Hacker News Reader

Comment by jdoliner | Hacker News Reader

It's not totally obvious to me that you can get the economics of this to work. A Google search costs ~.04 cents to serve, whereas a frontier reasoning LLM request costs about 2 cents. The revenue from a Google search is also around 2 cents. So the margins are dangerously thin on an LLM.

Now there's lots of variables that can be tweaked on this. So it's possible to get it to work. But there's a lot less room for error.

View on HN

The obvious knob to turn here is that the floor price of ad auctions will be incredibly high, with the justification that AI is expensive.

As someone outside of the ad-tech space it blows my mind how much Instagram and Google ads cost these days, and OpenAI would certainly want to price their ad offering as more “premium” (see: $$$).

The pricing doesn't really work like that - it's more an auction amongst different companies wanting to sell their services so the price comes down to how much the companies make from customers acting on the ads. Which is why you have to pay crazy money to advertise a mortgage for example - the mortgage companies make a lot if they sell one. I think OpenAI could do well at it.

Every Google search request triggers a Gemini request tho.

Which is great… that's why I don't use chatGPT at all, having a LLM summary + a list of websites to deepen the search if I need, is just a superior user experience IMO.

I don't know anything about Google's architecture but I would guess that the average Gemini request per search query is < 1, surely there's a lot of caching that can be done and a lot of money to be saved by doing so.

Except Gemini has lied to me about local events, it told me that in my city (specifically in my city, mentioning it by its name) a musical event was happening and I lost transportation time and cost, so it can be pretty spotty.

Every unique Google search triggers a Gemini request.

I suspect that just like with Search, LLMs are used for a number of different action types. One being specifically web search, one being product search and so forth.

Within the web-search and product-search requests there is undoubtedly A LOT of overlap between peoples queries. It would not be unfeasible to have on nice long good answer generated by e.g. ChatGPT 5.1 cached, and first throw the initial user request into some kind of classifier and use a smaller LLM to judge whether the cached answer is close enough to the initial query.

Maybe they'll gamify it with credits and make you watch ads for in order to gain them and thus use the service for free.

They will do both

going off my example - chatgpt knows everything about me including desires, goals and issues im struggling with.

combine this with the fact that i have disposable income.

i can't fathom how much advertisers are willing to pay to put themselves in front of my eyes vs a google search for "dining table"

This isn’t accurate even for API prices for a request/response.

Go on something like openrouter with gpt 5.1 and use the chat then check the billing and you’ll see an average joe query is like 0.00102 or something.

You’re quoting figures from articles for initial ChatGPT release in 2022

The price of inference has been dropping like a rock. I wouldn't expect that 2c to be true in a couple of years.

Likewise, was the cost of a Google search 20 odd years ago those amounts?

Presumably they'll be embedding commercial influence in the interaction where you have no clue ad dollars are steering the conversation.

That will no doubt have higher value than Google's $.02/search revenue, since the users will be completely incapable of separating the wheat from the chaff.

Reply on news.ycombinator.com