AI Product Title Generator: How to Test One Before You Trust It

An AI product title generator can be trusted for first drafts, not for final titles. Test it by running 20 of your own listings through it, scoring each new title against the old one for accuracy, keyword fit and length, and changing only the titles that pass a fact check by a person.

The reason to test rather than trust is that a title has two jobs at once: it must be accurate, so a customer gets what the words promise, and it must be findable, so the right customers see it. A generator is quite good at the second and unreliable at the first, because it fills gaps with plausible details. A title that says "waterproof" for a water-resistant jacket brings returns, complaints and in some marketplaces a policy problem. The test below takes an afternoon and costs nothing beyond the chat assistant you already use.

Follow me on Instagram@sagnikteaches

What can go wrong in an AI-written title

Before building the test, know the failure types, because the scoring sheet is built to catch them.

Connect on LinkedInSagnik Bhattacharya

• Invented attributes. The tool adds a material, size, compatibility or certification that is not in your data. This is the most damaging error.
• Dropped attributes. It leaves out the size, count or model number that the buyer actually filters on.
• Keyword stuffing. It repeats the product word three ways to please search, and the title reads like spam.
• Length problems. It overshoots the limit for the channel, so the end, often the variant detail, is cut off.
• Wrong order. It puts decoration before the product word, so the first few words, which are often all a shopper sees on a phone, say nothing about what it is.
• Style drift. Capital letters on every word, emoji, claims like "best" or "premium" that you cannot support.

Step one: pick a sample of 20 listings

A sample that is all easy listings will flatter the tool. Build your 20 from four groups of five:

Subscribe on YouTube@sagnikteaches

1. Best sellers. They have the most to lose from a bad change.
2. Slow sellers. They have the most to gain, and the weakest current titles.
3. Simple products. One variant, a few facts.
4. Awkward products. Many sizes, compatibility lists, bundles, or technical specifications.

Put them in a sheet with these columns: product ID, current title, the facts you hold (material, size, colour, count, model, compatibility, anything on the label), the primary keyword you want it to rank for, and the length limit for the channel. The facts column is the heart of the test. It is the only source the generator is allowed to use, and it is the reference you will check the output against.

An illustrative row for a gift shop selling ceramics: current title "Blue mug"; facts "ceramic, 350 ml, dishwasher safe, microwave safe, hand-glazed, blue speckle, comes in a gift box"; keyword "speckled ceramic coffee mug"; limit 150 characters. A pet shop might use "Salmon dog treats" with facts "single ingredient, baked, 150 g, no added salt or sugar". A sports club shop might use a team scarf with colours, material and size in cm. Use the same structure for every product, whatever the trade.

Step two: use a fixed prompt, not a fresh one each time

A fair test uses one prompt for every listing, otherwise you are testing your prompt writing rather than the tool. Here is a prompt you can copy.

Write 3 alternative product titles for an online shop.

Use ONLY the facts below. Do not add any material, size, feature,
certification or benefit that is not listed. If a useful detail is
missing, write [MISSING: what you would need] instead of guessing.

Rules:
- Maximum 120 characters each.
- Start with the plain product name a buyer would type into search.
- Include the primary keyword once, naturally. No repeats.
- Put size, count or model before decoration.
- Sentence case, no emoji, no words like best, premium or perfect.

PRIMARY KEYWORD: [keyword]
CURRENT TITLE: [title]
FACTS: [facts]

Two features of this prompt matter. The "MISSING" instruction gives the model a safe way to say it lacks information, which cuts invented details. And the rule to put the plain product name first protects the opening words. Run it in ChatGPT or Claude, whichever you use for other jobs. If you want to compare them, run the same 20 listings through both with identical prompts and score both; the differences between assistants are usually smaller than the differences between good and bad prompts.

If your plan lets you upload a spreadsheet, you can attach the sheet and ask for the output as a new column. Otherwise paste five listings at a time. Keep the three suggestions per listing, because the best of three is often better than the first.

Step three: score each title on four checks

Add columns to the sheet for each suggested title, then score on four checks. A person does the scoring, ideally the one who knows the products best.

Check one: accuracy, pass or fail

For each claim in the new title, find it in the facts column. If a word is not supported, the title fails, whatever else is good about it. This is the strictest check and it is not negotiable. A subtle case: "dishwasher safe" is in your facts, but the title says "dishwasher and oven safe". That is a fail, and the kind of one that costs a return and a bad review.

Check two: keyword fit, 0 to 2

Score 2 if the primary keyword appears once and reads naturally in the first half. Score 1 if it appears, but late or awkwardly. Score 0 if it is missing or repeated. Do not chase exact-match phrases at the expense of readability; a title like "Mug ceramic coffee mug speckled mug" scores 0.

