Key takeaways
RankNet is a pairwise learning-to-rank neural network introduced by Microsoft Research in the 2005 ICML paper "Learning to Rank using Gradient Descent." It learns to order search results by training on pairs of documents using a probabilistic cross-entropy cost and gradient descent. RankNet evolved into LambdaRank and LambdaMART, and Microsoft says its ranking algorithm is currently used by Bing for web search.
- RankNet was introduced by Chris Burges and colleagues at Microsoft Research in their 2005 ICML paper, and it later won ICML's Test of Time award.
- It is a pairwise learning-to-rank model: it learns from pairs of documents (which should rank higher) using a probabilistic cross-entropy cost and a neural network trained by gradient descent.
- RankNet's lineage runs RankNet to LambdaRank to LambdaMART, the boosted-tree version that won Track 1 of the 2010 Yahoo Learning to Rank Challenge.
- Microsoft Research states RankNet's ranking algorithm is currently used by Bing for web search, with continuous retraining on fresh labeled data.
- To rank in Bing, optimize for its published factors (relevance, quality and credibility, user engagement, freshness, location, page load time) rather than for RankNet's internal math.
What RankNet is
Definition
RankNet is a machine-learned ranking algorithm developed at Microsoft Research and introduced in the 2005 paper Learning to Rank using Gradient Descent, presented at the 22nd International Conference on Machine Learning (ICML) in Bonn, Germany. The authors were Chris Burges, Tal Shaked, Erin Renshaw, Ari Lazier, Matt Deeds, Nicole Hamilton, and Greg Hullender.
It belongs to a family of techniques called learning to rank: instead of a human hand-tuning a formula, the system learns a ranking function from labeled training data. RankNet takes a feature vector describing a query-document pair and maps it to a single relevance score, then orders results by that score. According to Microsoft Research, RankNet was one of the earliest neural-network ranking models to be used in a commercial search engine, and the original paper later won ICML's Test of Time award, given to the paper from ten years earlier judged to have had the highest academic impact. RankNet is one of several systems we cover in our pillar overview of Bing's ranking algorithms.
RankNet at a glance
- Developed by
- Microsoft Research
- Lead author
- Chris Burges and colleagues
- Introduced
- ICML 2005, Bonn, Germany
- Approach
- Pairwise learning to rank
- Cost function
- Probabilistic cross-entropy
- Training
- Neural net via gradient descent
- Lineage
- RankNet to LambdaRank to LambdaMART
- Status today
- Used by Bing for web search
How RankNet works (the pairwise approach)
RankNet's central idea is to learn from pairs of documents rather than trying to predict an absolute score for each one. During training, query-document pairs are labeled for relevance by human judges (for example, "excellent match" or "good match"). For any two documents returned for the same query, RankNet learns which one should rank higher.
The mechanics, as described in the original paper, are:
- Feature vectors in, score out. Each document is represented as a list of numeric features. A neural network maps that vector to a single real-valued relevance score.
- A probabilistic cost function. RankNet models the probability that document A should rank above document B using a sigmoid of the score difference, then applies a cross-entropy cost that penalizes the gap between the model's predicted ordering probability and the known correct one.
- Gradient descent training. The paper extends backpropagation to handle ordered pairs; the blog retrospective describes training the deployed model with Stochastic Gradient Descent (SGD), nudging the two scores in the right direction.
Because the loss depends only on the difference between two scores, RankNet avoids the harder problem of mapping to fixed rank values or rank boundaries. The original paper reported results on toy data and on data from a commercial internet search engine consisting of 17,004 queries.
From RankNet to LambdaRank to LambdaMART
RankNet was the first step in a well-documented lineage, laid out in Chris Burges's 2010 Microsoft Research technical report From RankNet to LambdaRank to LambdaMART: An Overview.
- RankNet (2005): pairwise neural network trained with a cross-entropy cost and gradient descent.
- LambdaRank: its key observation is that to train the model you do not need the costs themselves, only the gradients with respect to the scores. Those gradients (the "lambdas") can be scaled by the change in a ranking metric such as NDCG when two documents are swapped, letting the model directly optimize a non-smooth IR metric.
- LambdaMART: the boosted-tree version of LambdaRank, combining the lambda gradients with MART (gradient-boosted decision trees). An ensemble of LambdaMART rankers won Track 1 of the 2010 Yahoo Learning to Rank Challenge.
This progression matters because it shows RankNet is not a frozen 2005 artifact. The same pairwise, gradient-based foundation underpins the tree-based models that became standard across the search industry. For the vector-search infrastructure that retrieves candidates before this ranking core scores them, see our guide to Bing SPTAG.
