AI search is the ability to find an asset by describing what it depicts, rather than by guessing the words somebody typed into a metadata field when they uploaded it. A photograph with a camera-generated filename and no tags used to be findable only by someone who already knew where it was; described instead, it can be found by anyone. This article covers how the search understands meaning, what gets described and by what, how results are ranked, and what it still cannot do.
The most useful thing to know first is that there is nothing to switch on and no second search field to learn. Meaning-based matching runs inside the ordinary search bar, alongside the keyword matching that was always there.
A product reference and the phrase "two people shaking hands in a bright office" both belong in the same box, and you do not have to decide in advance which kind of query you are making. That is deliberate, and the consequence is the point: there is no wrong mode to be in, and nobody has to be taught which one to pick.
A single search runs several kinds of matching at the same time and merges them into one ranked list:
Because these run together rather than one after another, an asset that matches strongly in one way still ranks well when it matches in no other. The exact code and the half-remembered impression are served by the same query, which is why there is no mode to choose.
Once a file has been processed and a preview generated, it is passed to a vision model that writes a description of what it contains: subject, setting, composition, dominant colours. That description is stored on the asset, included in search, and readable on the asset itself if you want to see what the platform thinks it is looking at.
It is worth being exact about what kind of text this is. It describes appearance, not significance. It will tell you that a photograph is a wide shot of a harbour at dusk with boats in the foreground. It will not tell you that the image was the hero of last spring's campaign, or that it is cleared for paid social, because neither of those is visible in the picture.
That division is the useful one, and it is the argument for both halves of the system. The description covers what a picture obviously shows and nobody would think to write down. Your metadata covers everything the picture cannot possibly show.
The practical effect lands hardest on the assets nobody ever described. Most libraries contain a large stock of files that arrived faster than anyone could catalogue them, and a backlog like that was never going to be tagged retrospectively — the work is real, nobody owns it, and it never becomes this quarter's priority.
Those assets become searchable without anyone touching them. It changes the arithmetic of an old library: content that was effectively lost, and was quietly being recommissioned because nobody could find it, comes back into circulation.
Merged results are then re-examined more carefully. A second model scores each candidate against your query while reading the asset's own text — its name, tags, assigned metadata and extracted document content — and candidates that do not hold up are dropped rather than left to pad out the list. Queries are also read for dates and reference codes, which are applied as constraints instead of being matched as loose text.
This is why metadata discipline still pays after AI search arrives, which is not the conclusion people expect. Recorded metadata is what the ranking reads, and it carries the things that decide whether an asset is usable at all: which campaign it belongs to, whether it was approved, which markets it is licensed for, when it expires.
AI search will find you the photograph of the harbour. Only your metadata knows whether you are allowed to publish it.
Access control is applied to the whole query, so every kind of matching is filtered before anything is ranked. Folder permissions, visibility levels, restricted metadata fields, and embargoed or archived assets are all accounted for at that point.
The ordering is the part that matters. Nothing is found and then hidden from you afterwards. An asset you are not permitted to see never competes for a place in your results, so it cannot surface as a blurred thumbnail, a stray total or a suggestion that something exists just out of reach.
Four limits are worth knowing before you rely on it:
Because meaning is being matched, the habits that make a good query are the opposite of the ones a keyword box teaches:
Meaning-based search depends on your workspace's search index having been prepared for it, and matching on images directly requires further setup. Keyword search works either way, so the change is invisible when it is not yet active. To check what is enabled for your workspace, contact your administrator or Data Dwell support.