Check three: length and cut-off, pass or fail

Count characters against the limit for the channel. Google's product-data guidance for Shopping, for example, allows up to 150 characters but says shoppers usually notice only the first 70 or fewer, depending on screen size. So look at the first 70 characters on their own: the product name and the single most important attribute should be inside them. Channel rules vary and change, so check the current title guidance in each channel's help pages; for Google Shopping and Amazon in particular, see Google Shopping title rewrites and Amazon's 75-character title rule.

Check four: fit to the shopper, 0 to 2

Read the title as a customer would. Would they know what it is, how big, how many? Score 2 if yes, 1 if one thing is unclear, 0 if the title would make them click and then leave. This is subjective, which is why one person scores all 20, so the standard is at least consistent.

A filled-in example: five listings from a garden centre

An illustrative garden centre tests five products. The table shows the old title, the best of the three AI titles, and the scores. Accuracy is pass (P) or fail (F); keyword and shopper fit are 0 to 2.

Read the failures. "Stainless steel" and "ergonomic" were not in the facts column for the secateurs; the blade material was listed as "carbon steel". "Premium" and "squirrel-proof" were invented for the bird feeder. And "organic" is a claim that certification schemes and labelling rules restrict for many products; the facts only said "peat-free", so the title is both inaccurate and a claim you could be asked to prove. Two of five passed, a 40 per cent pass rate in this illustration, and exactly the kind of result you want to find out before changing 500 titles.

Now the useful step: take the failures and fix the prompt, not just the titles. Add "Do not use the words organic, natural, premium, best or proof unless they appear in the facts", and rerun the failed three. In the illustrative rerun, the secateurs title comes back as "Bypass secateurs with carbon steel blade, lightweight" and passes. The point of the test is to learn which of your rules the tool needs spelling out.

Step four: the keyword check beyond your own guess

The generator does not know what customers search for, so do not let it choose the primary keyword for you. Get your keyword from sources you can look at: your store's own search box data, the search terms report in your ads account if you run ads, the autocomplete suggestions on a search engine or marketplace, and the words customers use in emails and reviews. Put the result in the sheet before running the prompt.

A quick sanity check for each keyword is to type it into the marketplace search and see whether the top results are the same kind of product as yours. In the illustrative gift-shop case, "speckled ceramic coffee mug" returns mugs; "blue mug" returns mugs, travel flasks and a lot of unrelated items. The more specific keyword is the better title anchor.

If you want the tool's help with ideas, ask for variants after you have supplied the real keyword: "Give 10 other phrasings a buyer might type for this product". Treat the output as a list to check against autocomplete, not as evidence of demand.

Step five: put the winners live in a small batch, with a control

Pass in the sheet means "good enough to try", not "proven". Choose 10 listings with passing titles and change them, leaving a comparable 10 untouched as a control. Record the date, and keep the old titles in the sheet so you can revert in minutes.

After three to four weeks, compare views, clicks and orders for the changed group against the control, using whatever analytics your channel gives you. Small shops will see noisy numbers; unless the difference is large, do not draw conclusions from it. What you can say is whether the change did harm: if the changed group lost orders or got more questions asking for details the old title provided, revert them.

Also watch returns and messages. If "doesn't fit" or "not what I expected" notes increase on changed listings, the new title promises something the product does not deliver. Spotting patterns in complaints, returns and defects shows how to catch this early.

A platform's built-in tool versus a chat assistant

Many shops use a title or description assistant built into their store platform or marketplace rather than a chat assistant. The test above works for both, since it checks outputs against your facts. The built-in tools often see your product data directly, which can reduce typing, but they also tend to hide the prompt, so you cannot fix a recurring problem by changing instructions. If a built-in tool fails the accuracy check repeatedly, a chat assistant with your own prompt gives you more control. Shopify Magic compared with ChatGPT for descriptions shows the same trade-off on longer copy, and Etsy's AI title suggestions covers one platform's built-in option.

Two more tests worth running on the same sheet

The swap test for variants

Products with colours or sizes cause a particular problem: the tool writes a title for the parent product, then a variant title that quietly contradicts it. In an illustrative case, a sports club shop sells a training top in three colours and four sizes. Ask for the parent title and the title for the "navy, size L" variant in the same prompt, and check that the variant title adds only what the facts column lists for that variant. If the parent says "unisex fit" and the variant says "men's", that is a fail. Run this on two or three of your multi-variant listings, because variants are where a mistake multiplies across the most pages.

The blind read

Print or hide the sheet so the old and new titles are unlabelled, and ask a colleague or a friendly customer which of each pair they would click on and which describes the product more clearly. Five pairs are enough. It guards against the one thing the four checks cannot see: a new title that is accurate and well optimised but plainer or stranger than your regular customers expect. If your colleagues prefer the old title in most pairs, look at what the old titles do that the new ones do not, and put that into the prompt, for example "keep the product nickname customers use".

