Competitor assortment shifts
What a defined set of competitors added, dropped, or moved to the front page since last month, checked on a schedule rather than when you remember.
Using an AI Agent for Product Research: What to Sell Next
Guide — — by Mahmoud Zalt
How to put an AI agent on product research without ending up with the same dropshipping list everyone else already has.
Split the job before you automate it
Product research is two jobs wearing one name. The first is gathering: finding what exists, what it costs, who sells it, what buyers complain about, whether interest is rising. That is slow, repetitive, and exactly what software should do. The second is judging: deciding whether this fits your brand, your margins, your supply chain, and your appetite for risk.
Almost every disappointment with AI product research comes from tools that blur the two, presenting a judgment as if it were a finding. A ranked list of trending winners looks like an answer and is mostly a repackaging of the same public signals every other subscriber to that tool is seeing on the same morning. If the list is the product, the edge is gone before you read it.
The version that actually helps is less exciting: an agent that does your gathering, against your criteria, in your categories, and hands you a comparable table. The judgment stays yours, and the advantage comes from covering more ground than a competitor who is doing this manually on a Sunday.
What to have it gather
Benefits
Complaints in reviews
The recurring gripe across reviews of an existing product category. Unmet needs are more useful than popularity, and far less contested.
Price bands
What the category actually sells for across sellers, so you know whether your intended margin exists before you commit to stock.
Question and forum demand
What people ask for and cannot find, in the places your buyers actually gather rather than in a generic trend index.
Supplier reality
Minimum order quantities, lead times, and whether a listing is a real manufacturer or a reseller, gathered from the portals themselves.
Your own data
What people search for on your store and find nothing. This is the highest-signal source you own and the one most stores never read.
That last one deserves the emphasis. Your internal search log is a list of things customers already wanted from you specifically, which is a fundamentally better signal than a global trend chart. No competitor has it. Most store owners have never looked at it. An agent can summarise it weekly in about a minute of your attention.
The brief that makes it useful
The difference between a research agent that saves you a day and one that produces noise is entirely in the constraints you give it. Vague in, vague out, and the vague version reads plausibly enough that you can waste a week on it.
- Name the category narrowly. "Home fragrance under 40 dollars sold direct to consumer" beats "home goods" by an enormous margin.
- Name the competitor set. Give it the eight stores you actually watch. A generic scan of the whole internet returns the generic answer.
- State your disqualifiers up front. Fragile, oversized, regulated, cold chain, anything with a battery. Half the ideas die here and you save the reading.
- State your margin floor. So the output arrives pre-filtered by whether the economics could ever work, instead of you doing that arithmetic later.
- Ask for evidence per row. Every claim comes with where it came from, so you can check the three that matter rather than trusting all forty.
That last constraint is what separates research you can act on from a confident-sounding summary. Insist that each row carries its source. You will not check every one, but you will check the ones that would cost you money to be wrong about, and knowing you can check changes how much weight the whole document deserves.
Making this a standing job rather than a one-off search is where it compounds. That is the shape hiring takes at Sistava: an AI Employee that runs the same gather every fortnight builds a picture of movement rather than a snapshot, this competitor has added three items in the category two months running, this complaint has grown, this price band is drifting down. Movement is the signal worth acting on, and it only exists if somebody is looking consistently, which is precisely the thing humans are worst at sustaining.
Where it will mislead you
Worth knowing before you trust an output. None of these are reasons not to do it, they are reasons to keep the decision on your side of the line.
| Failure mode | What to do about it |
|---|---|
| Popularity read as opportunity | A crowded category is evidence of demand and of competition. Weight the complaint data over the volume data. |
| Confident numbers with no source | Require a source per claim. Treat any unsourced figure as a hypothesis, not a fact. |
| Supplier listings taken at face value | Resellers present as manufacturers routinely. Verify anything you would commit stock against. |
| Trends that already peaked | Public signals lag. If you can see it clearly in a trend chart, the early margin is usually gone. |
| Fit assumed rather than judged | It does not know your brand, your customers, or what you are willing to be known for. That part is not delegable. |
The healthy way to hold all of this: the agent expands how much you can look at, and being able to look at more is a real edge when your competitor is checking three sites by hand. It does not tell you what to sell. Anything promising that is selling the same answer to everyone in your category, which is the definition of not an edge.
That split is worth building into the setup rather than re-deciding it every month. The way it works at Sistava, the gathering, the sourcing and the repeat schedule sit with the employee, while the call on brand fit and risk stays with you. You write the category, the competitor set and the disqualifiers once in plain English, and the same brief runs again without you rewriting it.
If your category has particular sources that matter, a trade publication, a specific forum, a supplier directory that has no API, you can train a custom AI Employee to watch exactly those. That is usually where the non-obvious findings come from, because it is the ground nobody else is covering systematically. If you would rather see how the gathering reads before setting anything up, the free tools at Sistava run in the browser and cost nothing to try.
FAQ
Can an AI agent tell me what product to sell?
It should not, and be sceptical of anything that claims to. It can gather what exists, what it costs, what buyers complain about, and what is moving in your competitor set, which is the slow half of the job. Whether something fits your brand, margins, and supply chain is a judgment with your name on it, and a tool that answers it is giving the same answer to every subscriber.
What is the best signal for product research?
Unmet need beats popularity. Recurring complaints in reviews of existing products point at gaps that are not yet crowded, whereas a trending list points at a category where everyone is already competing. The single best source most stores ignore is their own internal search log: searches on your site that returned nothing are requests from customers you already have.
How do I stop the research from being generic?
Constrain it. Narrow category, a named competitor set, explicit disqualifiers, and your margin floor. A generic brief scans the whole internet and returns the answer everyone else got. A specific brief covers ground your competitors are covering by hand, which is where the advantage actually is.
Can it check suppliers too?
Yes, including supplier directories and portals that have no API, using browser control. Ask it to gather minimum order quantities, lead times, and whether the listing looks like a manufacturer or a reseller. Verify anything you would commit real stock against, since resellers present as manufacturers routinely and that is an expensive thing to get wrong.
How often should this run?
Fortnightly is a good default for competitor and category watching, because what you want is movement rather than a snapshot. A one-off scan tells you the state of the world; a repeated one tells you which direction it is going, and direction is the part worth acting on.
Is this different from a dropshipping product finder?
Yes, in the thing that matters. A product finder gives every subscriber the same ranked list from the same data, which competes it away immediately. A research agent works your sources against your criteria and hands you evidence rather than a verdict. The output is less exciting and considerably more useful.
The realistic promise is not that software will find you a winner. It is that you get to look at fifty candidates properly instead of five badly, with the boring gathering already done, and that you keep the one part of this job that was ever going to be your advantage.