No stock agency lets you run a true A/B test. There is no way to split search traffic between two versions of the same file, so every keyword experiment on Shutterstock, Adobe Stock or Getty is really a before-and-after comparison across a matched group of images. That is workable, but only if you control for the things that move downloads on their own.
That distinction is why the tool question is usually the wrong first question. Editing metadata at scale is a solved problem with half a dozen mature products. Knowing whether the edit helped is not, because contributors do not get the measurement surface that a real experiment needs. Below is what the platforms actually expose, what each tool in this category is genuinely for, and how to design a test that produces a signal you can trust.
Why a real A/B test is not available
An A/B test needs three things: two variants live at the same time, traffic split between them, and an outcome measured against a shared denominator. Stock agencies give you none of the three. One file has one set of metadata. Search traffic is allocated by an algorithm you cannot address. And the denominator that would make the comparison meaningful, impressions for a given query, is not published to contributors by any major agency.
You could upload the same image twice with different keywords, but that is a duplicate submission and agencies reject or remove duplicates. So the honest ceiling on what a contributor can do is a sequential test: measure a cohort, change one thing, measure again, and be disciplined about everything else that changed in the meantime.
What you can actually measure
Adobe Stock's contributor portal has an Insights tab covering sales, downloads and withholding for a custom date range, with a CSV export. Shutterstock's contributor dashboard reports downloads and earnings. Neither gives you per-keyword impressions or your position in a specific search result, which means you cannot separate a metadata effect from a demand effect using agency data alone.
The metric that gets closest is sell-through rate: downloads divided by portfolio size over a defined window, calculated per cohort rather than per file. It is imperfect, because it still contains seasonality and agency ranking changes, but it normalises for the one thing you control, which is how many files are in the test group.
Attribution at the keyword level is a separate problem and worth understanding before you design a test, because a keyword that appears on your best sellers is not necessarily the keyword that sold them. We covered the difference in which keywords drive your downloads.
A test design that survives contact with reality
Test cohorts, never single files. A typical microstock file produces a handful of downloads a year, which is far too few events for a before-and-after comparison to mean anything. Group thirty to a hundred files that are genuinely comparable in subject and age, split them into a treatment group and a control group you leave untouched, and compare the two groups against each other rather than against their own past.
Keeping a control group is the step most contributors skip, and it is the step that makes the result readable. If both groups rise, the market moved. If only the edited group rises, the edit plausibly did something. Without the control you are comparing summer to autumn and calling it a keyword effect.
Change one variable. Rewriting titles and swapping keywords and adding a category in the same pass produces a number you cannot interpret. Pick one: title structure, the first few keywords, conceptual versus literal terms, or category assignment. Everything else stays identical.
Run it long enough to clear the noise floor. Weekly download counts on a small cohort swing wildly for reasons that have nothing to do with metadata. A window measured in months, covering the same calendar period for both groups, is the minimum that produces a difference worth acting on. Anything shorter mostly measures your own impatience.
Finally, write down the hypothesis before you edit. Not as ceremony, but because a test you can restate in one sentence is a test you can act on. Something like: replacing generic emotion keywords with specific object nouns on this cohort will raise sell-through relative to the control. If you cannot say what result would change your workflow, the experiment is not worth the risk to the files.
What each agency lets you change
Shutterstock allows edits to the title, keywords and categories of approved content through Catalog Manager, reached from the dashboard under Portfolio. Files still pending review cannot be edited at all, only removed from the queue. Rejected files cannot be amended and have to be re-uploaded, even for a one-word title change. Commercial and editorial designations are fixed once approved, so a file in the wrong bucket has to be deleted and resubmitted.
Adobe Stock permits metadata edits on live files through its contributor portal, and its Insights tab is where the outcome data lives. The practical consequence for testing is that your treatment group has to be built from approved files on both platforms, and that any file rejected mid-test drops out of the cohort rather than being fixed in place.
The tools, and what each one is actually for
No product in this category runs an experiment for you. They split cleanly into tools that change metadata at scale and tools that measure outcomes, and a serious test usually needs one of each.
Native agency dashboards
Best for: the authoritative number, and for small tests on one agency. Cost: free. Limitation: no cross-agency view, no impressions, and export quality varies. Adobe gives you a CSV for a custom date range, which is enough to build a cohort comparison in a spreadsheet. If you sell mainly on one platform, this is genuinely all you need.
