ChatGPT lost 22 points of web share in a year | Hacker News Reader

ChatGPT lost 22 points of web share in a year

aicharts.grok.me

25 pointsechohive4254 comments

Gemini is actually good enough for pretty much everyone I know who is not in tech and uses Android. These people are downright pleased with Google's AI summaries.

The fact that Google can push Gemini out in so many ways with their existing marketshare is mind boggling. The fact that ChatGPT hasn't lost 50-100% of web share shows it's Google's game to lose at this point.

> These people are downright pleased with Google's AI summaries.

Consider me part of "those people", I guess, but are you guys not, at this point? There was a period near the start where they were bad, but I find them perfectly usable now. I'll often open a tab, Google something, and just read the summary because it's basically instant and I don't need any more than just that, a summary.

I'm genuinely confused why people take so much issue with them. Sure they're still wrong sometimes, but it's blatantly obvious when that's the case, and I can just scroll past.

have you ever done the fun trick of asking any model to adversarially review itself? highly recommended practice for anyone who trusts a one-shot output

Sure, I've done this for more complicated questions and sometimes it reveals flaws. But I don't use Google for anything complicated. Generally it's just to get a quick vague definition of a term, pretty much.

is that functionally different than just using the search field in dictionary.com? I guess I'm just not seeing what the point of using Gemini is over a much more efficient feature that already exists

Sorry, I misspoke there. I do occasionally use it for outright word definitions (and in that case I find it more useful than, say, dictionary.com because my search bar defaults to it, it usually defaults to the most common definition, and it puts it in more easily-readable terms) but I more so meant that I'll use it for quick inquiries, often just to get a sense of what a thing is. For instance, some random recent searches:

  - cloudflare workers free tier
  - tubular lock
  - hasl finish
  - js null coalesce

For all of these it gives a perfectly acceptable result. So does the top search result, and if I want more info I'll click through to it, but it does the job of a summary just fine.

I'm not saying it's the best thing in the world, I'm just generally surprised by the amount of people that seem to hate it. It seems to do its job fine for me.

yeah, I think some of the vitriol is well-meaning but misplaced. when I think about the externalized costs (pollution, energy use, water use, rare earth mining, data center buildouts, etc) I think it's worthwhile to remain critical but people like to extend it, in their black and white thinking, as multi-categorical failures. for basic questions it's serviceable (though the number of responses that cite just Youtube videos means my skepticism for its output validity remains)

They're ok. But when I use a search engine it's because I want search results. If I wanted an LLM's interpretation I'd have just asked a search-enabled LLM in the first place.

True, but I find it's significantly faster than even "instant" chatgpt, and comes with the advantage that if the summary isn't what I need, I can just scroll down to see the real results (granted, which are usually useless these days).

I agree they should make them disable-able though.

> If I wanted an LLM's interpretation I'd have just asked a search-enabled LLM in the first place.

When you search on Google now that’s exactly what you’ve done.

Right, that's the problem I'm describing.

I rarely ever want to search. Searching is a means to an end, which is finding a piece of information, an answer. Whenever I'm scrolling through sites on Google Search, I'm just wasting my time filtering out misses. With Gemini and other AIs, I go straight to the answer and request a link (if not provided automatically) if I really want to check the source.

eh why would you check the source, who bother with minor things like that these days :)

Yeah, as someone used to searching for specific turns of phrase, having the query automatically interpreted as a question is quite annoying - for looking up song lyrics, game walkthroughs and articles by quotes I want a page list, not a summary

There's a RAM crisis on, you know, let me save you the compute!

yeah i'm genuinely baffled by this too. I read this attitude frequently and I'm tending towards the conclusion that its simply a loud minority of people. I don't know anyone who acts like this in real life.

How is an LLM summary really any worse than reading through random blogs, forum posts, or reddit? Those are somehow so much more reliable? If you want authority it gives you links, and if you don't find good enough links you can ask it a follow up--no need to be coy--just straight up ask it for more authoritative sources and it will usually surface them. It's so blazingly fast relative to doing genuine naive search, I really don't see the downside.

> How is an LLM summary really any worse than reading through random blogs, forum posts, or reddit?

For me it's that I find LLM prose annoying to read through, so if it's anything I want to really sit down and learn about, I prefer a proper blog/forum/etc writeup. But yeah, for quick summaries it kinda does just work at this point, fairly reliably and rapidly. I'm pretty confused by the attitude as well.

Web search and a chatbot AI are two vastly different interfaces with different aims.

At this point, with a mature Web and a mature toolbox, I know a lot of research tools and resources that I can tap before I get out the sledgehammer of Search or LLM. So, if I know a priori that it won't work to plug a term into Wikipedia, or Wiktionary, or YouTube or something, then I can choose Gemini or Google Search.

If I go to Search, then my query means "I want to find external links and find a resource to research this term". If I go to Gemini, that means that the term is probably out of my wheelhouse, and/or I really want to delve in for some insight about this term, and I probably do not want to poke 3rd-party sites about it.

The trouble is that Google sort of melds them into the same-same, either way I choose. It wastes resources and browser-page-real-estate on that AI summary that I've got to skip and sift through the Search misses. The LLM still insists on a RAG web search and refer me to some dumb links that I didn't want.

