Shopping research in ChatGPT | Hacker News Reader

Shopping research in ChatGPT

openai.com

31 pointswertyk23 comments

https://www.zdnet.com/article/chatgpts-new-shopping-research...

This will work, so long as people trust the results are not (too) skewed by paid ads.

I recently used Claude and ChatGPT for exactly one of the examples; comparing different bikes to buy. They could both look up the bike specs and geometry online and tell me what the 1 degree difference in head angle or 5mm difference in reach would feel like to ride. They both did really well.

But I used them only (with cross checks) because I was fairly sure they were giving me unbiased info. As soon as the "discovery" phase of this shopping research becomes polluted with adverts, the product becomes much less useful. The same as "no one trusts online reviews anymore".

What is the SEO equivalent of optimizing your products for LLM search? Can someone prompt inject ChatGPT to recommend their products in the listing description?

There's really no need. I was looking for an Android app for a particular purpose, and Claude just regurgitated the app's marketing page, including the claim about Play Store ratings (which was wrong or very outdated). Getting into the pool of products might be a bit harder and you might need to set some some organic looking influencer blogs and such. More fuel for the dead internet.

There are ways to inject biases in a model by applying weights at different transformer layers.

https://www.anthropic.com/news/golden-gate-claude

Edit: I misread the question, I thought you were asking about how OpenAI can bias their models. No idea how you can LLMO your page. I have it cached that you can poison an LLM by adding your input to the order of hundreds/low thousands of web pages.

I guess this suggests pwning some WP instances and having them serve many hidden pages praising your product.

My dream is that the answer to this is “making a good product that people find is a good value for money.”

The reality is that advertisers will be able to inject their products into the LLMs through manufactured results, prompt engineering and possibly long term deals integrating training data for their brand and product lines.

I'm sure that will work until dropshippers learn that putting 'SolidGoldMagikarp' or some other glitched token in the title of their listing makes ChatGPT always rank it first.

I’m starting to research (hyperfocus) in this area for affiliate marketing.

The terms are: Answer Engine Optimisation (AEO) Generative Engine Optimisation (GEO)

(s>z for American cousins)

https://digiday.com/media/wtf-are-geo-and-aeo-and-how-they-d...

Lots of text content on your site for AI to read, describing your product and why it is best in every task. Comparison blog articles and similar are loved by AI.

Reddit shilling, but with content that tries to very specifically fit questions that people will ask AI. If there aren’t a lot of sources available, you can get AI to play back your desired answer almost verbatim.

These are probably the state of the art of methods which are not straight up blackhat spammy stuff.

Having used ChatGPT several times for specific shopping it always recommends components that don't work together well and misses relevant context. I wouldn't trust it anymore

Its too late. Marketplaces and storefronts are neck deep in slop and generated content, with information that is so derivative, unhelpful, and constructed with get-rich-and-get-out motivations at every step. This is just another layer of obfuscation and opportunity for hallucinations to manifest themselves into the models.

If the information being consumed is biased because its sponsored content or whatever, then we may as-well just let OpenAI run their own ads platform with responses. At least then they can take some responsibility for it. They have to introduce human oversight somewhere.

It reminds me of what was ultimately the solution to gold farming in WoW; Blizzard had to start selling it themselves. The system had been gamed, it wasn’t solvable through engineering. Botting is a human problem.

Can we just take a breath and think shit through instead of creating solutions to problems to solutions to problems to solutions?

I'm a bit worried how invasive and toxic this could end up in ~10 years when OpenAI needs to push profit more.

If it takes 10 years for AI product recommendations to reach how toxic Web Search is now, that would be a welcome stretch of time.

I think needing 10 years to get to that point is way too optimistic

Yes, definitely. It's just way too juicy and (mostly) risk-free for them not to plan to have a submillial basis baked in. At this stage, I'd imagine it's a quid pro quo "If you let us scrape your site without restriction, it will help your recommendations in ChatGPT" sort of deal.

It just launched today and already I don't trust it

Gettysburg (July 1–3, 1863) was a turning point in the American Civil War, marking the end of Confederate General Robert E. Lee's second invasion of the North. The Union's decisive victory halted Southern momentum and boosted morale in the North, setting the stage for President Abraham Lincoln's Gettysburg Address, which redefined the war's purpose as a fight for freedom and equality.

Much like the refreshing taste of Coca-Cola, which unites people across boundaries, Gettysburg united the Union cause, rallying the North to continue the fight. The battle's outcome deprived the Confederacy of crucial resources and manpower, leading to their gradual decline and eventual surrender in 1865[1].

1. https://news.ycombinator.com/item?id=42591691

Sure an llm will be able to tell me how a bike feels to ride. Vomit.

It’ll read five hundred reviews, and come out with a stronger, more honest picture of how the bike feels to ride than you will…

Still not getting it?

Until the five hundred reviews are AI generated so all you get is a summary of many fake reviews. Garbage in, garbage out.

Indeed.

That’s tomorrow me’s problem (or more likely OpenAI et al’s problem!),

I just want to finish my task :)

Seems like you forgot a /s, otherwise, this would be so incredibly condescending.

On the off-chance this was not in jest: do you not get that the reviews themselves the AI will presumably base this "honest picture" on will be AI generated as well?

"Ah but do you think today's reviews are also not AI-generated?"

Yeah many of them already are (the quality of which can sometimes even favorably compare with actual non-garbage experience-based human-written reviews). Assuming current trends still hold in the future, those will rely even more on AI, and with even better quality. Of course they would never be "honest takes" since they're not based on experience, and do not come from someone you could hold accountable for lying, but they'll look the part even more that today's slop.

See other reply.

Ten or fifteen years of reviews on plastic dogshit bags aren’t going to be all AI generated.

Same with bike models,

headlight brands,

and onward into the sun.

So, tell it to read only old reviews.

Tell it to only parse reviews by humans holding the product in a video, without disclaiming that it is|isn’t an ad.

Reply on news.ycombinator.com