RankNet's role in Bing today
Microsoft-confirmed: the Microsoft Research retrospective states plainly that RankNet's ranking algorithm is currently used by Bing for web search. The retrospective also notes that, compared with the earlier ranking system known as The Flying Dutchman, RankNet could produce a ranking model in a couple of hours on a single machine rather than days on a cluster, which made frequent retraining practical.
Practitioner inference (not official): SEOs commonly describe Bing's ranking core as continuously retrained on fresh labeled data. While continuous retraining is a reasonable inference from how learning-to-rank systems operate and from the speed advantage Microsoft describes, Microsoft has not published the exact retraining cadence or the current model architecture. Treat any specific claim about how often Bing retrains, or about RankNet being the sole ranker, as inference rather than confirmed fact. Modern Bing ranking almost certainly blends many models and signals on top of this foundation, including the large language model work behind Bing Prometheus.
History of RankNet: a timeline
RankNet went from a faster machine-learned replacement for Microsoft's earlier ranking system to an award-winning paper whose successors still power commercial search.
-
2004
Development begins
Microsoft Research and Microsoft's web search team begin developing RankNet as a faster, machine-learned alternative to the earlier ranking system known as The Flying Dutchman.
-
2005
RankNet paper at ICML
Chris Burges and colleagues present "Learning to Rank using Gradient Descent" at the 22nd International Conference on Machine Learning in Bonn, Germany, introducing RankNet.
-
2006-2010
LambdaRank and LambdaMART
The pairwise approach evolves into LambdaRank (optimizing gradients scaled by metric change) and LambdaMART (the boosted-tree version).
-
2010
Yahoo Learning to Rank Challenge
An ensemble of LambdaMART rankers wins Track 1 of the 2010 Yahoo Learning to Rank Challenge, validating the lineage on a public benchmark.
-
2015
ICML Test of Time award
The 2005 RankNet paper wins ICML's Test of Time award for the highest academic impact among papers presented ten years earlier.
What RankNet means for SEO and Bing optimization
You cannot optimize for RankNet directly. It is an internal algorithm trained on human relevance labels and behavioral data you do not control. What you can do is align your pages with the signals Bing officially says it uses, which are the features such a model is trained to reward.
Per Bing's published webmaster guidance, those factors are relevance, quality and credibility, user engagement, freshness, location, and page load time. Notably, Bing is more openly explicit than Google about using engagement: its guidelines describe looking at whether users clicked a result and whether they stayed or quickly returned to Bing. The practical takeaway is that satisfying searcher intent and earning sustained engagement is the most durable way to align with a learning-to-rank system. For a deeper look at how those behavioral signals are measured, see our guides to Bing's Clarity-derived UX signals and Bing's social signals.
Bing's published ranking factors
Because RankNet is trained on labeled data, the most reliable way to influence it is to align with the factors Bing publicly names. These six factors come from Bing's own webmaster guidance.
| Factor | What it covers |
|---|---|
| Relevance | How closely the content on the landing page matches the intent behind the search query. Bing-confirmed published factor. |
| Quality and credibility | An evaluation of the page itself, including author or site reputation and content completeness; Bing notes content that cites data sources is considered higher quality. Bing-confirmed. |
| User engagement | Bing explicitly looks at whether users clicked a result and whether they stayed on it or quickly returned to Bing. Bing-confirmed, and stated more openly than Google states it. |
| Freshness | Bing generally prefers content that consistently provides up-to-date information, weighted more heavily for time-sensitive topics. Bing-confirmed. |
| Location | Where the user is located (country and city), where the page is hosted, and the language of the document. Bing-confirmed. |
| Page load time | Slow load can cause visitors to leave, which Bing may treat as a poor user experience and an unsatisfactory result. Bing-confirmed. |
The practical takeaway is that a fast, credible, intent-matching page that earns sustained engagement aligns with exactly the kinds of features a learning-to-rank model like RankNet is trained to reward.
How to optimize for Bing's RankNet
You cannot target RankNet directly, so optimize for the public signals it is trained to reward: tight intent matching, demonstrable quality, sustained engagement, freshness, speed, and clean indexation.
-
Match content tightly to query intent for the page's target queries.
Relevance is Bing's first published factor and is exactly the kind of feature a learning-to-rank model is trained to reward.
-
Strengthen demonstrable quality and credibility: clear authorship, citations to primary data sources, and complete coverage of the topic.
Bing states quality and credibility evaluate the page and the author or site reputation, and explicitly favors content that cites its data sources.
-
Earn sustained engagement by satisfying the click: fast-loading, well-structured pages that answer the query so users do not bounce back to Bing.
Bing openly says it observes whether users stay on a result or quickly return, making dwell and pogo-sticking a direct, confirmed signal.
-
Keep time-sensitive pages updated and date-stamped.
Bing prefers fresh content, especially for topics where information changes over time.
-
Reduce page load time across mobile and desktop.
Bing names slow load as a poor experience that can produce an unsatisfactory search result.