Google was famed for its fast searches and rapid results. They used to be good results. But today it's obvious why everyone let Search products decay into garbage: because LLMs were coming, and LLMs are going to supplant your standard searches anyway.

If only they could make it be wrong less? I know they need to have these results appear in record time, but I wonder the collective damage being done by just how horribly wrong the answer often is.

Can you point an example? For me, recently, it had the wrong price for cable-car fare in SF, but it did show (2015) in the result. That seems reasonable to me, it's not horribly wrong. Other stuff it's been accurate on.

I've been using Gemini as an advance search engine. Their main search with AI isn't that useful but Gemini is. They're the only AI online I will use and just for searching/compiling. For work/development it is Qwen3.8. Then again now that I found out Pi has a web search function I've been using Gemini less and less.

What do you mean by "Pi has a web search function"? Pretty sure Pi doesn't have one, except if you install some extension, but then Pi has every function.

This really doesn’t match my experience, as Gemini was bafflingly bad at any online [re]research when I tried it 2-3 months ago. Not that google’s ai models seem particularly good at anything of late, but given the fact that google is still the king of the hill in search (even if they mostly serve you garbage, they still crawl everything) Gemini being very noticeably worse at searching things online than claude or chatgpt was a read head scratcher. I wonder how our usage patterns differ.

But are they paying for Gemini?

It’s good. Returns a bespoke response is some ridiculous 0.3s or something like that which is mind blowing and I have no idea how they are pulling that off at all in the first place let alone the sheer scale of it.

> I have no idea how they are pulling that off

Google has the most AI compute by far. I don't have the link handy or I would share it.

https://epoch.ai/data-insights/google-custom-tpus-ai-compute

appears to be that link.

> Google's AI summaries

i really doubt the AI summaries use Gemini or even Gemma

Worldwide AI-chatbot web visits: ChatGPT 76% → 54%. Gemini 6% → 28%. Claude 1% → 9%. This is not monthly users and not revenue.

I understand Claude being niche (until they started running Superbowl ads?), but why'd it take Google this long to start catching up?

Because Google's AIs are absolutely awful?

they are good enough for 89.56% of ChatGPT users

Really? This just tells me you haven't tried it in a long time. I'd argue it's a much better experience for the average user.

It lies constantly and is only useful for shallow information. At least the non-paying version.

Because they were slow rolling them out, and heavily focused on the profitable corporate customers.

I think it'd be more informative to see total users per provider, rather than % shares.

I used to use Gemini (the product) because it's free. But its responses are so incredibly bad, and the interface has so many dumb bugs, that I stopped using it and switched to OpenCode Go models + Web UI. Gemini models can recite deep knowledge, but they can't do constructive things, can't do extended web research, can't provide more than a couple options for an answer. Either their system prompt is nerfed like crazy or the models just suck at agentic work. Even ChatGPT from 2025 was better than Gemini today.

Gemini's Deep Research is pretty good. In contrast, strangely enough, ChatGPT nerfed its Deep Research offering, although a custom GPT or Work mode can trivially make up for it. I admit that Google's offerings generally have too many serious bugs to be usable.

This source-ier chart is much better: https://cdn.prod.website-files.com/6213ddd7bd3eb8fb93bf1da4/...

Stacked charts might be 2nd only to pie charts in badness.

54% to go, good work team.

Meme: "I'm doing my part" Starship Troopers.

Service guarantees citizenship!

Why? Who is the replacement? Some state backed Chinese thing that is running loss leaders for a rug pull in the near future like Deepseek just did?

What's the methodology? Is Gemini actual gemini.google.com or mixed with google.com and Android searches?

Given this anyone still thinking Google lost needs to update their belief system

Lost at what, though? Waymo is gonna pan out but they sold their domain business to squarespace despite having GCP, and Google workspace deletes data when changing billing which is absolutely unacceptable.

The domain business is a distraction. It was wise to sell it. They still own several gTLDs so they have plenty of revenue from this avenue.

Selling .coms and Donut's gTLDs? Peanuts.

Godaddy makes all their money from other stuff, and Google still does lots of other stuff. Google Workspace still prints money.

Distraction? It's about vertical integration and being a one-stop-shop for people that want to have a website/SaaS business.

Lost at “AI” or some vague statement

Google has lost $302.8B while only making $25B on AI.

That's not only losing, but record-setting losing.

How much money was “lost” on the three gorges dam or any major infrastructure project?

Google has a completely different payoff timeline than you seem to

They are bigger than most governments and invest in decade horizons

Alternate headline:

“After three years of unprecedented investment including governments, Facebook, and Google dumping money into the promise of AI, ChatGPT still has over 50% of the traffic.”

Also, grok isn’t been on that list, further proving that no serious people use it.

Grok is there, just a little bit too small to label. Dark brown on the chart.

4.6 is a good model, by the way. Legitimately frontier-level and a good communicator.

Whose worldview is it communicating, though?

Didn’t I just see on HN the other day that it has “be humorous and irreverent” baked into its system prompt? Something like that?

Reply on news.ycombinator.com