---
title: "Bias in Matchmaking Algorithms and How to Audit It"
description: "Learn how bias enters matchmaking algorithms, how to conduct an algorithmic bias audit, and how organizations can build fairer, more transparent matching systems."
canonical: "https://meetwho.app/blog/bias-in-matchmaking-algorithms-and-how-to-audit-it"
language: "en"
published: "2026-08-07T02:57:43.954+00:00"
updated: "2026-08-11T07:19:57.252312+00:00"
reading_time_minutes: "17"
author: "Yağız Gürbüz"
author_url: "https://meetwho.app/author/yagiz-gurbuz"
source: "MeetWho — the networking layer for events and communities"
license: "Quote with attribution and a link to the canonical URL."
---

# Bias in Matchmaking Algorithms and How to Audit It

## TL;DR

- Learn how bias enters matchmaking algorithms, how to conduct an algorithmic bias audit, and how organizations can build fairer, more transparent matching systems.
- Bias in a matchmaking algorithm occurs when the system produces systematically uneven, distorted, or unfair recommendations.
- One major source of bias is historical data.
- An algorithmic bias audit should distinguish among several forms of bias because each requires a different response.
- Bias can reduce the quality of recommendations even when users never recognize it directly.

## Key questions

**What Is Bias in Matchmaking Algorithms?**

Bias in a matchmaking algorithm occurs when the system produces systematically uneven, distorted, or unfair recommendations. This may affect which people appear in search results, who receives more introductions, whose profile is ranked highly, or which users are repeatedly left outside valuable networking opportunities.

**Why AI Matching Systems Can Produce Biased Results?**

One major source of bias is historical data. A model trained on past interactions may learn patterns created by existing inequalities rather than patterns that reflect the platform’s desired future behavior.

**How Algorithmic Bias Affects Recommendation and Networking Platforms?**

Bias can reduce the quality of recommendations even when users never recognize it directly. Repetitive suggestions, narrow professional circles, and overreliance on popular profiles may prevent participants from discovering people whose expertise, goals, or resources genuinely complement their own.

**What Is an Algorithmic Bias Audit?**

An algorithmic bias audit is a structured evaluation of the data, design choices, model behavior, and real-world outcomes of an automated system. In matchmaking and recommendation platforms, the audit examines whether some users consistently receive fewer, lower-quality, or less relevant opportunities than others.

**Practical Framework for Auditing Matchmaking Algorithms**

A practical audit can be organized into four workstreams: data quality, model behavior, user experience, and governance. Each workstream should have an owner, documented evidence, identified risks, and a clear remediation process.

**How to Reduce Bias in AI Matchmaking Systems?**

Reducing bias begins with treating fairness as a product requirement rather than a final-stage test. Teams should review whether each data point improves recommendation relevance, whether historical interaction data reflects unequal visibility, and whether the system rewards popularity instead of genuine compatibility.

## Full article

Title: "Bias in Matchmaking Algorithms: Audit Guide"

 Description: "Discover how bias affects matchmaking algorithms and learn practical steps for conducting an algorithmic bias audit with transparent, fair AI practices."

# Bias in Matchmaking Algorithms and How to Audit It

 **Bias in matchmaking algorithms** can influence who is recommended, ranked, noticed, or excluded across professional networking platforms, event applications, social products, and other digital matching systems. An effective **algorithmic bias audit** helps teams identify these risks, examine how recommendations are produced, and improve the fairness, transparency, and usefulness of automated matching.

 A matchmaking algorithm does more than compare two profiles. It selects signals, assigns importance to those signals, filters potential connections, and ranks the results presented to each user. Bias can enter at any of these stages. Even when a system does not intentionally use sensitive characteristics, its data or proxy variables may still produce uneven outcomes.

 Auditing is therefore not a one-time compliance exercise. It is a structured process for understanding whom a system serves well, whom it may overlook, and whether its recommendations remain aligned with the platform’s stated purpose. In networking contexts, that purpose should not simply be to maximize interactions. It should be to support relevant, consensual, and mutually valuable connections.

## What Is Bias in Matchmaking Algorithms?

 Bias in a matchmaking algorithm occurs when the system produces systematically uneven, distorted, or unfair recommendations. This may affect which people appear in search results, who receives more introductions, whose profile is ranked highly, or which users are repeatedly left outside valuable networking opportunities.

 Not every difference in output is automatically unfair. A professional networking platform may reasonably prioritize shared goals, complementary expertise, event relevance, or mutual interests. The problem arises when differences cannot be justified by the system’s intended purpose, are driven by unreliable data, or create persistent disadvantages for particular users.

