When the AI Shopping Assistant Recommends the Fake

The email arrived at 06:40 on a Tuesday, which is how Mara found out.

"The clasp snapped on the second wear," the customer wrote. "I know it's handmade, but 40 dollars is 40 dollars." Attached was a photograph of a brass cuff with a hinge that had sheared clean through the solder.

Mara makes brass cuffs. She has made them for 9 years, in a converted garage outside Leeds, and she has never in her life sold one for 40 dollars, because the brass alone costs more than that. She sells them for 148.

She wrote back asking for an order number. There wasn't one. The customer had asked an AI assistant where to buy a hand-forged brass cuff like the one she'd seen on Instagram. The assistant had given her three options. She'd picked the cheapest.

That's the whole story, and it took Mara about 20 minutes to find the listing, and about a week to understand what had actually happened to her business.

The thing that broke wasn't search

Here's what Mara assumed, because it's what everyone assumes. She assumed she'd been outranked. Some copycat had done better SEO, climbed above her in Google, and intercepted a buyer on the way to her shop.

She hadn't been outranked. Her shop was still the top organic result for her own brand name and for most of the phrases people use to find what she makes. The copy wasn't competing with her in search at all. It had never needed to.

What had happened was that a buyer skipped search. She described what she wanted in plain language to an assistant, and the assistant did the shopping. It read listings, compared them, and named a seller. Mara's shop was one of the three it considered. It just wasn't the one it put first, because the assistant had nothing to work with except what the listings themselves said, and the fake's listing said everything hers did, in slightly better English, for a quarter of the price.

The assistant wasn't tricked. It did exactly what it was asked. It found the cheapest item matching the description, and the description matched because the copy had been built from Mara's own photographs and her own product titles.

This is the part that takes a while to sit down properly. A counterfeit listing isn't just stealing a sale from a customer who got confused. It's feeding a machine that has been asked to make a recommendation, and it's feeding it better than you are.

How big this already is

The numbers on counterfeits are old news to anyone in the business, and worth restating because they are larger than most founders assume. The OECD and the EU Intellectual Property Office put global trade in fakes at 467 billion dollars in 2021, or 2.3% of all global imports, in their report Mapping Global Trade in Fakes 2025, published in May 2025. Imports of fakes into the EU alone came to 99 billion dollars, 4.7% of EU imports.

Those are customs figures. They count physical goods crossing borders and getting seized. They don't count the listing that sits on a marketplace for 11 weeks, sells 300 units direct from a warehouse in another jurisdiction, and never meets a customs officer at all.

And they predate the thing that's changed. In 2021, a copycat listing had to win a search result or an ad auction to reach your buyer. Now it can win a sentence in a conversation.

Why an AI answer is a harder problem than a search result

A search results page gives the buyer ten options and a visible brand name attached to each. The buyer does the sorting. They can see that one of those results is a shop with 4,000 reviews and an 11-year history, and that another is a shop opened in August.

An answer gives them one option, or three, with the sorting already done. The reasoning is invisible. If the assistant got the wrong idea about who makes a thing, there's no SERP for the buyer to scroll past and correct it.

Three consequences follow, and they're the reason brand protection stopped being a legal chore and became a visibility problem.

The copy inherits your credibility. If a listing uses your photographs, your product name and your phrasing, an assistant reading that listing sees a product that matches the brand it has learned about from your site, your press coverage and your reviews. The trust you built attaches to the listing, not to you.

Price does disproportionate work. Ask any assistant for the best value option and it will weight price heavily, because price is the one attribute in a listing that's unambiguous and comparable. Craft, provenance and durability are adjectives. 40 dollars is a number.

The correction is slow. A wrong search result is gone when the page reloads. A wrong association between your brand and a seller who isn't you persists in the material these systems read, and it keeps being read.

What brand protection means now

The old version of this work was a legal function. You found out about a copy because a customer complained, you forwarded it to a solicitor, the solicitor sent a letter, and 6 weeks later something happened or didn't.

The version that works now is closer to monitoring than litigation. Five things, in this order.

Know what your fingerprint is. Your photographs are the most copyable asset you own and the most traceable. A copycat has to show the buyer what the product looks like, and the cheapest way to do that is to use your picture. That's a gift, because an image is searchable in a way that a paraphrased product description isn't. Your product photos are your fingerprint goes through how to find out who's using yours.

