Image tagging is the process of attaching descriptive keywords to a photo, illustration or video clip so that a search engine, or a stock agency's internal search, can retrieve it later. For a contributor it is the single step that decides whether a technically perfect frame earns money or sits unseen inside a catalogue of hundreds of millions of files. Get the tags right and buyers find the image. Get them wrong and the file effectively does not exist.
For years this was manual work: open a file, look at it, type twenty or thirty words, move to the next one. AI image tagging changes the economics of that step by having a vision model read the picture and propose the keyword set for you. What it does not change is who is responsible for the result. This guide covers how the technology actually works, where it is reliable, where it quietly fails, and how to decide which files still deserve human attention.
What image tagging actually means
Tagging and keywording get used interchangeably, but tags live in two separate places. On an agency they are a structured field attached to your submission. Inside the file they are stored as IPTC and XMP metadata that travels with the JPEG wherever it goes. Those two copies are not automatically the same, and contributors who only fill in the agency form lose their keyword work the moment they deliver the file somewhere else.
A working tag set does three jobs at once. It describes what is literally in the frame. It describes the concept a buyer is shopping for. And it covers the vocabulary variations those buyers actually type into a search box. A photo of a woman at a laptop in a kitchen is also remote work, freelancer, work from home and small business owner. Literal description alone is not enough, because the concept layer is where most of the commercial search volume sits. How you structure that set matters as much as how many terms it contains.
Titles and descriptions are part of the same job. Agency search reads them alongside the keyword field, and a title written as a comma-separated keyword dump is a documented quality problem at several platforms. The title is a sentence. The keywords are a set. Treating them as one undifferentiated blob is one of the fastest ways to look like a low-effort submitter.
How AI image tagging works under the hood
Modern automatic image tagging software runs on a vision-language model, the same class of system behind image captioning. Understanding the three stages it moves through is what tells you where to aim your review time.
Stage one is recognition. The model encodes the image into a numerical representation and matches it against concepts learned during training: objects, settings, actions, lighting, colour, apparent emotion. This is the part that has improved most dramatically over the past few years, and on ordinary commercial subjects it is now more consistent than a tired human at two in the morning.
Stage two is generation. Recognised concepts get turned into language: a title, a description, and a candidate keyword list. Good systems deliberately generate more candidates than you need and then rank them, because a ranked list can be truncated cleanly to fit different agency limits. A flat unranked list cannot, which is why some tools produce fifty words that all feel equally arbitrary.
Stage three is constraint, and this is where tools separate from toys. A raw model output is a list of words. A production tool has to enforce per-agency rules: Shutterstock accepts a minimum of seven keywords and a maximum of fifty, agencies reduce search visibility for repetitive or padded keyword sets, and every platform maintains its own banned and mandatory terms. Software that skips this stage hands you a compliance problem dressed up as a time saving.
There is a fourth stage most contributors forget: writing the result back into the file. Metadata that exists only in a web form is metadata you will retype. Metadata embedded in IPTC and XMP fields is metadata you own.
What automatic image tagging software gets right
Three things, consistently:
- Speed at volume. A shoot of 300 frames is a full working day of manual keywording and a few minutes of processing. That gap is the entire argument, and no amount of craft closes it.
- Consistency. Humans drift. By file 180 the keyword sets get shorter and lazier, and the last third of a shoot is routinely described worse than the first. A model does not get tired, so file 300 is treated exactly like file 1.
- Vocabulary recall. Most contributors reach for the same few hundred words over and over. A model proposes terms you would not have thought of, and incremental search visibility tends to come from exactly those long-tail entries rather than from the obvious ones.
None of this requires the model to have better judgement than you. It requires the model to be tireless, which it is.
Where AI tagging falls short
The failure modes are predictable, which is good news, because predictable failures are checkable failures. There are four.
- Named entities. Models guess at landmarks, brands and people. A confident wrong guess, a generic European tower labelled as a famous one, is worse than no tag at all, because a reviewer can verify it instantly and it becomes a rejection reason.
- Editorial context. Who, what, where and when are facts that live in your head or in the camera's GPS field, not in the pixels. No vision model can recover the date and location of an event from the frame alone.
- Intent and usage. The model sees a hand holding a card. It does not know you shot it for a fintech brief, and that concept is usually the one carrying the commercial search volume.
- Over-description. Left unchecked, models pad keyword lists with terms that are visually true but commercially useless. Fifty weak keywords now rank worse than twenty strong ones on platforms that actively measure keyword relevance.
The practical conclusion is a division of labour. The model owns recall, proposing everything plausible. You own precision, cutting what does not belong. That split is faster than either party working alone, and it is the difference between tools that help and tools that create cleanup work.
How to review an AI-generated tag set in sixty seconds
Review is only worth doing if it takes less time than it saves. A structured pass takes under a minute per file and catches almost everything above.