Stock Performer
Best for: the measurement half of a test. It analyses at the level of files, productions, keywords and themes across roughly fifteen agencies including Getty and iStock, Adobe Stock, Shutterstock, Pond5, Alamy, Dreamstime and Canva, and it exposes sell-through rate and revenue per image directly, which are the metrics a cohort comparison needs. User-defined collections map neatly onto test and control groups. Its pricing page lists a basic Sparrow tier at nine euros a month and the full Eagle analytics tier at twenty-nine, both cheaper billed annually, with a free trial that includes everything. Limitation: it reads your sales, it does not touch your metadata, and it cannot see impressions any more than you can.
Microstockr
Best for: contributors who want measurement and editing in one desktop app. It is a Mac and Windows application combining sales analytics, AI keywording and uploading to 20+ agencies, with per-image download history, collections and comparison charts, on three tiers from hobbyist through professional to a small team plan, each with a free trial. The vendor's own site documents the agency list. Limitation: being desktop-bound means the data lives on one machine, and combining the tool that edits with the tool that scores can make you less sceptical of your own results, not more.
StockSubmitter
Best for: pushing metadata changes to a lot of agencies at once. It has been running for over a decade, supports around forty agencies, stores credentials locally with an optional master password, and tracks both sales and review outcomes. Pricing is per submission rather than per file: a free tier gives 33 submissions a month to each agency with unlimited uploads, and paid tiers on its pricing page run from four euros a month at the entry level up to roughly sixty-two for unlimited, with the lowest monthly rates on multi-year terms. Limitation: Windows-first, with a separate online version for other platforms, and the interface rewards patience.
Xpiks
Best for: making the edit itself precise. It is a desktop XMP, IPTC and EXIF editor with batch editing, partial changes, keyword reordering, batch keyword deletion, find and replace, CSV import and export, and a pre-upload requirements check, uploading over FTP, SFTP or FTPS or through its own cloud. For applying exactly one controlled change to exactly one cohort, that toolset is close to ideal, and the vendor documents which features sit behind the Pro tier. Limitation: no analytics of any kind, so it can only ever be half of a test.
Rastock AI
Best for: keeping a variation consistent and reversible across agencies. Metadata is generated against each agency's own rules rather than one generic output, with banned and mandatory keyword lists you define, control over ordering, delivery to 10+ agencies over FTP and per-file status tracking in one place, plus a rollback if a batch goes out wrong. In test terms, that covers the treatment application and the undo, which matters because an experiment you cannot reverse is a change, not a test. The feature overview walks through the workflow. Limitation, stated plainly: it is not an analytics product and will not tell you which variant won. Pair it with agency data or a dedicated analytics tool for that half.
When to use which
If you sell on one agency and want to know whether a title format works, use the agency's own dashboard and a spreadsheet. Adding a paid analytics tool to a single-platform test buys you convenience, not accuracy, because the underlying numbers are the same ones you can already export.
If you sell across several agencies, the calculus changes. Stitching Shutterstock, Adobe, Getty and Pond5 exports together by hand every month is where most self-run tests die, and that is the specific problem Stock Performer and Microstockr solve. Choose between them on whether you want a web tool focused purely on analysis or a desktop app that also handles keywording and upload.
If the hard part is applying the change cleanly to hundreds of files with different rules per destination, the answer is a metadata and delivery tool rather than an analytics one. Xpiks and StockSubmitter both do this well from a desktop, and both assume you are supervising the run. A managed pipeline makes sense when the same test has to reach many agencies at once and you want per-agency rules enforced before delivery rather than discovered through rejections.
The risks nobody puts in the marketing copy
Editing live metadata is not risk-free. Shutterstock states that editing published keywords to introduce repeated words or phrases is not allowed, and that it regularly audits the live collection for spam. It also warns that adding back keywords a reviewer or administrator removed may result in a warning, account suspension or closure. The full set of rules and how each one is enforced is in our breakdown of Shutterstock keywording rules.
This rules out the most tempting experiment, which is padding a cohort toward the keyword ceiling to see whether more terms means more downloads. That is precisely the pattern the spam rules target, and the downside is asymmetric: a marginal gain against a possible account action. Design tests that swap terms rather than accumulate them.
