Skip to main content

AI content

Test prompts on real products Nothing saves until you apply.

Load a real product, generate a field, and read the generated text next to the value that is stored today.

Start freeBook a demo
  • 8 prompt modes
  • Current and generated side by side
  • Your recent results stay switchable

Key features

  • Eight modes to work in

    Description, short description, title, SEO meta, product attributes, FAQ, image alt text and web research, each with its own settings.

  • Tone, industry and length

    Set the voice, say what you sell, and cap the length. The character count and the field's limit sit under the output.

  • Nothing is applied by accident

    Generating copy writes nothing. Apply is a separate button and it only touches the product you loaded. Web research is the exception: it starts a real research job that follows your research settings.

Why test first

Why the first prompt rarely works

Wording that produced a good description for one product often collapses on the next one, and a catalog is where you find that out the expensive way.

  • The first draft is always too generic

    A prompt written in the abstract produces copy that would fit any product in the category. Reading it next to a real record is what makes that obvious.

    Result: you fix the prompt instead of four hundred descriptions.

  • Length rules break quietly

    A meta description that runs past the limit gets cut by the search engine rather than by you, and nothing in the catalog flags it.

    Result: the character count and the limit are visible while you tune.

  • Nobody wants to test on live data

    Trying a prompt on real products usually means writing to them and undoing it afterwards.

    Result: generating and applying are two separate actions.

How it works

Load a product, choose a field, compare

Three steps, and the third is where you decide whether the prompt was any good.

  1. 01

    Load a product

    Step 1

    Pick a real product from your catalog. Its current values fill the left panel, so every comparison is against what is actually stored.

    The lab is working on your data instead of an example.

  2. 02

    Choose a field and configure it

    Step 2

    Pick the field to work on, set tone, industry and target length, and start from a saved prompt or one of your presets.

    A configuration you can save as a preset for next time.

  3. 03

    Generate and compare

    Step 3

    The generated text appears next to the current value with its character count. Quick refinements nudge the result without rewriting the prompt.

    You apply the version you picked, or change the prompt and run it again.

What you can work on

Eight modes, eight sets of settings

Each mode knows what its field is for, so the controls and the limits change with it.

  • Description

    The long product description, with tone, industry and a target length you set before generating.

  • Short description

    The teaser that has to work inside a listing, held to the length the channel expects.

  • Titles

    Product titles built from the attributes the record already carries, in the order your category reads best.

  • SEO meta

    Meta title and description with the character count and the limit visible while you tune.

  • Product attributes

    Structured attribute values instead of prose, derived from what the product already holds.

  • FAQ

    Buyer questions and answers for the product, up to 20, with the question types, tone and point of view you choose.

  • Image alt text

    Alt text per image, written from the product it belongs to rather than from the file name.

  • Web research

    Start a real research job for the loaded product from inside the lab. It follows your web research settings, including automatic applying.

Compare

Compare before you apply

Generated text lands beside the value it would replace, never on top of it, so the decision is a comparison instead of an undo.

  • Current and generated, side by side

    Both panels carry a character count, and the generated one warns when it runs past the field's limit.

  • Refine without rewriting

    Quick refinements adjust the result, so a small correction does not mean editing the prompt again.

  • Your recent results stay

    The last few results per field stay switchable, so you can go back to the one that read best.

The two panels never merge: your stored title and description stay on the left, and anything the AI writes appears on the right, where you read it before it is saved.

Why test on one product

A prompt is cheapest to fix before the bulk run

Every weakness in a prompt is multiplied by the number of products you run it on. Testing on real products first is how you keep that number at one.

Generating copy across a catalog uses 1 AI credit per product. A prompt that turns out to be too generic on a run of 5,000 products costs 5,000 credits and an afternoon of cleanup, and the second run costs the same again. Text tests in Prompt Lab use no credits at all, so the sensible order is obvious: get the prompt right on a handful of real products, then run it on all of them.

Real products matter because prompts fail on the products you did not picture when you wrote them. A prompt written with a rich, well-described bestseller in mind will invent filler for a product that arrived with a name and an EAN. A prompt tuned on a jacket will produce odd copy for a belt in the same category. Prompt Lab loads the actual record from your catalog, with its real attributes and its real gaps, and shows you what the prompt makes of it.

Nothing you generate in the lab touches the product until you press Apply, and Apply only writes to the one product you loaded. You can test as often as you like on live data without a staging copy of your catalog.

Eight modes

What you can test in Prompt Lab

Each mode knows its field, so the controls, the limits and what Apply writes change with it.

