Best AI for Product Research: Test Demand Before You Buy Stock

Use a search-enabled AI assistant that cites sources, such as Perplexity, ChatGPT with search on or Gemini, to build a shortlist, then check demand with Google Trends and with your own customers before you order. Judge the AI by whether its links open and say what it claims.

The assistant is good at finding and summarising: what people are asking about, what competitors charge, which complaints recur in reviews. It is poor at numbers it cannot see. Search volumes, market sizes and "growth rates" stated without a link are often invented, and they look exactly like real ones. So the method below uses AI for breadth, Google Trends for relative interest, and a small, cheap test with your own customers for the thing that matters: whether anyone will pay.

Follow me on Instagram@sagnikteaches

The method assumes an existing shop choosing new lines, not a founder picking a first product. You already have customers, a sales history and a supplier or two, which makes the test cheaper and the signals better.

Connect on LinkedInSagnik Bhattacharya

Which AI for which research job

Different tools suit different parts of the job. A rough division, using current list prices:

Subscribe on YouTube@sagnikteaches

You do not need a paid assistant to start. A free search-enabled tier is enough for a first pass on ten candidates, though free plans carry usage limits that change often. Paid plans become worth it when you run this monthly with long prompts and file uploads. If you handle supplier prices or customer lists, read the plan's data settings first; business plans of the main vendors do not train on your content by default.

Worked example: a vet practice's shop picks three new lines from ten

Suppose a veterinary practice runs a small shop with a website beside the clinic. It sells food, bowls and a few toys, takes about 40 online orders a week, and wants three new non-medical lines for the next quarter. The practice manager writes down ten candidates from supplier catalogues, customer requests and staff suggestions: slow-feeder bowls, lick mats, cooling mats, calming diffusers, travel water bottles, grooming gloves, puzzle toys, dental chews, harness sets and a joint supplement range. The next sections show how she narrows ten to three, with illustrative scores.

Step one: ask the AI for evidence, not opinions

A prompt that asks "what are the best new pet products" returns a glossy listicle. A prompt that demands evidence returns something you can use. Run each candidate through the same request, and make it name its sources.

I run a small pet-care shop. Research this candidate product line:
[lick mats for dogs]. Using web search, report on:
1. What shoppers say they want and complain about (quote short
   phrases from reviews or forums, with the link for each)
2. The typical retail price range from at least three sellers
   (give each seller's page link)
3. Any safety rules or labelling rules that apply to this product
4. Anything suggesting demand is falling or seasonal
Rules: every claim needs a link. If you cannot find a source, say
"no source found". Do not give search volumes or market sizes unless
you can link to the page that states them.

Notice the last rule. It stops the assistant from padding the answer with numbers. Even so, open the links. Some links will fail to open or say something different from the summary, and an assistant can quote a review that is not on the page. Log what you find in a sheet with the columns candidate, claim, source link, verified (yes or no). Anything unverified does not count.

A sample of illustrative output for lick mats: shoppers praise silicone mats that are dishwasher safe and complain about suction cups that fail and mats that curl; typical prices sit between $8 and $18 across three sellers; no special rules found beyond general product safety. The "no special rules found" line should send you to your own rules, not end the check, because the assistant cannot know which regulations bind your shop. Ask your trade body or a product-safety adviser where in doubt.

Step two: use Google Trends for direction, with an anchor

Trends does not give counts. It scales a term's share of searches to a 0 to 100 range within the comparison you set up, so numbers from different comparisons cannot be added. Google's help page says low-volume terms may show as 0 and that small values may be noise. That makes it a tool for direction (rising, flat, falling, seasonal) and for comparing a few terms, which suits product research well.

The anchor trick fixes the main weakness. Add a term you already know sells, such as the best-selling bowl type in your shop, to every comparison. Then a candidate scoring a third of your anchor's interest means something. Without an anchor, "75" only tells you the term was at 75 per cent of its own peak.

1. Compare each candidate with your anchor term over five years, looking for steady, rising or collapsing patterns.
2. Check twelve months for seasonality (cooling mats peak in warm months; buying in late summer is a mistake).
3. Look at related queries Trends lists; they often show the words customers use, which differ from supplier catalogue names.
4. Note the date and a screenshot of each result, since results shift.

In the example, puzzle toys and lick mats trace gently rising lines above the anchor's level, cooling mats spike each summer, and travel water bottles are flat and low. Illustrative, but typical of what you find.

If you want rough counts, Google Ads Keyword Planner needs an account with billing information entered before it gives you ideas and monthly search estimates. Treat numbers there as estimates too. For the wider approach to search terms, see the tutorial on keyword research with AI for a small business website.

Step three: score the candidates on one rubric

Put the ten on a scoring sheet. Six criteria, each scored 1 to 5, with weights that add up to 100 per cent. Scores come from your verified evidence, not from the assistant's opinion, though you can have it draft the first column from your notes.

