- Every AI product research tool is reading the same free demand sources underneath, and those sources publish limits that change what their numbers mean.
- Google Trends is a sample rather than a census, normalised against total searches in that place and period, then scaled so the highest point in your chart becomes 100.
- Terms searched by very few people are shown as zero. For anyone hunting a niche, the zero is the most misread number on the screen.
- Trends also strips repeat searches from the same person and excludes searches made by Google's own products, including AI Mode and AI Overviews.
- Keyword Planner volumes are rounded and reflect exact matches, so two tools disagreeing about the same product is normal rather than evidence that one is broken.
- The checks that actually stop a costly mistake are not demand checks. They are the trademark register and the safety rules that decide whether you may sell the thing at all.
Ask an assistant what to sell next and you will get an answer in seconds, with numbers in it. The numbers are the problem. They look like measurements of demand, and most of them are ratios, samples or rounded buckets whose documented behaviour makes them useless for exactly the question a small merchant is asking.
This is not an argument against using AI for product research. It is very good at the part humans find slow: reading two hundred reviews of a competing product and telling you what people complain about. It is being asked instead to do the part it cannot do, which is to tell you how many people want a thing.
So the useful skill is reading the underlying sources correctly. There are only a handful of them, they are free, and they all publish how they work.
What is a Google Trends number actually measuring?
Relative popularity, scaled twice, from a sample. Google's FAQ about Trends data describes the pipeline without hedging. The service works from a largely unfiltered sample of real searches, anonymised and grouped by topic. Each data point is divided by the total searches for the geography and time range it sits in, so that places with more total search volume do not automatically rank highest. The resulting figures are then scaled from 0 to 100 based on a topic's proportion of all searches on all topics.
Two consequences follow and neither is intuitive. The first is that 100 means the peak of the chart you are currently looking at, nothing more. Change the date range and the 100 moves. A product that looks like it is exploding across a five year view can look flat across twelve months, and both charts are correct.
The second is that two regions showing identical interest do not have identical search volumes. Google states this directly. Interest is a proportion of local searching, so a small market where a niche product is unusually popular can outscore a huge market where it is merely common.
Why does a real product show zero demand?
Because Trends only displays data for popular terms. The FAQ lists this among the filters: searches made by very few people appear as zero. A term is not absent from the world because it is absent from the chart. It is below a display threshold.
For a merchant hunting an underserved niche, that is a trap laid precisely where you are standing. The entire premise of finding an unserved product is that few people are searching for it yet, and the tool is designed to render exactly that condition as a flat line at zero. An assistant reading the chart will report no demand, which is a factually different statement from what the data says.
Three other filters matter for the same reason. Repeated searches from the same person over a short period are removed, so a handful of obsessive buyers do not register as a trend. Queries containing apostrophes and other special characters are filtered out, which quietly destroys some product names. And searches made by Google's own products and services are excluded, a category the FAQ says includes internal searches made by AI Mode and AI Overviews.
That last one deserves a pause. A growing share of the questions people ask about products are now answered inside an AI surface, and the internal searching that generates those answers does not appear in Trends. The tool is measuring a shrinking slice of how people look for things, and nothing on the chart tells you so.
Why do two tools give different numbers for the same product?
Because they round differently and count differently, and both say so. Google's documentation for Keyword Planner statistics notes that search volume figures are rounded, which is why volumes for several locations often fail to add up the way you would expect. It also notes that historical statistics such as average monthly searches are shown for exact matches, so the same figure appears whether you look at broad, phrase or exact match.
Put a rounded, exact match, bucketed average next to a normalised, sampled, peak scaled index and disagreement is the expected outcome. Neither is wrong. They are answering different questions, and an AI layer that averages them into one confident figure has produced a number that corresponds to nothing.
| Source | What the number is | Documented limit | Safe use |
|---|---|---|---|
| Google Trends | Relative interest, 0 to 100 against the chart peak | Sampled, normalised per region, low volume shown as zero | Direction and seasonality, never size |
| Keyword Planner | Average monthly searches for exact matches | Rounded, so multi location figures do not sum | Rough order of magnitude between two ideas |
| Your own site search log | Exact strings your visitors typed | Only covers people already on your site | The strongest signal you own, and free |
| Marketplace review dates | Evidence a competing product sold | Reviews lag purchases and can be manipulated | Confirming a category is alive, not sizing it |
The third row is the one small merchants systematically ignore. Your own site search log records the exact words real people typed while standing in your shop, and every one of them is a person who wanted something and may not have found it. It is a smaller dataset than Trends and an infinitely more honest one, because it has not been sampled, normalised or scaled against anything. If your search log shows fourteen queries for a size you do not stock, that is fourteen more units of evidence than a Trends line.
What is AI genuinely good at here?
Reading. Specifically, reading volumes of unstructured text and telling you what is inside it, which is the slow part of product research and the part where a model earns its cost.
Feed it the one and two star reviews of three competing products and ask what people complain about. You will get a usable list of failure modes, and failure modes are product opportunities in a way that demand curves are not. The complaint that a jacket's pockets are too shallow is a specification. The observation that jacket searches rose 12% is not.
The same applies to your own inbox. Pre sale questions are a record of what buyers could not work out from your page, and the ones that repeat are either a product gap or a copy gap. Sorting a year of them by theme is an afternoon of human work and a minute of machine work, with no numerical claim involved and therefore nothing to misread.
