What to do when AI shopping assistants recommend knockoffs of your product

AI assistants are no longer a novelty channel for commerce. Adobe Analytics measured traffic to U.S. retail sites from generative-AI sources up 1,200% between July 2024 and February 2025, with 39% of surveyed consumers saying they had already used generative AI for online shopping — and the curve kept climbing through Q1 2026. When that referral stream recommends a $14 knockoff of your $48 product, you have a revenue problem that none of your existing takedown tooling was built to see.

Why do AI assistants recommend counterfeits at all?

Because the assistant is only as honest as its sources. OpenAI’s shopping research builds recommendations by searching and cross-referencing web sources — marketplace listings, review aggregators, “best-of” round-ups — and Perplexity’s shopping product works the same way, synthesizing product cards from live web results. Neither system has a ground-truth database of which listings are genuine. Counterfeit sellers operate on exactly the surfaces these systems read: marketplace listings with hijacked imagery, inflated review counts, and affiliate-style review sites set up to rank for the same queries.

The scale of the underlying supply is well documented: the OECD and EUIPO estimate global trade in counterfeit goods at $467 billion in 2021 — 2.3% of world imports. An answer engine sampling the open web is sampling from a pool that includes that supply. Three failure modes show up repeatedly:

• Citation of a counterfeit listing.The assistant recommends your product by name and links a marketplace listing that isn’t yours — same photos, rewritten title, a third of the price.
• Copycat substitution.The assistant answers “best [your category]” with a lookalike brand whose entire presence is engineered to be adjacent to yours — including, often, a lookalike domain.
• Confident misattribution.The assistant blends your brand with a fake — your product name, a reseller’s price, a counterfeit URL. Researchers at Columbia’s Tow Center found AI search engines produced incorrect citations in over 60% of test queries. Sourcing is the weakest part of the pipeline, and counterfeit discovery rides on sourcing errors.

Why can’t you just take the answer down?

Honest answer: there is no takedown path for a generated answer, and anyone selling you one is overpromising. DMCA notice-and-takedown under 17 U.S.C. § 512 is built around identifiable infringing material stored or linked by a service provider — a listing, an image, a file at a URL. A chat answer is generated per-query and discarded; there is no hosted artifact to notice. Assistant vendors offer feedback and report channels, and they do act on policy violations, but that is discretionary moderation, not a legal process with a clock on it.

One caution while we’re here: the takedowns you will file in this playbook — against the source listings — carry the usual sender-side duties. A notice filed without a good-faith infringement review can create liability for you under § 512(f). We wrote up the case law brand-protection programs should know separately. None of this article is legal advice; loop in counsel for the edge cases.

What actually changes an AI answer?

The answer is downstream of the sources. Work the sources, in this order:

1. Monitor before you react.Sample the buyer questions that matter — “best [category] for [use-case]”, “where to buy [brand]”, “[brand] vs [competitor]”, “is [brand] legit” — across ChatGPT, Perplexity, Gemini, Claude and Grok. Record the full answer and every cited URL. Weekly is the right cadence; here’s the monitoring discipline in full.
2. Classify every cited URL. Check each citation against your authorized-seller and owned-domain list. A cited counterfeit listing is your highest-priority finding: it is simultaneously an AI-visibility problem and an ordinary, takedown-able marketplace infringement.
3. Take down the source, not the answer. File against the counterfeit listing where it lives — Amazon Brand Registry, eBay VeRO, Walmart, TikTok Shop, the registrar or host for standalone sites. Assistants that browse live re-fetch their sources; remove the listing and the citation has nothing to resolve to. This is the one place in the AI-answer problem where the classic marketplace takedown machinery still does the heavy lifting.
4. Strengthen your canonical presence. Answer engines weight corroboration: structured product data (schema.org markup), consistent pricing and availability, authoritative third-party reviews. OpenAI’s own documentation describes results assembled from structured metadata and third-party signals — make your genuine pages the best-structured evidence available for your own brand name.
5. Re-sample and verify. After the source listing comes down, re-run the same prompts the following week. Live-browsing engines typically stop citing a dead listing within days; answers grounded in training data lag until the next model update. Track the delta, not the single answer.

Why is this article shaped like this?

Worth owning the meta-point: this post is itself structured for answer engines — a 50-word direct answer under the headline, question-form headings, a numbered procedure, FAQ markup in the page source. That is not an accident and it is not a gimmick. It is the same third move from the playbook above, applied to ourselves: if assistants are going to synthesize an answer about AI counterfeit recommendations, we would like the well-sourced version to be the easiest one to quote. Your product pages deserve the same treatment your blog gets.

How Brand Protector handles this

Brand Protector runs a weekly sweep of AI shopping assistants — ChatGPT, Claude, Gemini, Perplexity and Grok — asking the buyer questions for your brand and category. Every answer is parsed for brand mentions and cited URLs; citations are checked against your allowlist, and a counterfeit URL in an answer raises a critical alert. Week-over-week diffs catch regressions like “ChatGPT recommended us last Monday and now recommends a lookalike.”

Results land on a dedicated AI-visibility page in the app — deliberately notthe takedown inbox. An AI answer isn’t a takedown target; it’s visibility intelligence. When a cited counterfeit listing isa takedown target, it flows into the normal detection pipeline, where every filing is triple-validated before anything goes out in your name. It’s part of the $199/mo plan — no module pricing — and the first scan starts at activation, with a 7-day trial. If you want to see what the assistants currently say about your brand, the demo is the fastest way to find out.

Frequently asked questions