A record of what the test taught you

After the rerun, write four lines at the bottom of the sheet. Which kinds of listing passed on the first run? Which rule did the tool need spelling out? What is the pass rate now? What will you not let it write? An illustrative set of notes for the garden centre reads: "Simple plants and pots pass first time. Tools and anything with a material word needed the 'only use listed facts' rule restated. Pass rate 4 of 5 after the fix. The tool does not write anything with the word organic, ever." Notes like these are what let a second person run the same process next month without repeating the discovery.

Setting a pass mark and recording it

Decide before you score what counts as good enough. A workable bar for this kind of test:

• Every accepted title passes the accuracy check, with no exceptions.
• At least 70 per cent of the generator's best titles pass accuracy on the first run with the fixed prompt. Below that, the prompt or the facts column needs work before you scale.
• Average keyword and shopper scores of 1.5 or better.
• The product name and main attribute sit inside the first 70 characters.

Write the pass mark, the prompt and the date on the first tab of the sheet. When a new model version or a new tool arrives, rerun the same 20 listings and compare. That makes the sheet reusable: it turns "does the new version seem better?" into a measurable question.

Where the tool does well and where it needs a person

From the types of listings tested, a pattern usually appears. Simple products with complete facts pass easily; the tool is quick and the titles are clear. Products with long technical specifications or compatibility lists fail more often, because the tool shortens or paraphrases them and changes the meaning. Products where the facts column is thin produce invented detail unless the MISSING rule is followed. Bundles and multipacks get counts wrong. A reasonable policy is to let the tool draft titles for the simple group, review everything with a person, and write the technical group's titles by hand using the tool only for phrasing options.

A realistic mistake to look for: a title for a "pack of 6" that becomes "6 pairs" because the tool inferred the product was a sock. It was a pack of six single items. The error is easy to miss when scanning quickly, which is why the accuracy check goes claim by claim against the facts column, with the numbers read twice.

A short routine once the test passes

When you trust the process, use a light weekly routine for new products. Add the facts to the sheet first, because a listing with no facts column cannot be checked. Run the fixed prompt, pick the best of three, run the four checks, and publish. Keep the sheet, because it doubles as a record of what each title was based on, which helps when a customer queries a claim. Once a quarter, rerun a random five of the old test listings to see whether results have drifted with the tool's updates. If you want the broader picture of writing product copy with AI beyond titles, writing Shopify product descriptions with AI covers the longer text that sits under the title.

If you would like help designing a test around your own catalogue, that is the sort of job my AI implementation consultation handles; you can reach it through the contact page. Either way, the test is the same: your facts as the only source, a fixed prompt, four checks done by a person, a small live batch with a control, and a written pass mark.

Questions about testing AI product titles

How many listings should I test an AI title generator on?

Twenty is enough to see patterns without turning the test into a project. Pick them across your range: best sellers, slow sellers, simple items and complicated ones with sizes or variants. If every listing in the sample is the same kind of product, the test tells you little about the rest.

Should I change all my titles at once if the test goes well?

No. Change a batch of about 10 to 20 listings, wait a few weeks, and compare against the untouched ones. Changing everything at once means a drop or a rise cannot be traced to the titles. Keep a copy of every old title so you can revert quickly.

Can the AI see my product page or only what I paste?

Usually only what you paste or upload, unless the tool is connected to your store. Give it the facts (material, size, model number, compatibility) and tell it to use nothing else. Anything it adds beyond that is a guess, and guesses are the main accuracy risk.

Is a longer title with more keywords always better?

No. Marketplaces and shopping feeds have length limits and rules against promotional text, and long titles get cut off in results. A clear title with the main product words early, followed by the key attributes, usually does better than a long string of repeated terms.

Further reads

• How to Write eBay Listings With AI That Sell Faster — Titles and descriptions for a marketplace with its own rules.
• How Small Amazon Sellers Use AI for Listings and Reviews — How small sellers handle listings and reviews with AI.
• Taking a Local Shop Online With AI: Product Listings in a Weekend — Getting a first catalogue online with AI help.
• How to Use AI for Etsy Titles, Tags and Descriptions — Titles and tags for a handmade-goods marketplace.
• How to Give AI Your Business Context So Answers Fit — Reusable context so titles keep your wording.
• Viator and GetYourGuide Listings: Rewrite Titles With AI — A title formula that fits both platforms, prompts that stay inside each one's rules, and a fair way to compare views and bookings before and after.
• AI Product Name Generator for a Shopify Range: Shortlist and Checks — Prompt an AI for 30 product names, cut to five, then check trademark registers, domains and Shopify handles with a worked example.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Google's merchant help article on the title attribute; trade-press reports of Amazon's Seller Central notice on 75-character titles (checked 5 Nov 2026).