This post is meant to sell AI results monitoring software but I found it interesting. LLMs are less likely to provide predictable or repetitive results, meaning citations and authority of responses change wildly even if you look at them MoM.
Replies
What I am curious about in the intro is this statement: "When you ask a question, these models predict what comes next while throwing in controlled randomness to keep responses from becoming repetitive or predictable." ARE the major search LLMs disincentivized from providing identical responses?
sep 2, 2025, 3:13 pm • 7 1 • view
Is this just the line between LLM and RAG?
sep 2, 2025, 3:17 pm • 3 0 • view
Like from a simple "these are probability based word machines" standpoint, an LLM is still pretty likely to give similar responses, and responses that are the same *in substance* in response to a question. But if the LLM is trained to inject randomness that fluctuates MoM...
sep 2, 2025, 3:19 pm • 4 0 • view
... that seems to me like it's going to lead searchers to the absolute rock bottom of the internet by mid-2026. This thread is half SEO primal scream, half "how do we live in a society with no shared reality" primal scream
sep 2, 2025, 3:20 pm • 11 1 • view
As an SEO if I am told that LLM search is going to be disincentivized from showing my domain repetitively to searchers.... then I simply don't care. Subject me to the winds of fate at that point. I'm sticking to what I can do that will bring reliable, measurable volumes of traffic to my pages.
sep 2, 2025, 3:22 pm • 7 0 • view
Why would I bust my ass trying to get perplexity senpai to notice me in April of 2026 when I could be working on long tail pages that already bring in 4x the traffic?
sep 2, 2025, 3:23 pm • 5 0 • view
I *think* the function they are talking about in this post is something like "Temperature" as described here. A commenter explains it in this thread as "If the temperature is set appropriately, this results in more varied writing without going off the rails"
sep 2, 2025, 3:27 pm • 2 0 • view
The problem with this is, of course, when we are talking about a function where people are trying to understand a condition within reality like "Are vaccines safe", a topic with a lot of quantity of writing about "no they aren't" with zero medical authority but topic domination on sheer volume.
sep 2, 2025, 3:29 pm • 6 0 • view
But if we're incentivizing LLMs to be quirky girls, like, this is so unhelpful from a basic utility standpoint. It's very funny how often posts about LLMs include statements like this that make the technology fundamentally unworkable, and just breeze past that lol
sep 2, 2025, 3:30 pm • 5 1 • view
Yeah, it relates back to simulated annealing, which has been a technique in AI and ML models for decades; the idea is that if an optimizing model uses strict and deterministic descent to come to a result, it can get stuck in local minima.
sep 2, 2025, 3:43 pm • 1 0 • view
But it's an art, and specific to both the model and problem. Not enough heat, and you get a bad answer; too much and you get nonsense. I don't know that you can pick a single value for all LLM queries - if for no other reason than you have no idea what the shape of the surface is.
sep 2, 2025, 3:43 pm • 2 0 • view
And I think this goes back to using an LLM for a sum of general knowledge search engine. With those volumes and the training you're able to do on it, an incentive or temp function that results in 70-90% domain drift over 6 months is... wild
sep 2, 2025, 3:47 pm • 1 0 • view
That sounds absolutely dire from a reproducibility standpoint
sep 2, 2025, 3:14 pm • 0 0 • view
it is :) and everybody wants me to measure it :) :)
sep 2, 2025, 3:17 pm • 1 0 • view