 Matchmaking systems commonly rely on information such as:

 
- Professional interests and expertise.
- Stated networking goals.
- Profile completeness.
- Previous platform activity.
- Shared communities or event participation.
- Availability and communication preferences.
- Explicitly stated requests for help or collaboration.

 Each signal represents a design decision. Teams must determine whether a signal is relevant, whether it is measured accurately, and whether it could act as a proxy for a characteristic that should not influence recommendations. A fairness review must therefore examine both the model and the assumptions behind the model.

### Why AI Matching Systems Can Produce Biased Results

 One major source of bias is historical data. A model trained on past interactions may learn patterns created by existing inequalities rather than patterns that reflect the platform’s desired future behavior. For example, if certain profiles historically received more messages because they were already more visible, an algorithm may interpret that visibility as evidence of higher relevance.

 Incomplete data can create similar problems. Users who write detailed profiles may give the system more information to work with, while users who are less familiar with the platform, write in a second language, or prefer greater privacy may appear less relevant. A ranking model can then reward data availability rather than genuine networking compatibility.

 Proxy variables also require careful examination. A system may exclude a protected or sensitive attribute while still using signals strongly correlated with it. Location, education history, employment gaps, language patterns, job titles, and platform activity can all shape results in ways that are not immediately visible.

 Feedback loops can amplify these effects. People who are recommended more often receive more profile views, connection requests, and interactions. Those outcomes then become new behavioral data, making the same people appear even more desirable to the algorithm. Without monitoring and corrective controls, early ranking advantages may become self-reinforcing.

### Common Types of Bias in AI Matchmaking

 An **algorithmic bias audit** should distinguish among several forms of bias because each requires a different response.

 Bias type Definition Matchmaking example 
 Data bias Input data does not accurately represent the intended user population Recommendations favor users with more complete historical records 
 Representation bias Certain groups or user patterns are underrepresented The system performs poorly for uncommon job roles or networking goals 
 Selection bias Collected data reflects only a limited subset of users Highly active members dominate the interaction data used for training 
 Measurement bias A variable is an unreliable substitute for the quality being assessed Message frequency is treated as evidence of connection value 
 Confirmation bias Designers or reviewers favor evidence supporting existing assumptions Positive examples are emphasized while failed matches receive less scrutiny 
 Feedback-loop bias Previous recommendations shape future training data Frequently recommended profiles continue gaining visibility 
 

 Data bias often appears before a model is trained. If the dataset excludes certain user experiences or overrepresents highly engaged members, the resulting system may optimize for a narrow definition of success. Improving the model architecture alone will not correct a flawed dataset.

 Measurement bias appears when a platform chooses the wrong success metric. A high number of connection requests may look positive, but it does not necessarily indicate useful professional relationships. A better evaluation may consider mutual acceptance, relevance feedback, follow-up activity, or whether both participants understood why the introduction was suggested.

## How Algorithmic Bias Affects Recommendation and Networking Platforms

 Bias can reduce the quality of recommendations even when users never recognize it directly. Repetitive suggestions, narrow professional circles, and overreliance on popular profiles may prevent participants from discovering people whose expertise, goals, or resources genuinely complement their own.

 This affects more than individual satisfaction. Event organizers and community leaders may see weaker engagement when networking opportunities feel generic or unbalanced. Users may lose trust when they cannot understand why certain people are repeatedly promoted, particularly when the platform provides no explanation or meaningful control over visibility.

 In professional environments, **fair AI recommendation systems** should support relevance without turning popularity into the default measure of value. A participant who is less active online may still be the most useful person for a specific collaboration, investment question, technical challenge, or mentorship need.

### The Impact of Bias on Professional Networking

 Professional networking systems operate in a sensitive context because recommendations can influence access to knowledge, partnerships, employment opportunities, funding, and community visibility. An opaque ranking process may unintentionally reinforce existing networks rather than helping people form new and valuable relationships.

 Privacy and consent are equally important. A technically accurate recommendation is not appropriate when it ignores whether users have agreed to be discoverable or contacted. Responsible matching must combine relevance with clear participation choices, controlled visibility, and transparent explanations.

 MeetWho applies this principle by showing permission-based recommendations rather than exposing an unrestricted public participant list. It analyzes participant-provided goals, interests, areas of work, requested connections, and potential ways to help others. Recommended introductions explain why two people may benefit from meeting, how they could support each other, and how a conversation might begin.

 This approach does not remove the need for auditing. It illustrates why explainability and user control should be part of the system design from the beginning. A recommendation that can be examined, understood, and questioned is easier to evaluate than a ranking produced by an entirely opaque process.

