Where a licensed brand is on sale, and who made it
I’m David McMillan, a forward deployed engineer. For a brand licensing agency, I built a platform that finds one brand’s products for sale across seven marketplaces and search sources, works out which listings fall inside the agency’s programme and who made them, and sets them beside what licensees report. I built it in July and August 2026. It is live, behind a sign-in, and the agency uses its data.
was Where the products sell, as licensees report it now Where they sell, seen in 60,841 listings from seven sources
- Sector
- Brand licensing
- When
- July to August 2026
- Systems
- Seven marketplaces and search sources, the agency’s licensing exports
- Status
- Live
- Rev
- B · 14.10.2026
The work
A licensing agency signs companies to make products under a brand, then watches the programme: whether licensees pay the right royalties, who sells the brand without a licence, and where licensed products sell. This agency wanted all three for one brand, whose products it licenses in toys and collectibles. Its licensing system held the approved products and what each licensee reported. Its brief asked for the rest: which retailers and countries sell the products, at what price, and which listings look like copies.
Constraints
A royalty check built from public listings is a reason to ask a licensee a question, never proof. The platform sees only what is listed online, and it reads titles, the only listing text it collects.
What I found
Most listings that carry the brand’s name are outside the agency’s programme. Of the 44,392 listings about the brand, the remit check on 19 August 2026 placed 12,207 inside it. The rest were another company’s products, other product categories, products licensed outside the programme, or titles the remit check could not place.
What I built
Seven sources feed it: Google Shopping, eBay and AliExpress, and 1688, DHgate, Taobao and Shopee, marketplaces in China and South-East Asia. The agency’s exports of approved products and royalty reports sit beside them. Every listing keeps its source, its address, a reliability tier, and how and when it was fetched.
A language model reads each listing’s title and marks it relevant, borderline or not relevant to the brand. The rest of the platform is rules, with no model in them. Sellers are separated from the marketplace they sell on, makers are read against a registry, and the product codes in a title narrow each listing to a short list of the agency’s 651 approved products.
For listings inside the programme, resale left out, a queue sorts each by its maker: a maker whose rights the agency can show, a licensed maker named on a marketplace in China or South-East Asia, a maker whose rights it cannot show, or no maker the registry knows. AliExpress listings show where the brand sells and stay out of the queue. The platform flags likely counterfeits, each with the evidence behind the flag, and a person decides what happens next.
The royalty check runs product by product: what each licensee reports for each product in each country, against where that product is on sale. It also lists licensees who hold approved products and filed no royalty in the period loaded.
Controls
Every billed model call is logged with its tokens and cost, and a run will not start if its estimated cost is over the limit set for it. Paid searches are capped per run unless the cap is lifted on purpose. A named person can switch the model off, for every call or for one model, and every call checks the switch first.
Nothing goes to a seller until a person at the agency approves it on a counsel-approved template; the evidence is then frozen, and sending is a second, separate action. The database refuses to record rights for a maker until a product code or a person confirms them.
Each brand’s data is kept apart: each person signs in with their own account and sees only the brands they are assigned, and the database itself keeps one brand’s rows from another’s. Every operator action and every external call is written to an audit log that cannot be edited.
What happened
It went live behind its sign-in on 10 July 2026, two days after the build started. On 20 July the sort of 60,841 listings stood at 44,392 relevant, 1,329 borderline and 15,120 not about the brand, after one revision to the model’s instructions and a hand correction of 80 listings.
Three times on this programme, a rule that was right in general would have put someone’s lawful product in front of the agency, and each was caught by measuring it after it was written. So the tests hold 30 real listings that are legitimate and look suspicious, and fail if any of them is flagged. A new signal ships only if all 30 stay clear.
On 19 August the queue held 9,364 listings: 5,028 from makers whose rights the agency can show, 130 naming a licensed maker on a marketplace in China or South-East Asia, 2,109 a question of rights, and 2,097 naming no maker the registry knows.
Start with a call
The call takes 30 minutes. I ask how the work moves today, not which tools you use.
I’m based in Lisbon, Portugal, and work UK hours. I’ll be on site anywhere in the UK and Europe, with travel inside every fixed price.
bigbadbird.ai/work/brand-licensing-intelligence · dm@bigbadbird.ai · Built 15.10.2026