The second risk is quieter. Files that already sell carry download history, and history is understood to feed ranking on the major platforms. Running your experiment on proven earners means the momentum of those files masks whatever the edit did, and you may degrade something that was working in exchange for an unreadable result. Test on files that are not yet performing, where there is less to lose and the signal is cleaner.
A worked sequence
Take a hundred files from the same shoot that have been live at least six months and are selling below your portfolio average. Split them fifty-fifty by alternating rather than by subject, so both halves contain the same mix. Record downloads and sell-through for both halves over the previous quarter. This is also the starting point for a wider portfolio re-optimisation, so the work is not wasted even if the test is inconclusive.
Rewrite titles on one half only, using a single documented rule such as leading with the subject and naming the setting. Leave keywords and categories untouched. Push the change through whatever tool you already use, and keep a copy of the original metadata so the change is reversible; the mechanics of doing that safely at scale are in our guide to bulk editing keywords on a large archive.
Wait a full quarter, then compare the two halves against each other for the same period, not against their own previous quarter. If the treated half moved and the control half did not, you have something. If both moved together, you learned about demand rather than about titles, which is still worth knowing and cost you nothing but patience.
If your portfolio is small
Under a few hundred files, testing is mostly theatre. You do not have the event volume to detect a difference, and the time spent building cohorts is better spent applying rules the agencies already publish: unique metadata per file rather than a shared set across a series, conceptual terms alongside literal ones, titles that read as sentences, and no repetition. Those are documented and enforced, which makes them a surer bet than anything a small-sample test will tell you. Rastock AI applies them per agency automatically, with a 14-day free trial and no card required, no revenue share and no ownership claim on your files.
The uncomfortable conclusion for anyone hoping for a testing feature is that the agencies control both the experiment surface and the measurement surface, and they have not opened either. Until they do, the contributors getting the most out of their metadata are not the ones running the cleverest experiments. They are the ones following the published rules exactly, across every agency, on every file, and letting volume do the rest.
Frequently asked questions
Can I A/B test keywords on Shutterstock or Adobe Stock?
Not in the true sense. Neither agency lets you run two metadata variants of the same file simultaneously or split search traffic between them, and neither publishes per-keyword impression data to contributors. What you can run is a sequential test: take a matched cohort of files, edit one group, leave a comparable control group untouched, and compare sell-through rate between the two groups over the same period. Uploading the same image twice with different keywords is a duplicate submission and will be removed.
Can I edit keywords on files that are already published?
On Shutterstock, yes, through Catalog Manager under Portfolio, where you can change the title, keywords and categories of approved content. Pending content cannot be edited, only deleted from the queue. Rejected content must be re-uploaded rather than amended. Commercial and editorial designations cannot be swapped after approval. Adobe Stock also allows metadata edits on live files through the contributor portal.
Does re-keywording an old file reset its search ranking?
No agency documents that it does, and no agency documents that it does not. Ranking on the major platforms is widely understood to combine metadata relevance with download history, so an older file with sales carries momentum that a metadata edit does not erase. This is also why a before-and-after test on established files is harder to read than one on a fresh cohort: the history is doing work the edit cannot undo.
What tools let me compare keyword performance across agencies?
Stock Performer is the most focused option, with analysis at the level of files, productions, keywords and themes across roughly fifteen agencies, priced from around nine euros a month for the basic tier to twenty-nine for full analytics, cheaper annually, with a free trial. Microstockr covers analytics plus keywording and uploading in one desktop app. StockSubmitter tracks sales and approvals alongside its submission workflow. None of them run experiments for you; they supply the numbers you compare.
Is there a risk in editing metadata on live files?
Yes, and it is specific rather than vague. Shutterstock states that editing published metadata to introduce repeated words or phrases is not allowed, and that re-adding keywords a reviewer or administrator previously removed may result in a warning, account suspension or closure. It also audits the live collection for spam. A keyword test that pushes toward the fifty-keyword ceiling to see what happens is exactly the pattern those rules target.
How many files do I need for a keyword test to mean anything?
Enough that a swing of two or three downloads cannot flip the result. Since a typical microstock file earns only a handful of downloads a year, a single file never qualifies. Cohorts of several dozen comparable files per group, measured over months rather than weeks, are the realistic floor. Contributors with portfolios under a few hundred files are usually better served by applying known agency rules correctly than by trying to run experiments on thin data.