Skip to main content

AI owns the catalogue. It does not own the download ranking

The same generative-AI flag, counted twice: once over everything that exists for a subject, once over what actually gets downloaded.

Measured · SincereTrend

75.8%
AI share of the catalogue, for the median keyword
14.0%
AI share among that keyword’s most-downloaded works
1,198
keywords where both AI shares are measured

Every count of AI in a stock library measures the same thing: what has been uploaded. That is the shelf. It says nothing about what comes off it, and the two are not the same question — the first is about how much work exists, the second about what a buyer picks when shown it. Adobe happens to publish enough to ask the second one.

Its illustration results can be ordered by download count. Re-running an identical query that way and counting the same generative-AI flag over the first page gives the AI share of the most-downloaded work for that subject. Read against the catalogue share, the pair is a contrast no single ranking can produce.

The headline

Across 1,198 subjects where both readings exist, the median catalogue is 75.8% AI-flagged and the median download ranking is 14.0%. 82.9% of those subjects have a majority-AI catalogue; 8.6% have a majority-AI download ranking.

PercentileCatalogueBestsellers
25th57.0%7.0%
Median (50th)75.8%14.0%
75th87.8%27.3%
Adobe Stock illustration AI share, two rankings of the same results, across 1,198 subjects measured on both.

The widest gaps

Ordered by the distance between the two readings rather than by either one. The bottom of the bestseller distribution on its own is uninformative — dozens of subjects tie at 0-1%, and some of them have low catalogue shares too, so they show nothing. These do.

PhraseCatalogueBestsellers
chief96%3%
cold and flu93%1%
frontier94%3%
granddaughter98%8%
hardware96%6%
green meadows96%6%
blueprints95%5%
white horse98%9%
Each phrase links to its own keyword page, where both readings are free.

Where AI owns the ranking too

The obvious objection to everything above is that a download ranking is cumulative, so older work has an advantage from age alone. It is a fair objection, and this table is what survives it: if the figure only measured age, nothing could sit here.

What those subjects share is that the PHRASE is recent. "Chunky knit blankets", "muted winter palette", "biophilic design" are aesthetic terms whose demand appeared after generated work was already cheap, so there is no older catalogue for age to favour. Where a subject is genuinely new, generated work reaches the top of its ranking as readily as anything else.

By category

No category is anywhere near as generated at the till as it is on the shelf, and the ordering rhymes with the catalogue study: patterns and lettering least saturated, travel and people most.

CategoryKeywordsBestsellers (weighted)Bestsellers (median)Catalogue (median)
Lettering & Typography136.2%5.0%61.7%
Patterns & Backgrounds2910.3%5.0%47.0%
Technology & Internet2714.7%7.0%59.0%
Retro & Vintage159.7%9.0%71.1%
Nature & Landscape6415.1%9.0%67.7%
Education & Science2616.3%10.5%78.5%
Botanical & Floral4414.0%11.5%73.0%
Sports & Hobbies4118.8%12.0%83.7%
Business & Work5818.1%13.0%77.7%
Wedding & Celebration1121.6%14.0%75.0%
Health & Wellness4722.5%14.0%88.0%
Kids & Nursery2523.0%14.0%78.9%
Animals & Pets6321.6%15.0%76.3%
Holidays & Seasons4124.1%15.0%75.0%
Fantasy & Mythology4019.5%17.0%72.2%
Culture & Folk Art12223.6%20.0%79.1%
People & Lifestyle4424.7%20.0%86.7%
Food & Drink5025.5%21.1%76.4%
Travel & Places5927.8%22.0%79.6%
Weighted counts every sampled asset; median counts every keyword once. Two rows are thin — Lettering 13 keywords, Wedding 11 — because a subject needs both readings to appear here at all.

Method and limits

  • Sample: 1,198 subjects with at least 500 Adobe Stock illustrations, an AI reading on or after 2026-08-02, and a download-ranked page of at least 20 assets. Below 20, one atypical page swings the share by tens of points.
  • It is a download RANKING, not sales. Nobody here has access to anyone's sales figures, including our own users'. Adobe orders results by download count and we count the flag over that order; what we cannot see is revenue, licence type or who bought anything.
  • The ranking is cumulative, so work predating generated files carries an advantage from age alone. That confound is real and we have not corrected for it. The eight subjects at 84-99% are the reason we publish the figure anyway: they are all recent phrases, with no older catalogue for age to favour.
  • 89.0% of this sample is tier A, meaning subjects with demand we have already measured: the download-ranked read costs a second page load per subject, so it is rationed. This describes the measured head of the corpus, not the corpus.
  • Weighted across every sampled asset rather than every subject, the figure is 20.5% rather than 14.0%, so the subjects with the most AI among their bestsellers also tend to have the fullest sampled pages.
  • Adobe Stock only, illustration only. It is Adobe's own flag, so it inherits whatever Adobe's labelling misses, and Shutterstock, iStock and Etsy publish no AI flag at all — none of this extends to them or to photography.

Both readings are free on every keyword page — browse the corpus, or read how much of the catalogue is AI for the other half of this pair.

Questions people ask

Does AI-generated stock art actually sell?

Some of it clearly does — 8.6% of the subjects we measured have a majority-AI download ranking, and eight of them are above 84%. But for the median subject, generated work is 75.8% of the catalogue and 14.0% of the most-downloaded results. Whatever is driving that gap, "AI has taken the market" is not what the download ranking says.

Is this sales data?

No, and the distinction matters. Adobe lets you order results by download count; we count its AI flag over the first page of that order. We have no access to sales, revenue or licence data from any marketplace, and we do not claim any.

Older work has been accumulating downloads for years. Does that not explain the whole gap?

It explains part of it, and we cannot say how much. What it does not explain is the subjects where generated work is 84-99% of the download ranking. Those are all recent phrases with no older catalogue to accumulate an advantage — so the measurement does discriminate, rather than simply reporting asset age.

Should I pick subjects by this number?

Read it against the catalogue share for the same subject rather than on its own; the distance between the two is the informative part, and a low bestseller share on its own can simply mean a small sample. It is also one marketplace. Both readings are published free on every keyword page.

Why do only some keywords have this reading?

It costs a second page load per subject, so it runs against subjects whose demand we have already measured — 89% of this sample is in that tier. Where a keyword page shows no figure, it means we have not read the ranking for it, not that the share is zero.

Other measurements

These readings, for the subjects you care about

Demand, existing supply and AI share on every subject SincereTrend tracks — plus the 12-month history and the week to have the work indexed by.

14-day free trial · no credit card to start · then $5/month