- Read the first ten keywords only. Keyword order carries weight on most platforms, and if the opening ten do not describe the commercial concept of the image, nothing further down will rescue it.
- Delete anything you cannot see. If a term is not visible in the frame or unambiguously implied by it, it is noise, and noise is what relevance scoring punishes.
- Delete every proper noun you cannot personally verify. Landmarks, cities, brands, people. If you were not there or are not certain, it goes.
- Add one concept the model missed. There is almost always one: the use case, the mood, the audience, the season it will be licensed for.
- Check the title reads like a sentence a person wrote, not a keyword list with spaces in it.
Contributors who run this pass consistently end up with shorter keyword sets than the model proposed, and better performing ones. Cutting is the high-value step, not adding.
When to use AI tagging, and when to tag by hand
A three-tier rule covers most catalogues.
- AI with a light review: commercial lifestyle, food, objects, generic backgrounds, textures, abstract work. High volume, low factual risk, nothing a reviewer can fact-check against the world.
- AI with a careful review: travel and architecture, where landmark accuracy decides acceptance; anything containing recognisable products; anything shot under a model or property release.
- Manual, or heavily edited: editorial and news, documentary work, and niche technical subjects such as medical, industrial or scientific imagery, where the correct term is a term of art the model will not reach for on its own.
The category that surprises people is AI-generated imagery. It needs more metadata discipline, not less, because agencies require disclosure at submission and because generated files carry no camera metadata to fall back on. Every descriptive field has to be authored on purpose.
Building an image tagging workflow that scales
Tagging is one link in a chain, and optimising it in isolation rarely helps. The full chain runs: cull, tag, embed the metadata into the file, format per agency, deliver, then track what actually sold. A tool that only handles the middle step leaves you exporting spreadsheets by hand, which is where the time you saved goes back out again.
That whole chain is what Rastock is built around. It reads a folder, generates titles, descriptions and keywords, applies per-agency rules including banned and mandatory terms, writes the result into IPTC and XMP so it stays inside the file, and delivers over FTP and SFTP to 10+ agencies including Adobe Stock, Shutterstock, Magnific, Alamy, Depositphotos, Pond5, Canva, 123RF, VectorStock and Vecteezy. Any other agency that accepts FTP can be added manually. Upload status is tracked per file, and a batch that goes out wrong can be rolled back rather than re-keyworded.
Two things it deliberately does not do: take a share of your royalties, or claim ownership of your files. IPTC and CSV export stay open, so nothing you produce is locked to the platform. The free trial runs 14 days and does not ask for a card.
The bottom line
Image tagging is not a job worth doing entirely by hand at volume in 2026, and it is not a job worth handing over entirely either. The model is better than you at recall and worse than you at judgement. Build the workflow around that asymmetry, let it propose and you dispose, and the hour you used to lose after every shoot turns into a few minutes of review.
Frequently asked questions
What is image tagging?
Image tagging is the practice of attaching descriptive keywords to a photo, illustration or clip so that search can retrieve it later. On a stock agency the tags sit in a structured field on your submission; inside the file itself they live as IPTC and XMP metadata. Both copies matter, because only the embedded one travels with the file when you deliver it somewhere else.
Is AI image tagging accurate enough for stock agencies?
For recognition, yes. Modern vision models identify objects, settings, actions and mood reliably. For factual claims, no. Landmarks, brands, named people and editorial context are where models guess confidently and wrongly. The workable split is to let the model propose the full candidate list, then spend your own time deleting what does not belong rather than typing everything from scratch.
How many keywords should a stock image have?
Fewer than the maximum, and all of them relevant. Shutterstock accepts a minimum of seven keywords and a maximum of fifty, and agencies now reduce visibility for repetitive or padded keyword sets. Twenty to thirty precise terms, with the strongest concepts first, outperform a padded list of fifty. Keyword order carries weight, so decide deliberately what occupies the opening slots.
Does AI tagging work for AI-generated images?
It works, and those files need more metadata discipline rather than less. Generated images arrive with no camera metadata to fall back on, and every major agency requires the content to be disclosed as AI-generated at submission. So the descriptive layer has to be authored deliberately, and the disclosure step has to happen every time rather than occasionally.
Do AI-generated tags get written into the file itself?
Only if the tool writes them there. Some services fill in an agency web form and stop, which means the keyword work is lost the moment you deliver the same file elsewhere. Look for software that writes titles, descriptions and keywords into IPTC and XMP fields inside the image, so the metadata stays attached wherever the file goes next.
Can automatic tagging get my submissions rejected?
Yes, in two ways. Irrelevant or spammy keyword sets are a documented rejection and visibility-reduction reason at most agencies. Incorrect proper nouns are worse, because naming a landmark, brand or person wrongly is a factual error a reviewer can verify instantly. Both are review problems rather than model problems, and both are caught by a short manual pass.