Prompt Lab modes and their settings
ModeWhat it generatesSettings you controlWhat Apply writes
DescriptionThe long product descriptionTone, purpose, industry, point of view, target length, specification tableThe description of the loaded product
Short descriptionThe teaser shown in listings and summary blocksTone, industry, point of view, target lengthThe short description
TitlesOne, three or five title optionsNumber of options (1, 3 or 5), maximum length, include the brand, tone, industryThe title you pick
SEO metaSEO title and meta descriptionPrimary keyword, industryThe meta title and meta description
Product attributesStructured attribute values derived from the product's dataIndustry, how many attributes, which kinds to focus onThe attribute values you accept
FAQBuyer questions and answers for the productFAQ types, number of questions (up to 20), tone, point of view, use the latest web researchThe product's FAQs
Image alt textAlt text per product imageWhich image, maximum length, include the brand or a keyword, extra instructionsThe alt text of each image
Web researchSpecs, copy, images and prices from the webIdentifier type, research focusRuns a real research job; see the note below
Web research in the lab starts a background job straight away, costs 5 credits per product, and follows your web research settings, including automatic applying.

A test session

One test, start to finish

The loop most people settle into after their first few sessions.

  1. 01

    Load a real product

    Open Prompt Lab under AI and Automation, search for a product by SKU and load it. Its current values fill the left panel, so every comparison is against what is actually stored.

  2. 02

    Pick the mode

    Choose the field you are working on. The settings panel changes to the controls that field needs.

  3. 03

    Start from a saved prompt or a default

    Load a template from your Prompt Library, start from the mode's default, or write your own instructions.

  4. 04

    Set the controls

    Tone, purpose, industry, point of view and length for copy; a primary keyword for SEO meta; the number of variants for titles.

  5. 05

    Generate and compare

    The output appears next to the current value, each with a character count. Read both before you change anything.

  6. 06

    Refine or rewrite

    Use a quick refinement for a small correction, or change the prompt itself when the problem repeats across products.

  7. 07

    Repeat on other products

    Load a second and third product from different corners of the catalog and generate again with the same prompt.

  8. 08

    Save, then apply or run

    Save the prompt to the library, and apply the result to this product if it is good. The bulk run happens from the library or the product grid.

Choosing test products

Six products that break a prompt early

Testing on your best product proves very little. Pick a small set that covers the ways your catalog differs.

  • The thin one

    A product with a name, an EAN and almost nothing else. If the prompt invents specs here, it will invent them across every supplier import.

  • The rich one

    A product with full attributes and a good description. Check that the new copy uses the facts and is actually better than what is there.

  • The odd one out

    A product that sits in the category but does not look like the rest, such as an accessory among main items.

  • The bestseller

    The product where a worse description costs the most. Compare the output with the copy that already sells.

  • A variant parent

    A configurable product with sizes or colours, to see whether the copy talks about the range or about one variant.

  • Another category

    If the prompt will run across categories, test one from each. Vocabulary that suits shoes can sound wrong on bags.

Reading the result

What to check in the comparison

Most problems are visible in a careful read of one output next to the current value. Look for these before you look at style.

  • Every fact traces back to the product. Each number, material and claim should appear in the product's attributes or description. If it does not, the prompt is inviting the AI to fill gaps.
  • Nothing important was dropped. The current value may hold a certification, a warning or a compatibility note the new copy left out.
  • The length fits the field. Read the character count against the limit of the field and of the channel the copy is for.
  • The opening is about this product. A first sentence that would fit any product in the category means the prompt asks for tone, not substance.
  • No repetition across products. Generate for three products and read the openings side by side. The same phrase three times will be the same phrase 3,000 times.
  • The keyword sounds natural. For SEO meta, the primary keyword should read like part of the sentence, not like a tag.
  • The voice matches your other copy. If it does not, fix the Knowledge Library rather than the single prompt; the AI reads it on every run.

Iterating

Refine without starting over

The lab keeps the small corrections cheap, so you spend your edits on the prompt itself.

  • Quick refinements nudge the last result with one click, such as shorter, more technical, punchier, benefits-led, lead with the brand, or a stricter length for titles.
  • Recent results stay switchable per mode for about two hours, so you can go back to the version that read best. They live in your browser, not in the product.
  • Presets save a mode's settings, such as tone, industry and length, so they come back the next time. Presets are stored in your browser, so a colleague does not see them.
  • Two test variables: in a lab test, {{title}} and {{description}} are replaced with the loaded product's values. You rarely need them, because the product's data is sent along with every prompt anyway.

Titles and SEO meta

Testing titles and meta descriptions

Short fields are where a prompt is easiest to test and easiest to get wrong, because every character counts.

In Titles mode you can ask for one, three or five options at once and apply the one you like. Generate options for a handful of products, then look at what the winners have in common: usually brand and product type first, then the one attribute buyers filter on. Write that order into the prompt.