Weighted scores for the ten candidates, illustrative: lick mats 3.80, slow-feeder bowls 3.65, puzzle toys 3.55, joint supplement range 3.30, calming diffusers 3.25, cooling mats 3.20, dental chews 3.10, grooming gloves 3.00, travel water bottles 2.80, harness sets 2.55. The arithmetic is simple enough to check: for lick mats, 4 x 0.25 + 3 x 0.20 + 4 x 0.25 + 3 x 0.10 + 4 x 0.10 + 5 x 0.10 equals 3.80. Have a spreadsheet do these sums, not the assistant, for the reasons explained in the tutorial on why AI is bad at maths.

Add one rule that overrides the totals: any risk score of 1 is a gate, not a penalty. The joint supplement range scored 3.30 on a weighted basis, ahead of two other lines, but its risk score is 1 because health claims, ingredient rules and labelling are tightly regulated for products like that. It is not a quick quarter's test. It goes on a separate list, "check the rules first", and the shop moves to the next candidate. The three lines left standing are lick mats, slow-feeder bowls and puzzle toys.

Step four: price and margin check on real quotes

Margin is the criterion where assistants help least and spreadsheets help most. For each of the top candidates, get a real supplier quote with unit cost, minimum order, lead time and carriage, then compute landed cost per unit, including card or marketplace fees, packaging and an allowance for returns. Illustrative lick mat: unit cost $2.60, carriage and duty $0.55 a unit, packaging $0.30, payment fees $0.45 on a $12.50 sale; landed and fees total $3.90, gross margin per unit $8.60, or 69 per cent. At a minimum order of 200 units the stock outlay is about $690 (200 units at $3.45 for the mat, carriage and packaging, before payment fees), which shows the downside you are risking.

The assistant earns its place here by asking the questions you forgot: what about damaged goods, who pays for returns, do you need a product liability check, what is the shelf space. Paste your costings and ask "what costs have I left out?" It will not know your supplier, but it will list the usual ones. If you also sell through a marketplace, add its referral and fulfilment fees to the same sheet before you compare lines, because a line that clears 60 per cent in your own shop can fall well below that once those fees come off.

Step five: a demand test that costs a fraction of the stock order

A scoring sheet predicts; a test measures. Run one for each of the top three lines before ordering a full quantity. Choose from these, cheapest first.

• Ask the list. Send a short email to existing customers: "We're considering adding X. Would you buy one? Reply or click to join the waitlist." Count clicks and replies, not opinions.
• A waitlist page. A product page with photos from the supplier (with permission), a price and a "tell me when it's in" button. Judge by sign-ups per 100 visitors. Say clearly that the product is not yet available; do not take payment for stock you have not got.
• Pre-order with a ship date. Only where your terms allow it and you are confident of the date; otherwise a waitlist is safer.
• A sample order. Order the smallest quantity the supplier allows, or a sample pack, and sell it in the shop with a counter sign. Count how many are gone in two weeks.
• Staff and counter tests. Ask staff to mention the product to every relevant customer for a week and tally responses.

Decide the pass mark before you start, in writing. An example for a shop with 1,800 on its email list: pass if at least 45 people (2.5 per cent) join the waitlist or pre-order within ten days, and if the sample sells at least 15 units in two weeks. Those thresholds are illustrative; set yours from your list size and your past launches. The point of fixing them first is that people talk themselves into marginal results afterwards.

Use the assistant to write the email, the waitlist page copy and a short follow-up sequence. It is excellent at that. The tutorial on testing a new service idea with AI uses the same logic for services, and the one on spotting trends before buying stock shows how a clothing boutique applies it.

Mining reviews for the reasons people buy and return

Reviews are the best free evidence of demand, and an assistant can read far more of them than you can. Copy 30 to 50 reviews of the leading versions of a candidate into a document (selected by you, from pages you can see, with the star rating) and ask for themes. Do not ask it to find the reviews itself and summarise them unseen; that is where invented quotes creep in.

Below are 40 customer reviews of dog lick mats, each with its star
rating. Group them into themes for what people like and what makes
them return or complain. For each theme give the count of reviews
and two short quotes copied exactly from the text. If a theme has
fewer than three reviews, say so. Finish with three product features
a new supplier should offer to beat these.

Illustrative output: twelve reviews praise easy cleaning, nine complain that the suction cups do not hold, seven say the mat is too small for large dogs, five say the pattern is too shallow and the treat is gone in a minute. Now you have an offer, not just a product: a larger mat with strong suction, which you can check against supplier samples. Check the counts by searching the pasted text yourself for two themes; assistants miscount, though the themes are usually right.

