July 2026 · 20 min read
VISID - Small Test. Interesting Signal.
We ran a controlled test on Pinterest using six matched image pairs - one stamped with VISID structured metadata, one without. The results were uneven, unexpected in places, and worth documenting honestly.
What we tested
During the 2026 FIFA World Cup, we published six pairs of soccer images to a Pinterest account. Each pair contained the same scene: one version stamped with VISID metadata - title, description, subject keywords, creator, rights, AI training permissions - and one version with no embedded metadata. All images were generated in Midjourney - which matters, because AI-generated images carry no inherent EXIF or authorship data, making them a clean baseline: the only metadata present was what we deliberately added.
To prevent Pinterest's duplicate detection from collapsing the pair into a single pin, the unstamped version of each image had 30 pixels cropped from the top. This was enough to register as a distinct image while keeping the scene visually identical for practical purposes.
Both images in each pair were posted to the same account, to the same board ("Beautiful Imagery"), at the same time, with identical captions and hashtags. Every pin - stamped and unstamped - carried this description:
"Beautiful game. Beautiful imagery. Beautiful imagery deserves more than a like. It deserves context. Give your images the context they deserve."
#visid #metadata #photography #visualcontent #aimarketing
The stamps were written with the World Cup context in mind. Most carried "World Cup 2026" explicitly in the title, description, and keywords - making the event a deliberate signal in the embedded metadata, not just background noise. The accordion under each stamped image shows the exact fields used. The only variable was the presence or absence of that structured metadata embedded in the file itself. We measured impression counts over 60 days using Pinterest Analytics. Here are the six scenes.
The Net
soccer_01_c
soccer_01_stamped
The Pitch
soccer_02_c
soccer_02_stamped
The Ball
soccer_03_c
soccer_03_stamped
The Boot
soccer_04_c
soccer_04_stamped
The Crowd
soccer_05_c
soccer_05_stamped
The Goal
soccer_06_c
soccer_06_stamped
The results
Impression counts from Pinterest Analytics over 60 days. Each block shows the Analytics screenshot alongside the raw numbers.
2,930 impressions - unstamped/5,060 impressions - stamped
1.7x difference
Consistent but modest. The stamped version held a steady lead across the full 60-day window.
536 impressions - unstamped/6,320 impressions - stamped
11.8x difference
A strong and widening gap. The pitch is a clear subject - metadata appears to have reinforced what the algorithm was already seeing.
89 impressions - unstamped/16,600 impressions - stamped
186x difference
The outlier. Three blurred children, a makeshift ball of plastic bags tied together, somewhere in Africa - visually rich but categorically ambiguous. The stamped version carried an explicit declaration - soccer, World Cup, sport - and the algorithm had something concrete to work with.
75 impressions - unstamped/1,110 impressions - stamped
14.8x difference
Strong and consistent. The boot is a recognisable object - metadata reinforced what the algorithm likely already suspected.
2,070 impressions - unstamped/83 impressions - stamped
24.9x difference
The only reversal. The unstamped crowd image outperformed by a wide margin. We have a theory - see below.
310 impressions - unstamped/1,160 impressions - stamped
3.7x difference
Solid and steady. The goal frame is a clear compositional signal - metadata compounded it.
| Scene | Unstamped | Stamped | Ratio | Winner |
|---|---|---|---|---|
| The Net | 2,930 | 5,060 | 1.7x | stamped |
| The Pitch | 536 | 6,320 | 11.8x | stamped |
| The Ball | 89 | 16,600 | 186x | stamped |
| The Boot | 75 | 1,110 | 14.8x | stamped |
| The Crowd | 2,070 | 83 | 24.9x | unstamped |
| The Goal | 310 | 1,160 | 3.7x | stamped |
What Pinterest actually says about this
Before getting to theories, it's worth being honest about what is and isn't known from Pinterest's own documentation and engineering literature.
Pinterest has never confirmed that embedded file metadata - EXIF, IPTC, or XMP fields like title, description, and keywords - affects pin distribution or ranking. Their developer documentation focuses entirely on Open Graph and Schema.org markup in page HTML, not on embedded image metadata. Their engineering papers - including published research on search relevance, visual search, and home feed ranking - list their signals consistently as pin titles and descriptions, AI-generated visual captions, board context, engagement history, and graph embeddings. Embedded file metadata does not appear in any of them.
There is also no published independent research specifically testing embedded metadata against Pinterest performance. No controlled study. No credible marketing experiment. Nothing to cite in either direction - that we could find.
What is confirmed: in March 2025, IPTC announced that Pinterest had implemented support for the IPTC DigitalSourceType field. When Pinterest detects the value trainedAlgorithmicMedia in that field, it displays a visible "AI Modified" label on the pin. This is a confirmed, documented case of Pinterest reading an embedded IPTC field and acting on it. Their infrastructure can read embedded metadata. Whether they read descriptive fields like title, description, or keywords - and whether those fields influence distribution - they have never said.
One more practical note: a 2015 test by embeddedmetadata.org found that Pinterest strips IPTC and XMP from downscaled versions of uploaded images, retaining them only in full-resolution downloads. Whether Pinterest's ML ingestion pipeline reads the metadata before or after compression - before it would be discarded - is unknown.
All of this is the honest backdrop: Pinterest has never confirmed embedded metadata helps, no one has published a controlled test, and yet their infrastructure demonstrably reads at least some IPTC fields. We ran this test into that gap.