## What Is an Algorithmic Bias Audit?

 An algorithmic bias audit is a structured evaluation of the data, design choices, model behavior, and real-world outcomes of an automated system. In matchmaking and recommendation platforms, the audit examines whether some users consistently receive fewer, lower-quality, or less relevant opportunities than others.

 The purpose is not merely to confirm that a model is technically accurate. A system may perform well on an overall accuracy metric while still producing uneven outcomes for particular user groups, professions, experience levels, languages, or networking goals. A meaningful audit therefore combines quantitative testing with qualitative review, product context, user feedback, and human judgment.

 The scope should include every stage that can influence a recommendation:

 
- Data collection and profile design.
- Feature selection and weighting.
- Candidate filtering.
- Ranking and recommendation logic.
- User visibility and consent settings.
- Explanations shown with each match.
- Feedback and behavioral signals.
- Post-launch monitoring.

 A well-designed **algorithmic bias audit** also defines what fairness means for the specific product. In professional networking, fairness may involve equal access to relevant opportunities, comparable recommendation quality, appropriate exposure, and protection against popularity-based feedback loops. It does not necessarily mean that every user receives identical recommendations.

### Key Steps in an Algorithmic Bias Audit

#### 1. Define the System’s Purpose and Fairness Objectives

 The audit should begin with a clear description of what the algorithm is intended to achieve. A networking platform might aim to help participants find people with complementary expertise, shared interests, compatible goals, or mutual capacity to help.

 Teams should then translate that purpose into measurable fairness questions. For example:

 
- Do users with similar goals receive recommendations of comparable quality?
- Are less active users systematically ranked below highly active users?
- Do privacy preferences unintentionally reduce recommendation relevance?
- Are certain professional roles repeatedly underrepresented?
- Does one group experience a lower mutual connection rate?

 Without explicit objectives, auditors may collect large amounts of data without determining whether the system is fulfilling its intended function.

#### 2. Review Training, Profile, and Behavioral Data

 The next step is to identify every data source used by the system. This includes profile information, stated preferences, event participation, previous recommendations, connection requests, acceptance behavior, messaging activity, and feedback signals.

 Auditors should assess whether these datasets are complete, representative, lawful, relevant, and collected with appropriate consent. They should also identify missing values, inconsistent definitions, historical distortions, and variables that may indirectly reveal sensitive attributes.

 Data review should not assume that more information always produces fairer outcomes. Some fields may introduce noise or reflect social advantages unrelated to networking relevance. Every input should have a documented reason for inclusion.

#### 3. Test Recommendation Outputs

 Auditors should evaluate both individual recommendations and patterns across large groups of users. Testing may involve controlled profiles, historical simulations, synthetic cases, or comparisons between users with similar goals but different nonessential characteristics.

 Useful questions include:

 
- Who appears most frequently in recommendation lists?
- Who rarely appears?
- How diverse are the suggested professional backgrounds?
- Are explanations consistent with the actual ranking logic?
- Does changing an irrelevant profile field alter the results?
- Are recommendations useful to both people in a suggested match?

 Testing should cover different events, audience sizes, profile-completion levels, and networking objectives. A model that performs well at a startup conference may behave differently in a corporate workshop, online community event, or professional association meeting.

#### 4. Compare Outcomes Across Relevant Groups

 Aggregate performance can conceal meaningful disparities. Auditors should segment outcomes according to legally and ethically appropriate criteria, while respecting privacy and data-minimization requirements.

 The comparison may examine recommendation frequency, ranking position, mutual acceptance, reported relevance, connection completion, or follow-up activity. Where sensitive data cannot or should not be collected, teams can still investigate disparities across product-relevant segments such as role type, language preference, experience level, event format, or profile completeness.

 The analysis should focus on both exposure and quality. Giving every user the same number of recommendations does not guarantee fairness when some recommendations are consistently less relevant or actionable.

#### 5. Document Findings and Corrective Actions

 Every material finding should be recorded with its evidence, likely cause, user impact, severity, and proposed remedy. Documentation creates accountability and enables future teams to understand why a feature, signal, or threshold was changed.

 Corrective actions may include rebalancing data, removing unreliable features, changing ranking weights, adding human review, improving explanations, redesigning profile questions, or introducing exposure constraints. Each intervention should be tested to confirm that it improves the identified problem without creating a new one.

#### 6. Monitor the System Continuously

 A one-time audit provides only a snapshot. User behavior, event types, data quality, and product features change over time. Models may also drift as new interaction data enters the system.

