When imagining how AI search systems work, it’s easy to envision an ultrasmart ‘black box’ monolith model akin to 2001: A Space Odyssey that decides which brands get cited.
In reality, AI systems layer multiple ranking signals that each analyze a different aspect of the retrieved online content.
No single signal has full control, so they’re all equally important. Moreover, they function like a stack of evaluators.
How do we know?
We’ve been able to reverse engineer Google’s Vertex AI Search documentation to learn what makes generative AI models tick.
In our research, we uncovered seven distinct levels of scoring involved with ranking online content sources.
Understanding and optimizing for these signals is fundamental for improving your brand’s visibility on AI platforms like ChatGPT, Gemini, and Google’s AI Overviews.
In this guide, we’ll take an in-depth look at all seven AI ranking signals, including the best ways to optimize for them.
What are the Seven AI Ranking Signals? An Overview
Unpacking Google’s AI Discovery Engine was eye-opening, to say the least.
It turns out that AI doesn’t replace classic search signals; it stacks on top of them. While LLMs are fantastic at interpretation, semantic matching, and handling context, they aren’t the best at precision ranking.
That’s why Google’s AI still uses a classic search ranking algorithm as a base ranking layer.
This initial pass provides a structured ‘first draft’ of results that newer, AI-powered signals then refine, re-rank, and contextualize.
Do other generative AI models use base ranking algorithms, too?
It turns out that they do.
Whether you’re dealing with ChatGPT, Perplexity, or any generative AI model, all of them rely on a base ranking layer.
Also, their ‘ranking stacks’ closely mirror Google’s, so this list applies to any GEO strategy.
Here’s an overview of the seven-step signal stage from the four-stage AI search pipeline.
#1: Base ranking (core algorithm output) 
It’s crucial to note that while the base ranking layer uses a search algorithm, it does not use traditional PageRank, link-graph metrics, or measure raw backlink quantity.
Instead, base ranking is entirely focused on:
- Clarity
- Semantics
- Structure
The algorithm weighs things like topical expertise, relevance, scannable headings, and classic on-page SEO signals to provide an initial set of results that match the user query.
#2: Gecko score (semantic matching)
Gecko is the first true AI-powered signal that enters the stack. It’s where the system shifts from old-school relevance matching to semantic understanding.
It uses an embedding-based vector similarity score to determine if the query and the content actually mean the same thing.
To learn how Gecko works, we have to rewind and look at two previous stages in the four-stage AI search pipeline.
Here’s a quick refresher of each stage:
- Prepare
- Retrieve
- Signal
- Serve
Gecko occurs in the signal stage, where the system evaluates content relevancy, recency, and trust.
Before the system retrieves any content (in the ‘prepare’ stage), it normalizes the query into standard language and converts it into vector embeddings, which are numerical representations of word meanings.
During the retrieval stage of the AI search pipeline, the system pulls in content that was already chunked into 300–500 token segments during the indexing process (which occurs before the query is even asked).
Each of these chunks comes with a pre-computed embedding, because embedding generation also happens during crawling and indexing.
The Gecko layer consumes these embeddings and compares how close they are in vector space.
A condensed example of the Gecko score
Technical lingo aside, here’s an extremely simplified example of embedding similarity in action.
Note: We’re using a microscopic embedding space that only consists of 3 dimensions. Most embeddings have anywhere from 500 to 2,000+ dimensions:
Text 1: “I love basketball.”
Vector: [0.8, 0.1, 0.7]
Text 2: “Basketball is my favorite sport.”
Vector: [0.79, 0.12, 0.68]
Text 3: “I enjoy cooking pasta.”
Vector: [0.2, 0.9, 0.1]
When measuring the cosine similarity, texts 1 and 2 share a high similarity (~0.998), while texts 1 and 3 aren’t very similar at all (~0.22).
This is one of the ways AI systems are able to understand meaning instead of just matching words.
#3: Jetstream (cross-attention) 
Jetstream provides another layer of nuance when interpreting meaning. This is because embeddings struggle with nuance, especially negation and comparisons.
When Jetstream enters the picture, it reads the retrieved chunks with cross-attention to ensure that they correctly match the context of the query.
In a nutshell, embeddings can flatten meaning, but Jetstream reintroduces structure to interpret fine-grained context like:
- X is NOT the cause of Y (negation)
- The difference between X and Y is (comparison)
#4: BM25 (keyword matching) 
This layer likely comes as a shock to marketers who thought lexical keyword matching played no role in AI search.
