often called keyframes);
2. it sends those images to a vision model with instructions;
3. it turns the answer into a title, a description and a keyword list, sometimes in a CSV for each agency.

Two consequences follow. The model sees only the frames it gets: one frame from a slow dolly shot looks like a still photo, and nothing that happens between two frames reaches it. And it does not read your file's metadata. OpenAI says its model "doesn't process original file names or metadata", and Anthropic says Claude does not receive metadata from images. The GPS position your drone recorded never reaches the model unless the tool adds it on purpose.

What auto keywording does well

On a clear, well-exposed frame, a vision model is quick at the literal layer of a keyword list: the main subject, the setting, the colours, the time of day, indoor or outdoor, the number of people, the framing. That is the part of keywording that takes the most typing and the least judgment, and it is where an AI keyword generator for stock saves real time.

The vendors document where their models are weaker, and several points apply directly to footage:

Where an AI keyword generator goes wrong

The errors that cost clips are rarely about the main subject. They come from words the model adds to sound precise, or to fill the list.

Precise places

A frame of a harbour at dusk does not prove which harbour. Shutterstock says location information "must be relevant and applicable" to the content, and Adobe's keyword examples show why precision matters: "London, England or London, Ontario". Add the real place yourself, from what you know about the shoot, and delete any city the model guessed.

Species and plant names

Shutterstock accepts scientific Latin names of plants and animals as an exception to its English-only rule. A wrong species name is still a keyword that does not describe the clip. Keep a species only if you can confirm it; otherwise stay at a level you are sure of, such as "heron" or "pine forest".

Brands, products and camera gear

Models recognise logos and products, and they write them down. For commercial content, Shutterstock says descriptions and keywords "cannot contain trademarks such as brand names or product names", and it bans specific camera information. Adobe tells you to keep trademarks, brand names and personal information out and to avoid gear-heavy terms in titles. A drone or phone model name breaks both rules at once.

The motion between frames

Buyers search footage by camera move and speed: pan, dolly, aerial, slow motion, time-lapse. A still frame rarely shows a pan, and it never shows a frame rate. Unless the tool compares several frames, add these words yourself.

Padding the list

Ask a model for 50 keywords and it will produce 50, including mood words and near-synonyms that drift away from the clip. Shutterstock counts the same word stem repeated across keywords as spam, and unrelated terms too. Pond5 says the best results come from 10 to 20 descriptive, precise keywords. More words are not better words.

What the agencies say about relevance and spam

None of the pages read for this guide sets a rule on who or what wrote the keywords. What they regulate is the result: does each word describe the clip, and is it repeated or stuffed.

The exact limits, including Adobe's three different keyword counts, are in keyword limits by agency. Rules on clips made with generative AI are a separate topic, covered in AI video policies.

Automatic keyword suggestions at the agencies

Before paying for a tool, look at what the agencies already do at upload. All three offer automatic keyword suggestions, and all three leave the decision to you.

Shutterstock's method has a specific trap. Suggestions drawn from visually similar images can carry words that described those other images: a place, an event or a concept that is not in your clip. Treat them as a list to pick from.

How to review AI keywords before you submit

A fixed order makes the review quick and catches the errors above. Do it with the clip playing, not with the frames the tool used.

1. Title or description first. Shutterstock asks for a sentence that answers who, what, when, where and why, under 150 characters in its editor. Check that it names the action, not only the first frame.
2. Delete what you cannot confirm. Places, species, events, a person's job or relationship: if the clip does not show it and you do not know it, it goes.
3. Remove brands, product names and camera or drone models (on Shutterstock, trademarks are allowed only in editorial metadata).
4. Merge word stems. Keep one of run, running and runner.
5. Add what the frames could not show: the real location, the camera move, slow motion or time-lapse.
6. Order for Adobe. Put the 10 most important keywords first. Adobe does not let you change the order while a clip is in moderation; see Adobe Stock keyword order.
7. Count against each limit with the free title and keyword checker.

For what a good list contains in the first place, layer by layer, see how to keyword stock footage. If you write keywords with a chat assistant instead of a dedicated tool, the method and a full prompt are in ChatGPT keywords for stock footage.

Where RushPilot fits

RushPilot is a desktop app for stock video contributors, in private beta. It does AI keywording for stock footage as one step of a longer job: it reads five frames of each clip, writes the title, description and keywords in English to each agency's limits (in order of importance for Adobe Stock), checks the common rejection reasons such as resolution, duration, visible logos and upscaling, then waits for you to approve the batch before it uploads to Shutterstock, Adobe Stock and Pond5.

It works from frames too, so the limits on this page apply to it as well: it cannot know a place or a species the frames do not prove, and the review stays yours. It makes no promise of acceptance. New accounts get 25 clips free; the offers are on the pricing page, and you can join the waiting list. If you need a tool today, the options are compared in keywording tools for stock footage.