What we think is happening
The cold start problem - and what happens when a wave arrives
When a new image lands on Pinterest, the algorithm has a decision to make: where does this belong? It reads the pin caption, the board context, runs visual analysis, and - if it's there and readable - may consider embedded metadata in the file. That categorisation happens fast, near the moment of upload. For images where the visual content is clear and unambiguous, any additional signal matters less. For images where the subject is ambiguous, an embedded signal could be the one that breaks the tie - placing the image into a semantic bucket before it has any engagement history to speak for it.
The Ball (soccer_03) makes this concrete. The image shows three blurred children with a makeshift ball made from plastic bags tied together, shot somewhere in Africa - visually rich but categorically ambiguous. It could be poverty, documentary, childhood, Africa, street photography. The stamped version declared: this is soccer, this is football, this is sport. The unstamped version offered no such signal. The algorithm had to guess.
Looking at the Analytics graphs, the impression spikes on the stamped images weren't gradual - they were explosive, landing almost exactly during the World Cup window. This is where the cold start and the event meet: the stamped images were already sitting in the right semantic buckets when search and feed volume for "World Cup 2026" peaked. The unstamped counterparts largely weren't - they were never categorised in time, and the wave passed them by.
It's not a compounding-over-time story. It's a right-place-right-time story, where the "right place" was declared in the metadata before the event arrived. What a more evergreen test would show - images stamped with no natural search spike to catch - is an open question, and one worth running.
The Crowd reversal
The Crowd (soccer_05) broke the pattern entirely - the unstamped version outperformed by nearly 25x. We don't have a clean explanation, but we have a theory: the crowd image is visually rich and self-explanatory. Faces, colour, energy, scale. An algorithm looking at that image has plenty to work with without any additional signal.
In this case, our metadata may have actually narrowed the algorithm's interpretation - flagging it as a soccer image when the underlying visual could have surfaced across sport, events, celebration, culture, and travel feeds. By declaring intent, we may have constrained distribution.
This is the most honest finding in the test. Declared intent is valuable - but it's a specific kind of signal. It's the creator's statement about what the image means. That isn't always the same as what an algorithm can discover on its own.
What this doesn't mean
This was six image pairs on a single account, in a single category, during a major global sporting event driving search and feed volume in unpredictable ways. We are not claiming that VISID metadata improves Pinterest impressions by 186x. We are saying this signal is worth paying attention to.
Six pairs is not a sample size. It's a starting point.
We can't isolate the metadata variable perfectly. Pinterest's algorithm responds to engagement history, account authority, image quality scoring, and timing factors we can't fully control.
The World Cup context almost certainly inflated impression counts across the board. These numbers will not reproduce in a quieter category.
The Crowd reversal is a real result. Declaring a narrow context on a visually rich image can constrain rather than expand its reach - keyword strategy matters as much as having metadata at all.
Pinterest has never confirmed they read descriptive IPTC or XMP fields for ranking. The mechanism we're proposing is a hypothesis, not a proven causal chain.
A note on image format and metadata survival
All images in this test were JPEG. When we verified the pinned versions of the stamped soccer images through VISID View, the title, description, and keywords were present in the files Pinterest stored and served. The metadata survived the pin.
That said, we've observed inconsistency across other images we've tested on Pinterest - some carry title, description, and keywords through; some carry only keywords; some carry nothing. The variables aren't fully understood. Image resolution, upload method, file quality, and how the image was pinned all appear to play a role. We're not in a position to make a clean rule here.
What we can say is that for the images in this specific test, the metadata was present in what Pinterest served. If it's in the file Pinterest distributes, we believe it travels with every re-pin - carrying the structured identity of the image forward without any further action from the creator. We haven't been able to verify this directly yet, but it's a hypothesis worth testing.
One technical observation from verifying a pinned image through VISID View: the file returned a warning - IPTCDigest is not current. XMP may be out of sync. That warning indicates Pinterest modified the file at ingest - rewriting or stripping part of the metadata - without recalculating the IPTC checksum. Pinterest is not passing the file through unchanged. That may explain why some fields survive and others don't, and it suggests the inconsistency we're seeing is Pinterest's processing, not the original embed.
Why we ran this
VISID embeds structured metadata into images - title, description, keywords, authorship, rights, AI training permissions. The pitch is that this makes images more readable to machines, which makes them more discoverable, more attributable, and more durable as assets.
That's a reasonable argument to make theoretically. The research literature doesn't contradict it - but it doesn't confirm it either. Pinterest hasn't said. Nobody has tested it properly. We wanted to see if there was any observable signal in practice, in the wild, on a real platform, with real images.
Pinterest is also worth considering as a test environment in its own right. It's a closed platform - no page HTML, no JSON-LD in a document head, no surrounding body copy for a crawler to read. The embedded file metadata is close to the only structured signal you control. If it moves the needle there, in that constrained environment, what happens on the open web - where the same embedded metadata sits alongside a JSON-LD schema in the page head, a matching title tag, alt text, and surrounding copy all saying the same thing? The corroborating signals compound. Pinterest is the harder test. The open web should be easier.
Small test. Interesting signal.
We'll keep running these. Follow-up test coming soon - we'll keep you posted.
VISID embeds structured intelligence into images - giving them authorship, readability, and ethical visibility in an AI-native internet. Learn more at visid.app.