In SEO meta mode, set the primary keyword per test. Google does not set a fixed character limit for meta descriptions; it shortens the snippet to fit the device. A length you choose and keep consistent is more useful than chasing a magic number, and the character count under the output helps you hold it.

Marketplace titles are different, because the limits are hard. Amazon, for example, caps most product titles at 75 characters from 27 July 2026. If a prompt writes titles for a channel with a limit, put the limit in the prompt and check the count on your longest product names.

From lab to catalog

Turn a tested prompt into a bulk run

A prompt that held up on six products is ready for six thousand.

  1. 01

    Save it to the library

    Save as a new prompt, or update the library prompt you loaded. Its settings are saved with it.

  2. 02

    Make it the default for its field

    The default is preselected in the enrichment dialog and used when WISEPIM fills empty fields on its own.

  3. 03

    Run it on a first batch

    Select a filtered batch of a few hundred products and run the prompt from the grid. Each product uses 1 credit.

  4. 04

    Read a sample, then scale

    Open ten products from the batch. If they hold up, run the rest, or put the prompt in an automation so new products get it on import.

Good to know

What Prompt Lab does not do

So you know where its answers end:

  • No model picker. WISEPIM chooses and maintains the model behind each job. You control the instructions and the settings, not the model or its temperature.
  • One prompt at a time. There is no split view that runs two prompts side by side. Generate variants of a prompt in the library, then test each one in the lab.
  • Session history, not an audit log. Recent results are kept in your browser for a short time. Save anything worth keeping to the library.
  • Apply is per product. The lab never writes to more than the one product you loaded. Bulk runs happen from the library, the product grid or an automation.

Sources

Search and marketplace facts on this page

Snippet length and the Amazon title limit come from the sources below, checked in October 2026.

FAQs

Prompt Lab questions, answered

What you can generate, what testing does to your data, and how a prompt you tuned here reaches the rest of the catalog.

Still have questions?

Can't find the answer you're looking for? Please get in touch with our team.

Contact Support

Eight things: the long description, the short description, product titles, SEO meta, product attributes, FAQs, image alt text, and web research on the loaded product. Each one is its own mode with its own settings, because a title and a meta description do not need the same controls. It does not generate images, only the text about them.

No, with one exception. Generating copy produces output in the lab and writes nothing to the product, and Apply is a separate button that only touches the single product you loaded. Web research mode starts a real research job that follows your web research settings, so with automatic applying on (the default) its results are written to the product. Running a prompt across a selection is a different action, in the enrichment modal on the product grid.

Yes. Save it to your Prompt Library and it is shared with your project, and available to the bulk enrichment run. The per-mode settings can also be kept as a preset, so the tone, industry and length you settled on come back the next time you open that mode.

Two ways. Tone and industry are controls you set per run, which is enough for a quick test. Anything more specific comes from the Knowledge Library, which the lab reads along with your prompt: the phrases you use, the ones you avoid, who you are writing for.

It keeps your recent results per field so you can switch back to an earlier one while you are working. That is a working history for the session, not a permanent audit log of everything the AI has ever written. Anything you want to keep goes into the Prompt Library or gets applied to the product.

Name the audience in the prompt rather than describing the product, set the industry so the vocabulary matches your sector, and lower the target length: short copy forces the model to pick something specific. If it still reads like anyone's catalog, the Knowledge Library is usually the missing piece.

Text tests do not. Generating a description, short description, title, SEO meta, attributes, FAQs or alt text in the lab uses no credits. Web research in the lab is a real research job and costs 5 credits per product.

The same model WISEPIM uses for enrichment, so what you see in the lab is what a bulk run produces. There is no model picker; you steer the output with the prompt and the settings.

Not in one view. Generate variants of a prompt in the Prompt Library, then load each one in the lab on the same product and switch between the recent results to compare them.

Five or six that differ: a thin product, a rich one, a bestseller, a variant parent, an odd one out and one from another category. A prompt that works on all six rarely surprises you in a bulk run.

No. Presets and recent results are stored in your browser. To share a tuned prompt, save it to the Prompt Library, where it is shared with your project by default, or set it as the default for its field.

Usually because the products differ, not the prompt. A test on a rich product does not predict the output for thin ones. Test on a thin product too, and fill missing facts with web research before the run.

Prompt Lab is part of the WISEPIM app, and text tests use no credits, so you can tune prompts on the free plan's 100 products. Running a prompt in bulk uses AI credits, of which the free plan includes 50 a month.

Next step

Test a prompt on your own product

Load one product, generate the field, compare it with what is there. Nothing is saved until you apply it.

Start free
Book a demo
4.7

Prompt Lab documentation

The eight modes, the tone and industry controls, quick refinements, and saving a tuned prompt back to your library.

(opens in a new tab)

Keep exploring

Hand-picked next steps to go deeper.