Sign in
← back

Google’s Ranking Signals

Reverse Engineering

Overview of search ranking factors including popularity signals, PCTR models, semantic relevance, keyword matching, freshness, and various search modes.

Listen

How does a modern search engine decide which results you see first? It all starts with a base ranking, which is the initial relevance score of a document. From there, several sophisticated adjustments shape the final list. Keyword matching looks at how often your search terms appear, while embedding adjustments measure semantic similarity. For even deeper context, a semantic relevance model helps the system understand nuance and negation.

The system also relies heavily on user behavior. Popularity signals boost documents that get the most interaction. Predicted click-through rate models estimate the likelihood of a user clicking a result, and these can even be personalized using an individual's history once the search system has served at least one hundred thousand queries. Other ranking factors include document freshness, predicted conversions, and manual adjustments to promote or demote specific content.

When it comes to presenting these results, there are three main search modes. You can get a standard list of results, a generative summary written above the list, or a conversational search that supports follow-up questions. For generative AI searches, only the top five results are analyzed. These are shown either as short text snippets or longer, extractive passages. Finally, the system can block adversarial queries, preventing the underlying large language model from generating inappropriate answers.

Popularity

Popularity signals are derived from user interactions based on ingested user events. The more the users interact with a document, the stronger the boosts are. These data requirements check the overall readiness of your events to generate the popularity signals. This is regardless of the specific search app that you choose.

Predicted CTR model

PCTR models predict the chances of viewing a document under a given context based on historical user events. It is an important factor considered in ranking. Threshold and metrics values are aggregated over all linked data stores with events data.

Personalised predicted CTR model

Personalised PCTR models take user-specific signals, such as their metadata or user history, into consideration. Only takes effect if at least 100,000 queries have been served by VAIS.

  1. Position – This shows the final rank of the document in the search results.
  2. Query Name – This is the identifier or title of the returned document.
  3. Base Ranking – This is the initial relevance score of the document provided by the core ranking algorithm, before any adjustments are made.
  4. Embedding Adjustment – This score is adjusted based on the semantic similarity between the query and document embeddings. It is also known as the Gecko score.
  5. Semantic Relevance – This is a more advanced relevance score from a cross-attention model (Jetstream) that better understands context and negation compared to embeddings.
  6. Keyword Matching – This score is based on the frequency and relevance of query keywords found in the document, typically using an algorithm like BM25.
  7. Predicted Conversion – This score predicts the likelihood of a user engaging with the result (e.g., clicking), based on historical user interaction data (PCTR/PCVR).
  8. Freshness – This score is adjusted based on the recency of the document, which is especially important for time-sensitive queries.
  9. Boost/Bury – This is a manual adjustment applied to the score to either promote (boost) or demote (bury) a document based on business rules.


This is the initial relevance score of the document provided by the core ranking algorithm, before any adjustments are made.

Modes of search

  1. Search: Search with a list of results
  2. Search with an answer: A generative summary above the search results
  3. Search with follow-ups: Conversational search with generative summaries and support for follow-up questions

How many top results go into AI search?

Five.

Snippets vs Extractive Answers

Snippets – Short fragments of text from the search result content

Extractive answers – Longer passages of text from the search result content

Per Search Metrics Comparison

  1. Baseline Search Count
  2. Comparison Search Count
  3. Search Count Delta
  4. Baseline CTR
  5. Comparison CTR
  6. CTR Delta
  7. Baseline No Results Rate
  8. Comparison No Results Rate
  9. No Results Rate Delta

Ignore adversarial query – Prevents LLM answers on adversarial queries.

Dan Petrovic · Dec 24, 03:40

You did a great research, but I would add another most important Google Ranking factor

Engagement

The more user engaged with the content, the more it gives positive data to search engine about the website

And what to do for engagement

All the SEO things

Like

On page …. give unique information

Technical…. User Don’t see page loading too much or broken images

User experience

Off page.

Safi Ullah · SupportsSuggests · · Dec 24, 05:23

100% percent agree, just need to find the exact wording in Google’s docs and I’ll add it in as all of the above you’ve seen in the article comes straight from Google.

Dan Petrovic · SupportsExpands · · Dec 25, 05:06