Returns tell you about hidden costs. If a product's one- and two-star reviews are mostly about breakage or sizing, your margin check needs a higher returns allowance. A 4 per cent return rate on a $12.50 item costs about $0.50 a sale before you pay to ship the replacement, so the lick mat margin in step four falls from 69 per cent to somewhere nearer 65 per cent, still healthy but worth knowing.

What to do with the seven that missed the cut

Do not throw the leftover scores away. The sheet is a pipeline. Lines with weak trends but good margin (grooming gloves, at 3.00) are candidates for a bundle with something you already sell: a grooming starter set raises basket size without a new stock risk. Lines with good interest but poor margin (harness sets) may pay at a different supplier, so ask for fresh quotes next quarter. Lines gated by risk, like the supplement range, go to someone qualified to check the rules, and you decide then whether the work is worth it.

Review the sheet every quarter, rescore with fresh Trends comparisons, and retire anything that has sat at the bottom twice. Over a year you build a record of what you looked at, what you decided and what happened, which is more valuable than any single research round.

When the research misleads: three slips to watch

The invented number. A practice manager asks an assistant for the monthly search volume of "lick mat for dogs" and receives "approximately 74,000 searches a month, up 38 per cent year on year". There is no link. The same question in another session returns 49,000. If an answer changes between runs and cites nothing, it is a guess in a confident voice. Replace it with a Trends comparison against your anchor.

The stale source. A cited page turns out to be three years old, from a different market. Check dates, and be wary of "best sellers" pages that exist to sell you a product research tool.

The popular-so-everyone-wants-it fallacy. A product with rising interest and many sellers may leave no room. Interest tells you people are looking; the competition score and the margin check tell you whether you can win. For the rival-shop side, competitor research with AI in one afternoon gives a fast routine.

After the test: reorder rules and the next round

If a line passes, order the first full quantity based on the test's sales rate, not on the supplier's price break. Suppose the sample of 20 lick mats sold in 9 days; that suggests about 2 a day, so 60 units cover a month. Order one month's worth at first even if the price per unit is lower at 200, unless the savings clearly exceed the cost of holding stock for six months. Set a reorder point at lead time plus a week of sales. The tutorial on forecasting stock and reordering picks up from there.

If a line fails, log why: low sign-ups, price rejected, wrong audience. Keep the log with the original scoring sheet. After two or three rounds you will see which criteria predicted success best in your shop, and you can adjust the weights. That calibration is something no assistant can do for you, because the data is yours.

Run the whole routine once a quarter. The first round takes a day or two; with the prompts and the sheet saved, later rounds take an afternoon, and the assistant handles most of the summarising. What stays with you is the part that decides whether money leaves the bank: the sources you opened, the sums you checked, and the pass mark you set before you looked.

Product research questions shop owners ask

Which AI tool is best for product research?

A search-enabled assistant that shows its sources, such as Perplexity, ChatGPT with search on, or Gemini, paired with Google Trends. The assistant finds and summarises; Trends gives relative interest over time. The best tool is the one that gives you links you can open and check.

Can I trust search volumes an AI quotes?

Not unless you can open the source. Assistants often state monthly search numbers that trace back to nothing. Treat any figure without a working link as invented, and use Google Trends or your own shop's search data for relative interest instead.

How many units should a first test order be?

Enough to test the sales page and fulfilment, not enough to hurt if it fails: often the supplier's minimum, or a sample pack. Decide the pass mark before you order, such as a set number of pre-orders or waitlist sign-ups from your existing list.

Does Google Trends show how many people search?

No. It shows interest scaled from 0 to 100 relative to a topic's share of searches, and terms with little volume can show as 0. It tells you direction and comparison, not counts.

Further reads

• Amazon FBA vs FBM Cost: Model Your Own Products With AI — Fee structures if you also sell new lines through a marketplace.
• How to Get Your Products Recommended in ChatGPT Shopping — Get the new line found when shoppers ask ChatGPT.
• Best AI Tools for a Small Ecommerce Store: 7 Worth Paying For — Seven paid AI tools a small store may want once lines grow.
• AI for Ecommerce Analytics: Shopify Reports, GA4 or ChatGPT First? — Read your own shop data before and after a new line.
• Can AI Tell You What to Charge? The Limits of AI Pricing Research — The limits of AI pricing research, including for new lines.
• Online Shop AI Mistakes That Hurt Trust and Conversions — Mistakes that damage trust once the product is live.
• Taking a Local Shop Online With AI: Product Listings in a Weekend — Put the winning line online quickly with good listings.
• Market Analysis AI Tools for Small Business: Costs and Source Checks — ChatGPT, Claude, Gemini, Perplexity and Copilot compared for small-business market research, with list prices and a routine for checking cited sources.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Google Trends help page on how data is scaled; Google Ads help on Keyword Planner access; Perplexity, ChatGPT and Gemini vendor pages for plans. Product names, scores and thresholds in the example are illustrative.