Where it goes wrong is the moment you ask for a figure. A model asked how many people want a product will produce a plausible number, because producing plausible text is what it does. Unless that number is traceable to a source you can open, treat it as prose. We went through the general shape of this problem in how to keep an AI drafted product page factual, and the discipline is identical: the model may write, the numbers come from somewhere you can click.
Which checks actually prevent an expensive mistake?
Not the demand ones. A product that sells less than you hoped costs you a slow quarter. A product you are ordered to stop selling costs you the stock, the listing and sometimes a legal bill, and those risks are both checkable in advance and routinely skipped.
The first is the name. The USPTO's guidance on searching the trademark database puts the purpose plainly: a search done before you apply tells you whether your mark is available for your particular goods or services and whether an existing mark conflicts with it. The phrase carrying the weight is for your particular goods or services. Trademark rights are scoped by class, so a name being used elsewhere does not always block you and a name being free in a web search does not mean it is free in your class. That is a twenty minute check against a register, and no AI answer substitutes for it.
The second is whether you may lawfully place the product on your market at all. For anyone selling into the European Union, the General Product Safety Regulation has applied since 13 December 2024. It reaches a broader set of goods than the directive it replaced, including products sold online and items that are used, repaired or reconditioned, and it also captures products designed for professional use that end up reaching consumers.
The practical version for a small importer is that a product needs an economic operator established in the EU who is accountable for it. If you are buying from outside the bloc and selling to consumers inside it, that role has to be filled by someone, and discovering this after the pallet arrives is an expensive order of operations. The listing side of the same problem is covered in what an automated listing compliance check can and cannot catch.
The detail that catches resellers is the breadth. The regulation covers used, repaired and reconditioned goods alongside new ones, so a shop selling refurbished items is inside it rather than outside. There is an exemption list, and it is narrow in a way that rarely helps an ordinary retailer: medicines, food and feed, living plants and animals, plant protection products, antiques, and products clearly marked as needing repair before use. Vintage clothing is not antiques. Refurbished electronics are not exempt.
The regime also has teeth on the surveillance side. The rapid alert system formerly called RAPEX now operates as Safety Gate, national authorities have stronger inspection powers, and consumers have more direct routes to report an unsafe product. For a merchant the meaning is simple enough: an unsafe product is likelier to be noticed and the notice travels faster than it used to.
How do you test demand without buying stock?
By selling the thing before you own it, in a way that is honest about what you are doing. This is the step that replaces most of the research, and it produces the only data that settles the question.
The cleanest version is a real product page with a real price and a waitlist instead of a buy button. Not a survey. Surveys measure politeness. A page that asks for an email address in exchange for being told when something arrives measures a decision, and the conversion rate from your existing traffic tells you more than any index. If a hundred visitors see it and nobody signs up, you have your answer for the cost of an afternoon.
The stronger version is a pre order with payment taken and a stated dispatch date, which converts interest into a number with money attached. That raises obligations, since taking money for goods you do not hold means consumer protection rules on delivery timing and refunds apply, and the dispatch date has to be one you can hit. It is a real commitment rather than a test, and it should be treated that way.
Either way, the sample is small and local and yours, which are the three properties the free demand tools cannot offer. Twenty sign ups from your own traffic is a weaker statistical result than a national search index and a far better basis for a purchase order, because those twenty people already know your shop, your prices and your shipping.
A research order that does not waste money
Work from cheapest and most certain to most expensive and most speculative, which is the reverse of how most people do it.
Start with what you already own. Site search queries with no results, pre sale questions that repeat, and the products customers ask you to restock. This costs nothing and it is evidence about actual buyers rather than about the internet.
Then check whether you are allowed to sell it. Name against the register, safety and compliance obligations for your markets, any category restriction on the channels you use. Doing this second rather than last is the single change that saves the most money, because it can end the investigation before you have spent anything.
Then, and only then, look at demand tools, and look at them for shape rather than size. Is interest seasonal, and does the season match your cash cycle? Is the direction up or down across several years? Does a related term tell a different story from the one you searched for? Those questions survive sampling and normalisation. How many units will I sell does not.
Finally, price the downside rather than the upside. The number that decides whether a product idea is sane is not projected revenue. It is what happens if you sell a third of what you hoped: whether that stock is storable, returnable, discountable, or a write off. Suppliers will tell you the minimum order quantity. Only you can work out what that quantity costs you when the forecast is wrong.
What about categories rather than products?
Worth a separate thought, because the research is different. Adding a product to a category you already sell uses information you have: your own margins, your own shipping profile, your existing audience. Entering a new category discards all of that and rebuilds it from nothing, which is why it so often fails for reasons unrelated to demand.
A shop that sells candles and adds candle holders is doing product research. The same shop adding electronics is starting a new business with a familiar logo, and the questions it needs answered are about supplier reliability, return rates and compliance, not about search interest. Keeping your catalogue coherent also happens to be what makes it legible to machines, a point we covered in how a clean category structure changes what automated tools can do with your catalogue.
The deeper version of this is infrastructural. A shop that can query its own orders, its own search log and its own return reasons has better product research inputs than any external tool sells, and a shop that cannot has to buy its answers from someone else's sample. That is one reason to own the schema behind the storefront, which is the argument set out in what an AI built store gives you that a rented one does not.
Read the zero. Search the register. Ask your own customers first. The tools will still be there afterwards, and you will know what their numbers mean.