The Light That Wasn't What It Seemed and What Agentic AI Could Do About It
Yesterday I spent an hour looking for a ceiling light for my study.
I found it almost immediately, a beautiful wood and metal, low-profile LED fixture with a suitably Scandi name, on what looked like a reputable, well-designed UK lighting site. The kind of thing you'd see in a Sunday supplement and instantly want.
Then I did what I always do before hitting 'buy'. I searched around. And I found the exact same light. On five other sites. All with different Scandi-adjacent names. All with slightly different prices. All with the same product photographs.
A deeper dive, and the truth emerged: it's a Chinese product from an anonymous manufacturer, available on Amazon with a string of poor reviews. Complaints about flimsy construction, poor wiring, and a finish that doesn't live up to the photographs.
Dozens of UK and European lighting sites had dressed it up, given it a story and priced it like it had one. Yet none of them disclosed where it actually came from. (Dutch regulators recently fined a retailer €90,000 for exactly this kind of non-disclosure on Chinese drop-shipped goods.)
The real question it left me with wasn't about that one light. It was bigger than that.
Where Did All the Quality Go?
I know great British lighting exists. Tom Raffield makes extraordinary steam-bent wooden lights by hand in Cornwall. There are genuinely brilliant independent British makers, artisan brands, small-batch designers, things that are actually worth the price tag. But they were invisible to me during that search.
The web's current architecture rewards scale. The brands that win in organic search are the ones that have invested in SEO, in paid placements, in product feed optimisation, not necessarily in the quality of what they sell. The result is a kind of quality inversion: the products you see most are the ones most aggressively marketed, not the most deserving of your attention.
Dropshipping has industrialised this problem. Anyone can spin up a beautiful-looking website, source a product from a Chinese marketplace, give it a name, and rank it on Google. Increasingly, AI-generated product descriptions and AI-built storefronts are making this even faster and easier.
So: Will Agentic Search Fix This?
Agentic commerce. AI agents that don't just surface results but actually research, compare, validate, and purchase on your behalf are no longer theoretical. Google now has agentic checkout live in the US. Amazon's Rufus handles hundreds of millions of interactions.
The promise is compelling. Instead of opening ten tabs and getting fooled by white-label rebranding, you tell your agent what you want, and it does the legwork: comparing specs, cross-checking reviews, verifying origin, evaluating return policies, and surfacing options you'd never have found manually.
In theory, this should be the end of the Scandi-named Chinese lamp problem. An agent that reads every review, cross-references a product against multiple retailers, and checks the manufacturer's actual identity should see straight through the kind of deception I encountered yesterday.
But here's my concern…
The Bias Problem No One Is Talking About
Research published in 2025 from teams at Columbia University, Yale, and MyCustomAI found that AI shopping agents (including GPT-4.1, Claude Sonnet, and Gemini) show measurable biases in how they make purchase decisions. They favour products higher up on pages. They love "Best Seller" and "Overall Pick" badges. Different models prioritise different signals, one might weight ratings heavily, another might weight positioning or even price boundaries.
What this means in practice: agentic search may not liberate shoppers from the dominance of well-resourced retailers. It may simply automate the same popularity bias that already exists, but faster and at scale.
The products most likely to be recommended by AI agents are the ones whose data is most structured, most machine-readable, and most aggressively optimised for agent discovery. Brands that invest in AI sitemaps, product catalog APIs, rich structured data, and LD+JSON schema will surface. The independent artisan in Cornwall with a beautiful product and a modest website may not.
There is also the question of what training data AI agents are actually learning from. If the web is already saturated with cheap imports dressed in premium branding, the agents trained on that web will reproduce that bias in their recommendations unless they are specifically designed to interrogate provenance, quality signals, and independent verification.
The Case for Cautious Optimism
There are genuine reasons to think agentic search could improve things if it matures thoughtfully.
Agents that truly cross-reference reviews across multiple platforms will find the Amazon listing with the poor reviews, just as I eventually did, but in seconds rather than hours. Agents capable of verifying manufacturer origin, checking returns geography, and reading the small print that most consumers skip could make the dropshipping deception model significantly harder to sustain.
Some researchers argue that AI agents actually disadvantage big brands in one key way: they evaluate semantic specificity rather than brand recognition. A small maker whose product descriptions and reviews precisely match a buyer's requirements will beat a generic bestseller if the match is strong enough. That's a genuine opportunity for quality independents, but only if they make themselves legible to machines.
Only 12% of shoppers currently trust AI-generated recommendations outright, while 45% still trust human reviews. That trust gap is actually a window. Agents that are transparently sourced, demonstrably unbiased, and rooted in verified review data could earn the trust that keyword search never really deserved.
What We’re Trying to Drive
For agentic search to genuinely improve customer experience rather than simply accelerate the existing race to the bottom, a few things need to happen:
Provenance transparency: Agents should surface, not bury, the country of origin, manufacturer identity, and where returns go. This should be non-negotiable.
Review authenticity signals: Agents need to weight independently verified, cross-platform reviews over on-site testimonials that can be curated.
Discovery beyond the popular: The most valuable thing an agent could do is surface the product I didn't know to look for. The genuinely excellent independent maker who has never had the budget to compete on Google Ads.
Quality as a first-class signal: Not just ratings, but specifications, materials, longevity data, warranty terms, and carbon footprint. The things that actually matter for a considered purchase.
The irony is that the ceiling light that started all this will likely still be the first result - agentic or otherwise - until the agents are trained to see through the surface and interrogate what's underneath.
And that's entirely a design choice. Not an inevitability.
What's your experience? Are AI shopping tools making you more confident in purchase decisions - or just faster at making the same mistakes?