O
(Not source code, not reverse engineering, not a “hack” – but a defensible, mathematically approached explanation of how ranking and the operation of LLMs can be put in the same framework. Google algorithm deciphering examples)
There is a moment when you look at a SERP (Google search results) and feel that something is not working according to “tricks”, but a multi-step, predictive machine is working – just like large language models (LLMs). I wouldn’t describe it as “cracked the Google code”, but as follows: I put together a mathematical narrative that explains both the modern search engine and why an LLM is able to predict “always the next word”.
In this article, I’ll take you through a complete pipeline – with examples, simple formulas, and a “map” that can be told in a blog without slipping into fictional technical statements.
The big common denominator: representation → transformation → probability → selection
The search engine and the LLM can be drawn onto the same meta-principle:
- Embedding: the raw input (query, document, context) is represented in a large-dimensional vector space.
- Transformation: several layers “transform” this representation (with learned weights).
- Probability assessment: score/distribution is generated.
- Selection: the system “decides” (ranking or next token).
The mention of “Hilbert space” is correct if you use it not as a literal implementation, but as a formal abstraction : a vector space with an internal product, where similarity is measured.
2) The "Fourier transform" part is exactly: spectral viewpoint, not magic
The Fourier analogy is worth making “certain“ like this:
- We are not saying that “Google runs a Fourier transform in ranking.”
- We claim that feature space can be interpreted as bases, and certain phenomena (patterns, periodicities, “components”) can be described with a spectral eye.
In a blog, I would put it like this:
“Ranking marks are multidimensional components. If you want, you can look at them ‘spectrally’ as well: it’s not a single ‘SEO trick’ that counts, but a signal composition that has dominant and weak components.”
This approach is perfectly compatible with modern ML: models often work in learned bases , and the “good hit” is where the relevant components are strong.
3) "Schrödinger collapse": how can it be used intelligently and unassailably?
The quantum part is good if it is used as a pure metaphor :
- The system carries one distribution (uncertainty) based on many signals: there are several possible outcomes.
- In the end, however, a concrete decision is made: a ranking, a top-1, a top-k.
This is correct:
“The ‘collapse of the wave function’ is only figurative speech here: from the many possible interpretations and signs, the system chooses a single, ordered list.”
4) Markov logic and graph: why doesn't the link disappear, but only transforms?
The Markov character (and the classic PageRank spirit) is good in the narrative because ranking = graph + probabilities.
Links, mentions, and brand-strengthening relationships can all be thought of as a net:
- vertices: pages/entities
- edges: links / references / mentions
- “temporary probability”: where authority/trust “flows”
In the modern world, this is supplemented:
- Not only link = live, but also entity mention, source quality, authorial credibility, consistency .
6) Example 1: "eyelid surgery Budapest" – how does the machine "think"?
Suppose the user enters:
q = “eyelid surgery Budapest prices healing”
The system “mixes” several signal packets:
(A) Intent match
- Clinic page? Medical information? Blog? Price list?
If the query is “prices” and “recovery”, then the good result is: - provides detailed price limits,
- describes the healing process,
- not just a marketing text.
(B) E-E-A-T authenticity
- Is there an author (doctor/professional profile)?
- Are there references or professional sources?
- Is there transparent information (risks, contraindications, control, complaint handling)?
(C) Localization
- Signs tied to Budapest: address, map, local references, GMB/Google Business Profile coherence, local reviews.
(D) Usability
- Fast site, good mobile experience, easy appointment booking, understandable structure.
It’s not the specific weights that matter, but the message: the sum of components.
The “Fourier-like” narrative here:
If the Trust component is weak (no author, no medical framework, no transparency), then the keyword match is in vain: a dominant component is missing in the spectrum.
7) Example 2: "radiator valve" – e-commerce query, other components dominate
Query:
q = “radiator valve 1/2 inch thermostatic”
Here, the ranking favors other signs:
- Accuracy of product data (compatibility, dimensions, stock)
- structured data (Product, Offer, AggregateRating)
- Category→Product internal link system
- Filters, facets management (duplication control)
- Price, Shipping, Returns, Trust Signs
8) "Miklós Róth cracked the Google code" – the blog narrative that can be said acceptably
This is how I would describe it as strong, yet not dangerous or false:
“I didn’t hack Google’s secret algorithm. I ‘cracked’ what mathematical pattern should be used to think about it: ranking is not a set of tricks, but transformed representations + probabilistic decision. The same logic drives LLMs: the next token is selected from a distribution, based on the representation of the context.
This is a hit, and it is true.
9) What should you do with this in practice? (White hat, predictive SEO/AEO checklist) ✅
If the system is the sum of components, then you don’t need “an SEO hack“, but component reinforcement:
- Embedding from the side: clear entities, conceptual space, internal link logic, well-structured content.
- From the trust side: authors, references, transparent information, real evidence, consistent branding.
- From the local side: NAP consistency, local content, opinions, map presence, local entity relationships.
- From the UX/Tech side: speed, mobile, indexability, duplication, structured data.
And the most important “LLM parallel” idea:
“If you want the search engine and AI assistants to ‘predict’ that you are the best answer, you need to give signals that will help the model come to the same conclusion steadily.”
10) Closing twist: the search engine ranks, the LLM speaks – but both are predictive
The search engine “predicts” what the best result will be.
The LLM “predicts” what will be the best next token.
Azonnali segítség mindenben, ami AI
Kérdésed van az AI SEO-ról, az AI-láthatóságról vagy az AI marketingről? Ne írj űrlapot – hívj fel, és élőben átbeszéljük, mit lát ma belőled a Google és a ChatGPT, és mi lenne a leggyorsabb fogás a piacodon.
📞 +36 70 629 0690 Azonnali segítségkérés – minden, ami AIHétköznap munkaidőben azonnal veszem · nincs kötelezettség, nincs sablonduma