 Continuous monitoring should track fairness indicators alongside conventional product metrics. Alerts can help teams identify sudden shifts in exposure, recommendation quality, acceptance rates, or user complaints. Periodic reassessments are especially important after major model updates, onboarding changes, new data sources, or expansion into new regions.

### Metrics Used for Measuring Algorithmic Fairness

 No single metric can establish that a matchmaking system is fair. Different metrics reflect different values, and some may conflict with one another. Teams should select measurements that match the system’s purpose and explain why those measurements are appropriate.

 Fairness metric What it evaluates Limitation in matchmaking 
 Demographic parity Whether groups receive outcomes at similar rates Equal exposure may not reflect different goals or consent preferences 
 Equal opportunity Whether qualified users have comparable chances of positive outcomes “Qualified” must be defined carefully 
 Disparate impact Whether a process disproportionately disadvantages a group May reveal disparity without identifying its cause 
 Error-rate comparison Whether false positives and false negatives differ across groups Requires a reliable definition of a successful match 
 Exposure parity Whether users receive comparable ranking visibility Visibility alone does not measure relevance 
 Recommendation utility Whether suggestions create value for recipients Value may be subjective and difficult to measure consistently 
 

 A strong evaluation combines several metrics with interviews, user feedback, complaint analysis, and expert review. Quantitative parity should not be achieved by lowering recommendation quality or ignoring meaningful differences in user intent.

## Practical Framework for Auditing Matchmaking Algorithms

 A practical audit can be organized into four workstreams: data quality, model behavior, user experience, and governance. Each workstream should have an owner, documented evidence, identified risks, and a clear remediation process.

 Audit stage Primary goal Example evidence 
 Data review Identify distorted or inappropriate inputs Dataset profiles, missing-value analysis, consent records 
 Model testing Evaluate ranking and recommendation behavior Test cases, simulation results, feature sensitivity reports 
 Outcome analysis Detect uneven user experiences Exposure, relevance, acceptance, and feedback metrics 
 Governance review Establish responsibility and monitoring Decision logs, review schedules, escalation procedures 
 

### Data Quality Review Checklist

 
- Document every data source used in matching.
- Confirm that each feature has a legitimate product purpose.
- Measure missing data across relevant user segments.
- Identify variables that may function as sensitive proxies.
- Review whether historical behavior reflects existing inequalities.
- Verify consent, retention, and access controls.
- Test how incomplete profiles affect ranking quality.
- Check whether highly active users dominate the dataset.

### Model Transparency and Explainability Checklist

 
- Explain the main factors behind each recommendation.
- Ensure displayed explanations reflect actual model logic.
- Allow users to control networking participation and visibility.
- Provide a process for reporting irrelevant or inappropriate matches.
- Test whether small, irrelevant profile changes alter rankings.
- Record model versions, feature changes, and audit findings.
- Review unexpected recommendation patterns with human experts.
- Monitor whether feedback loops concentrate exposure among a small group.

 Explainability is especially valuable in networking because users need context before deciding whether to connect. A clear reason—such as complementary expertise, a shared project interest, or mutual capacity to help—supports informed decisions and makes questionable recommendations easier to identify.

## How to Reduce Bias in AI Matchmaking Systems

 Reducing bias begins with treating fairness as a product requirement rather than a final-stage test. Teams should review whether each data point improves recommendation relevance, whether historical interaction data reflects unequal visibility, and whether the system rewards popularity instead of genuine compatibility.

 Improvements may include balancing datasets, removing weak proxy variables, adjusting ranking weights, testing alternative success metrics, and introducing human review for high-risk decisions. Models should also be reassessed whenever onboarding questions, matching criteria, event formats, or target audiences change.

 A responsible mitigation programme should include:

 
- **Representative data:** Evaluate whether the dataset reflects the people, professions, languages, and networking goals the platform is designed to serve.
- **Purpose-limited features:** Use only signals that have a documented and defensible relationship to match quality.
- **Human oversight:** Give qualified reviewers the ability to investigate unusual patterns and challenge model assumptions.
- **User feedback:** Let participants report irrelevant, repetitive, or inappropriate recommendations.
- **Transparent explanations:** Show users why a connection was suggested without revealing private information.
- **Continuous monitoring:** Track exposure, relevance, acceptance, and complaints after deployment.

### Designing More Responsible Networking Recommendations

 Responsible professional matching should prioritize mutual value rather than the number of introductions generated. A high-quality recommendation should reflect what both participants are working on, what they hope to find, and how they may be able to help one another. It should also respect whether each person has chosen to participate in networking.

 User control is essential. Participants should understand what information contributes to recommendations, manage their visibility, and decide whether to initiate or accept a connection. These controls support fairness because they prevent relevance scores from overriding privacy preferences or personal boundaries.

