Keywording a photo library people can actually search

Every image library dies the same death. Not from a crash — from a search box that returns nothing, until people stop trusting it, stop using it, and go back to messaging a colleague who might remember where the photo is.

The library is still there. The images are still in it. They are just unreachable, and an unreachable asset is indistinguishable from one that does not exist.

The fix is keywording, and keywording is unglamorous work that nobody schedules. This guide is about doing it in a way that pays off — and about the parts you can skip.

Keyword for the search, not for the image

The instinct when keywording a photo is to describe it. This is the wrong instinct, and it is why so many keyword sets are useless.

You are not describing the image. You are predicting the search that will one day need to find it.

Those are different jobs. Faced with a photo of a woman presenting to four colleagues in a meeting room, the describing instinct produces: woman, laptop, whiteboard, chairs, window, coffee cup. Every one of those is accurate. Not one of them is a search anyone will ever run.

The predicting instinct asks: who will come looking for this, and what will they type? The answers are things like presentation, team meeting, collaboration, workshop, office culture — plus, crucially, the things your organisation specifically needs: the campaign it was shot for, the client, the year, whether there is a usable copy space on the left for a headline.

The coffee cup is in the picture. Nobody will ever search for it.

The four layers worth tagging

A keyword set that works tends to have four layers. Most bad keyword sets have only the first.

1. What is literally in it. beach, dog, laptop. This is the layer AI does well and humans find boring — which makes it the perfect thing to automate. It is also the least valuable layer, but you do need it, because it is what people type when they have no better idea.

2. What it is about. remote work, sustainability, celebration, failure. The concept, not the content. This is where most real searches live, and it is largely invisible to a classification model: a photo of an empty office is about remote work, and no amount of pixel analysis will tell you that.

3. What it is for. hero image, copy space left, vertical crop, low contrast background. Purely practical, entirely ignored by most people, and adored by whoever has to actually build the page. If a designer can search for “vertical, copy space, muted” and get twelve usable options, your library has justified its existence.

4. Where it came from. Shoot, campaign, client, photographer, date, usage rights. This is the layer that prevents legal problems, and it is the one nobody can reconstruct later. If you tag nothing else, tag this, because it is the only layer that becomes impossible to recover once the photographer has moved on.

Controlled vocabulary, or the library rots anyway

Two people tagging the same library will produce car, cars, automobile, and vehicle. Six months later, searching car finds a quarter of your car photos, and everyone concludes the search is broken. They are right.

A controlled vocabulary is just an agreed list of terms and the rule that you only use those terms. It sounds bureaucratic and it is the difference between a library and a pile.

You do not need a formal taxonomy to start. You need:

• One term per concept. Pick car, ban automobile. Write it down.
• A consistent number. Plural or singular, pick one, forever. (Plural, usually — people search cars.)
• A rule for new terms. Anyone can propose one; someone has to approve it. Without that gate, the vocabulary is not controlled, it is just a suggestion.

This is also the point where AI keywording and your taxonomy collide, and it is worth being clear-eyed about it: a general-purpose model uses its vocabulary, not yours. It will happily return sports car when your controlled term is vehicle — passenger. Which is why AI tags should enter your workflow as candidates, never as final values — a distinction that the tools promising “fully automatic tagging” tend to gloss over.

How to clear a backlog without losing a month

The backlog is the thing that makes people give up. Forty thousand untagged legacy assets is not a task, it is a project with a budget, and it never gets one.

The way through it is to stop trying to tag everything perfectly and instead split the work by confidence.

An AI tagging pass gives you, for every image, a set of proposed keywords with a confidence score. That score is not a curiosity — it is the tool that makes the backlog tractable:

1. Auto-accept the confident half. Set a threshold — 80% is a reasonable starting point — and accept everything above it in one action. The model is rarely wrong when it is that sure, and you have just tagged half the library without looking at it.
2. Review the uncertain half. One image at a time, quickly. You are not writing, you are judging: yes, no, yes, no. A few hundred images an hour is a realistic pace.
3. Add your layers. The AI gave you layer 1. You add layers 2, 3 and 4 — the concept, the usage, the provenance — which is where your knowledge is actually worth something.

The result is not a perfect catalogue. It is a catalogue that exists, which beats the perfect one that does not.

You can do exactly this in PicsTag: drop in a folder or a CSV of image URLs, run the tagging pass, auto-accept above a threshold, review the rest one image at a time, and export to CSV or JSON for your DAM. It runs entirely in your browser, so a backlog of client or unreleased assets never has to be uploaded anywhere — which, for a lot of libraries, is the only reason this is possible at all.

What to do this week

If your library is already in trouble, do not start with a taxonomy project. Start with the images people are actually failing to find.

Look at your search logs — the queries that returned nothing. That list is your keywording brief, ranked by demand, and it will be much shorter than you fear. Tag for those searches first. Then run an AI pass over the rest to give everything a baseline, so that the search box at least returns something for every reasonable query.

Perfect metadata on 100% of the library is a fantasy. Adequate metadata on all of it, and good metadata on the part people actually want, is a week of work — and it is the difference between a library and a haystack.