-
Get the page crawled and indexed in Bing via Bing Webmaster Tools and IndexNow.
No ranking model can score a page it has not ingested; indexation is the prerequisite, and Bing also powers ChatGPT search and Copilot.
RankNet myths vs. reality
RankNet attracts a lot of confident-but-wrong claims. Here are the most common myths and what is actually true.
Myth You can optimize your pages specifically for RankNet.
Reality RankNet is an internal model trained on human relevance labels and behavioral data. You influence it only indirectly by improving the public signals Bing names: relevance, quality, engagement, freshness, location, and speed.
Myth RankNet is the single algorithm that ranks Bing results.
Reality Microsoft confirms RankNet's algorithm is used by Bing, but modern web ranking blends many models and signals. The RankNet-to-LambdaMART lineage is the documented foundation, not the entire live system.
Myth Microsoft Clarity data feeds Bing's ranking algorithm.
Reality This is unconfirmed industry speculation. Microsoft documents Clarity only as a free behavioral analytics tool (heatmaps, session recordings, funnels). There is no Microsoft statement that Clarity data is used as a ranking input.
Myth Social signals are a confirmed, heavily weighted Bing ranking factor.
Reality Bing has referenced social context at times, but the magnitude and direct weighting of social signals in ranking is practitioner inference, not a published, quantified Microsoft ranking factor.
Myth RankNet is outdated 2005 research with no current relevance.
Reality The 2005 paper won ICML's Test of Time award, its successors won the 2010 Yahoo challenge, and Microsoft says the algorithm is currently used by Bing for web search.
Frequently asked questions
RankNet is a machine-learned ranking algorithm from Microsoft Research, introduced in the 2005 ICML paper Learning to Rank using Gradient Descent. It uses a neural network to score how relevant a document is to a query, learning from pairs of documents. Microsoft says its ranking algorithm is currently used by Bing for web search.
RankNet was created at Microsoft Research by Chris Burges, Tal Shaked, Erin Renshaw, Ari Lazier, Matt Deeds, Nicole Hamilton, and Greg Hullender. Their paper Learning to Rank using Gradient Descent was presented at the 22nd International Conference on Machine Learning (ICML) in 2005, and it later won ICML's Test of Time award.
RankNet trains on pairs of documents for the same query, learning which should rank higher. It models the probability that one outranks the other using a sigmoid of their score difference, applies a cross-entropy cost, and adjusts a neural network with gradient descent. This avoids predicting absolute scores or fixed rank boundaries.
RankNet is the original pairwise neural network. LambdaRank observed you only need the gradients (lambdas), scaled by the change in a metric like NDCG, not the costs themselves. LambdaMART is the boosted-tree version combining those gradients with gradient-boosted decision trees, and it won the 2010 Yahoo Learning to Rank Challenge.
Microsoft Research states that RankNet's ranking algorithm is currently used by Bing for web search. It is understood to be continuously retrained, though Microsoft has not published an exact cadence. Modern Bing ranking very likely layers additional models and signals on top of this documented foundation rather than relying on RankNet alone.
You cannot target RankNet directly. Instead optimize for the factors Bing publishes: relevance to query intent, quality and credibility, user engagement (Bing watches whether users stay or return), freshness, location, and page load time. Strong, intent-matching, fast pages that earn sustained engagement align best with any learning-to-rank model.
There is no Microsoft statement confirming this. Microsoft documents Clarity as a free behavioral analytics tool offering heatmaps, session recordings, and funnel tracking to improve user experience. Any claim that Clarity behavioral data is used as a Bing ranking input is industry speculation, not confirmed fact.
Yes, and Bing is unusually explicit about it. Its published guidance describes evaluating whether users clicked a result and whether they spent time on it or quickly returned to Bing. This makes dwell time and pogo-sticking a directly stated engagement signal, in contrast to Google's more guarded public position.
The bottom line
Bottom line
RankNet turned ranking into a learned problem rather than a hand-tuned formula, and that pairwise, gradient-trained foundation still underpins Bing's search ranking today. You cannot optimize for the model itself, but you can optimize for the public signals it rewards: relevance, credibility, sustained engagement, freshness, location, and speed. Build fast, credible, intent-matching pages, get them indexed in Bing, and you align with the system rather than chasing its internal math.
References
- RankNet: A ranking retrospective - Microsoft Research
- Learning to Rank using Gradient Descent (Burges et al., ICML 2005)
- From RankNet to LambdaRank to LambdaMART: An Overview (Burges, MSR-TR-2010-82)
- Learning to rank using gradient descent - ACM Digital Library
- Clarity Overview - Microsoft Learn
- Bing's search ranking factors - Search Engine Land
- Bing Ranking Factors Revealed in Update to Webmaster Guidelines - Search Engine Journal