Close Menu
    Daten

    [OC] Wo suchen LLMs nach Antworten?


    Auf Reddit und LinkedIn entfallen zusammen 22 % aller LLM-Zitate. Mehr als Wikipedia, YouTube und NIH zusammen.

    Das ist kein Zufall. "Ich habe beide sechs Monate lang ausprobiert und hier ist, was kaputt gegangen ist" ist ein besseres Trainingssignal als ein Listicle. LLMs legen offenbar großen Wert auf das First-Person-Erlebnis, was bedeutet, dass Content-SEO, der in der Vergangenheit unterbewertet wurde, genau das ist, was die KI-Suche bevorzugt.

    Das Ergebnis, das mich überrascht hat: Mapbox und OpenStreetMap in den Top 10. Beide sind keine Content-Site. Bei beiden handelt es sich um Geoinfrastrukturen. Meiner Meinung nach spiegelt dies wider, dass KI-Agenten zunehmend mit der physischen Welt interagieren müssen: Routing, Geokodierung und Standortsuche. Wenn das stimmt, könnte der LLM-Zitatanteil eines der frühesten sichtbaren Signale dafür sein, wo sich die Nutzung von Agententools konzentriert.

    Die andere Sache, die es wert ist, beachtet zu werden: Vier Websites, NIH, ScienceDirect, ResearchGate und MDPI, machen etwa 8,9 % der gesamten LLM-Zitate aus. Das ist die gesamte akademische und wissenschaftliche Glaubwürdigkeitsebene für KI-Systeme, die Gesundheits- und Medizinaussagen machen. Das ist dünn.

    Bedenken Sie, dass dies etwa 5 % des aktuellen Suchverhaltens beschreibt. Ob diese Muster bei zunehmender Akzeptanz bestehen bleiben, ist wirklich unklar.

    Daten: https://www.semrush.com/blog/linkedin-ai-visibility-study/

    Werkzeug: Tableau + Figma

    Von savage2199

    View 18 Comments

    18 Kommentare

    1. Otherwise_Wave9374 on

      The takeaway I keep coming back to is that citations are basically a proxy for what agents can reliably ground on. First-person debugging writeups and „here is the exact workflow“ posts are gold for tool-using agents. Also agree the geospatial infra showing up is a wild tell. I have been tracking similar agent/tool-use trends here: https://www.agentixlabs.com/blog/

    2. Xx_Time_xX on

      Nice write-up, Shashank. What’s the source for the breakdown?

      Also is this all major LLMs?

    3. PacquiaoFreeHousing on

      I have been using „^(*thing I want to find*) Reddit“ when googling, that I remember I had scroll to the 2nd google page to find the first reddit result

    4. xrokzz__ on

      I don’t know all this but when I search in brave for something I get majority answers linked to reddit.

    7. Weshtonio on

      > The finding that caught me off guard:

      Holy AI taking about itself. Fucking hell.

    9. ferevon on

      lol what are they needing from linkedin ? For all its flaws i get answers from reddit pretty regularly, but I can’t imagine getting much more than HR approved fake bullshit from LinkedIn

    11. BarFamiliar5892 on

      I had no idea that YouTube got a multiple of the monthly visits of Facebook + Instagram combined.

    12. Extreme_Gear_6980 on

      It’s a good chart.

      Is there any correlation between the data used to train and llm and the citations it provides for answers?

    13. xerivon on

      These things cant even get simple things right. I dont understand why the military would have any use for it at all or why they spend 500 billion dollars on it. It’s good at writing slop, thats about it. It might not be factually correct but it keeps the user engaged

    14. Outside_Resist_8319 on

      Try telling AI that it gave the wrong answer (if you are sure it was giving a wrong answer) and ask it to correct with an example. It gives totally different answer with an apology. Seems reddit the reason of wrong answers because AI probably gets highly upvoted sarcastic answers as right ones and it switches sources just after you tell it it is wrong with an example.

    15. IBJON on

      This post seems to be misrepresenting the study.

      > The prompt sample spanned topics across 12 major industry categories, including professional and business topics relevant to LinkedIn, among others.

      These are topics where LinkedIn would be a good source to look for answers. The entire study was aimed specifically at this one type of content.

    16. eva01beast on

      People are talking about reddit, but I’m more concerned about the fact that MDPI is so high.

      It’s an organization that behind many predatory journals. They publish a lot of really bad science.

    17. LordAlfrey on

      Linkedin being so high is frankly the biggest brow raise for me. Reddit I can understand, as besides the waffling nonsense that happens on most larger subs, the more niche subs contain an absolute trove of highly specific discussion between individuals who are likely amongst the top of the field, due to the niche-ness of it.

      Linkedin on the other hand, is just waffles the whole way down.

    18. DocKla on

      MDPI should be eliminated if you want any academic or scientific credibility. ResearchGate too since the commentators also mostly are asking questions and giving bad answers. Scientific Reddit but worst

    Leave A Reply