Watch the places copies actually appear. Not the whole internet. The handful of marketplaces where someone selling a copy of your product can list it in 10 minutes and reach buyers on day one.

Know which notice to send. A copied photograph and a copied brand name are two different legal problems with two different forms, two different evidence requirements and two different success rates. Sending the wrong one costs you weeks. DMCA or trademark? is the decision in one page.

Expect it back. A takedown that sticks is the exception, not the rule, and the seller who relisted is more valuable information than the listing that went away. The relisting problem covers what to watch after the first notice.

Make your own store unmistakable to a machine. This is the half of the job that has nothing to do with takedowns. If the assistant knows which store is officially yours, the copy has to beat you on merit rather than on confusion. Entity hygiene is the schema and sameAs work that does it.

Notice what isn't on that list. Nowhere does it say "find a lawyer". You'll want one eventually, for the cases that go past a form. Most cases don't.

The part the software can't do, and shouldn't

I want to be direct about this, because it's where most of the tooling in this category oversells itself.

Software can search. It can compare a photograph against a listing image and give you a similarity score. It can notice that a title repeats your brand name, and that the price sits 75% below yours, and that 6 other listings come from the same shop. It can save a screenshot with a timestamp so the evidence survives the listing being pulled. It can draft the notice.

Software cannot look at a listing and know it's a copy. It doesn't know which of your distributors you authorised last March. It doesn't know that the shop in Portugal is your sister's. It doesn't know whether the similarity is theft or coincidence, and in this category a false accusation is not a small error. A takedown notice is a sworn statement. In the United States it carries a penalty of perjury under 17 U.S.C. 512(c)(3), and 512(f) makes anyone who knowingly misrepresents liable for the damages they cause. In the EU, the Digital Services Act requires platforms to give the seller a statement of reasons and a route to contest it, so a bad notice doesn't vanish quietly either.

So the division of labour is not negotiable. The machine sorts and explains. The human confirms and files. Any product that offers to do the confirming for you is offering to put your name on a sworn statement it can't verify, and you should walk away from it. Human in the loop, done properly sets out what that line looks like when it's drawn honestly in the code rather than in the marketing.

This is also why we built the thing we built. Brand Protection Tools, which Verity AI runs, finds the listings, groups them by seller, saves the evidence and prepares the notice. Confirming and filing sit behind the customer's click, and that's a decision in the architecture, not a setting. It searches Etsy and Alibaba.com today, with AliExpress, Temu, eBay and Amazon next. We'd rather say that plainly than publish a coverage map we can't stand behind.

If you're a board, this is a control gap

For anyone reading this from the governance side rather than the founder side, the uncomfortable version is short.

Most organisations can tell you their trademark portfolio and their legal spend. Very few can tell you how many listings copied their products last quarter, how many notices went out, how many stuck, and how many of the removed listings came back under the same seller. Those four numbers are the only ones that describe whether the control works, and the absence of them is usually not a budget problem. It's that nobody owns it.

What a board should ask about marketplace risk is that conversation written down, including the questions that have no good answer yet.

Where to start if you're Mara

She didn't need a programme. She needed to know how bad it was.

The first useful hour is one product, reverse-image searched, and an honest count of what comes back. The second is the right notice on the worst offender. After that it's a cadence rather than a project, and the cadence is cheap. Brand protection on a small brand's budget is the version that fits a business of 1 person, and the 30-day plan is what the first month looks like if you'd rather have a sequence than a to-do list.

One more thing, because it's the piece founders skip. Fixing the copies does not fix the recommendation. If an assistant has learned that your product comes from several sellers and one of them is cheaper, removing the cheap one leaves the assistant with a gap, not an answer. You have to tell it who you are, in the structured, boring, machine-readable way it understands, and keep telling it. How copycat listings leak into AI answers is the mechanism, and the fix.

Mara's cuff listing came down 9 days after she filed. The same seller had three more up within a fortnight, two of which used a photograph she'd posted on Instagram in 2024.

She knows about those ones. That's the difference.

---

Brand Protection Tools, which Verity AI runs, is software for finding and documenting copies and preparing notices. It isn't legal advice, and it doesn't file anything on your behalf. If you want a second pair of eyes on how AI assistants currently describe your brand, that's the work we do at Verity AI. Start a conversation.