Why do AI systems still use BM25?
AI can misread content at times, and keywords create anchors that keep things grounded. This means it’s still important to use the exact terminology your audience uses.
If you don’t, BM25 will fail to fire, and you’ll lose a key vote.
#5: PCTR (engagement) 
PCTR stands for predictive click-through rate, and it answers the question, “Will users actually want to click on this?”
It gives a boost to content that users are more likely to engage with, and the formula uses a combination of:
- Past engagement signals
- Layout
- Metadata
- Snippet quality
There are three tiers involved with PCTR: cold, warm, and hot, with hot pages receiving consistent clicks (and thus, getting significant boosts).
#6: Freshness (recency) 
AI systems have a very strong recency bias, which reinforces the need to frequently update your content.
Even evergreen pieces are at risk of becoming stale due to things like outdated terminology and shifts in industry standards.
This layer detects whether a query has temporal intent or not. For instance, if you ask ChatGPT something like, “What’s the score for the Lakers game?” It will assume that you mean the latest Lakers game.
From there, it downranks stale content and upvotes fresh pages.
#7: Boost/bury (reputation, reviews, etc.)
This layer isn’t actually part of the semantic model; it’s an external system that includes enterprise-level business rules and safety overrides.
Basically, it’s the layer that determines, “Should this source be elevated or suppressed?”
It evaluates signals like brand reputation and trustworthiness, which include:
- Reviews
- Customer satisfaction metrics
- Content safety history
- Online mentions and citations
Positive signals can lead to a boost (a sign that you’re a trusted, safe brand) or a bury (for shady, low-quality brands).
On the business side, this layer informs the model:
- To boost business partners
- Unverified sources to demote
- To suppress any results flagged for legal or safety reasons
To flow through this part of the pipeline, you need to have a positive brand sentiment, a strong reputation, and an authoritative presence across the web.
What are the Signals Most Sites Struggle With?
Next, let’s look at the signals that most sites struggle with right now.
We’ve discovered that there are many websites that AI misunderstands.
It’s not that they lack outstanding content; it’s that it’s not formatted and optimized for AI search systems.
The signals most sites struggle with include:
- Embedding similarity (Gecko) – If your content isn’t separated into hyperfocused chunks, missing context may occur during the Gecko scoring layer. The best way to achieve this is to use clear headings (H1, H2, H3) that explicitly describe the content they contain. Do not deviate and include unrelated concepts in the same chunk.
- Semantic relevance (Jetstream) – While Jetstream excels at catching nuance missed by Gecko, it still requires clear entity relationships to function properly. Make sure important concepts like your products, services, locations, hours, and key terms are clearly defined, marked up with schema, and consistently mentioned.
- Freshness – Remember, your content needs frequent updates via timestamps and schema markup. Double-check your evergreen pieces and ensure they get periodically updated as well.
Why Keywords Alone Cannot Move the Needle
Remember, only one of the seven AI ranking signals has to do with lexical keyword matching.
While this means you should still conduct keyword research, it doesn’t mean that you can stick to a keyword-focused SEO strategy.
Semantic alignment is more important than keyword use, especially in terms of semantic clarity, topical structure, and vectorized meaning.
AI engines judge content primarily through embeddings and cross-attention. If your content fails to appeal to these layers, you’ll likely wind up with poorly ranked content.
How Next Net Strengthens These Core Signals
We’ve already developed a long list of optimization tactics that appeal to AI’s ranking stack.
Some of these techniques include:
- Improving embedding similarity with proper chunking.
- Enhancing semantic relevance by clearly distinguishing the relationships between concepts.
- Structuring content with semantic HTML and schema markup for cleaner parsing and chunk interpretation.
- Updating key signals to maintain freshness.
- Detecting and correcting topic drift.
Thanks to our research and experimentation, we’ve unlocked the secrets for improving your visibility across all generative AI platforms, not just Google.
Final Thoughts: The Seven AI Ranking Signals
To summarize, if you want better AI visibility, then you need to optimize for the seven AI ranking signals mentioned in this article.
Better embeddings result in stronger ranking positions.
Clearer meaning increases AI citation likelihood.
Stable freshness and trust signals lead to more consistent retrieval.
These are the main reasons why GEO is worth the effort.
Do you want to start winning with AI search?
Book a call with our team to lift your performance across the semantic layer.