 MeetWho applies this quality-over-quantity principle through its “Know who to meet” approach. Instead of presenting an unrestricted public attendee directory, it recommends relevant people among participants who have permitted networking. Each suggestion can explain why the individuals may benefit from meeting, how they could help one another, and how the conversation might begin.

 For organizers, this creates a more intentional networking experience while preserving participant choice. MeetWho also combines these recommendations with event-page creation, registration management, approval workflows, waiting lists, announcements, reminders, QR check-in, and configurable networking privacy settings.

> **Create a free event with MeetWho and help participants find the right people for meaningful, mutually valuable conversations.**

## The Future of Fair AI Matchmaking and Transparent Recommendations

 The future of **fair AI recommendation systems** will depend on stronger governance, better explanations, and more meaningful evaluation criteria. Overall clicks or connection-request counts are unlikely to provide enough evidence that a system is helping users equitably. Teams will need to measure relevance, mutual benefit, user control, and the distribution of opportunity.

 Frameworks such as the [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) and the [OECD AI Principles](https://oecd.ai/en/ai-principles) offer useful foundations for managing AI-related risks. They should be adapted to the product’s context rather than treated as universal checklists.

 Human-centred design will remain equally important. Users need understandable recommendations, practical ways to correct inaccurate assumptions, and confidence that their information will not be exposed or sold. Systems that combine technical fairness testing with consent, privacy, and accountable product decisions will be better positioned to earn long-term trust.

 For event networking, this means moving away from opaque popularity rankings and toward context-aware introductions. The most valuable system is not necessarily the one that recommends the most people. It is the one that helps each participant understand **who to meet**, why the meeting matters, and what both people could gain from the conversation.

## Frequently Asked Questions About Bias in Matchmaking Algorithms

### What causes bias in matchmaking algorithms?

 Bias can originate in historical data, incomplete profiles, underrepresented user groups, inappropriate success metrics, proxy variables, and feedback loops. Product decisions about filtering, feature weighting, and ranking can also create uneven results.

 The cause is rarely limited to one model component. A complete review should examine data collection, user experience, business objectives, system outputs, and post-launch behavior together.

### How do you perform an algorithmic bias audit?

 Begin by defining the system’s purpose and the fairness outcomes that matter. Then document the data sources, inspect input quality, test recommendations, compare outcomes across relevant groups, and investigate any material disparities.

 The final stages are remediation and continuous monitoring. Findings should be recorded, changes should be retested, and the audit should be repeated after significant product, model, or audience changes.

### Can AI matchmaking algorithms be completely unbiased?

 No complex matchmaking system can realistically be assumed to be permanently free from bias. Fairness depends on context, and different fairness goals may conflict with one another.

 The practical objective is to identify material risks, reduce unjustified disparities, make design choices transparent, and provide appropriate human oversight. Regular auditing is more credible than claiming perfect neutrality.

### Why is explainability important in AI recommendations?

 Explainability helps users understand why a person was suggested and whether the connection fits their goals. It also gives product teams evidence they can examine when recommendations appear repetitive, inappropriate, or inconsistent.

 An explanation should accurately reflect the factors used by the system. Generic text that does not correspond to the actual ranking logic may create the appearance of transparency without delivering meaningful accountability.

### How does privacy affect AI matchmaking systems?

 Privacy determines which information may be used, who can be shown, and whether participants have agreed to networking visibility. A relevant match should not override a user’s consent or expose private contact details.

 Privacy-first systems use purpose-limited data, configurable visibility, secure access controls, and clear participation choices. Paid access should never become a route to hidden profiles or private attendee information.

## Build Better Matching Through Accountability

 Bias in matchmaking algorithms cannot be managed through model accuracy alone. Teams need a repeatable process that connects data quality, fairness metrics, explainability, privacy, user feedback, and governance. A well-scoped **algorithmic bias audit** turns those principles into practical tests and documented improvements.

 For organizers seeking a more intentional event experience, MeetWho combines free event creation and participant management with permission-based networking recommendations. Its Event Networking Intelligence approach is designed to help people form fewer but more relevant connections—not simply collect more contacts.

> **Explore MeetWho to create a free event, manage participants, and help attendees discover the right people for meaningful networking.**

## Structured Data

### Article Schema

```
{
"@context": "https://schema.org",
"@type": "Article",
"headline": "Bias in Matchmaking Algorithms and How to Audit It",
"description": "Learn how bias enters matchmaking algorithms and how to conduct an algorithmic bias audit using practical fairness, transparency, and monitoring methods.",
"author": {
"@type": "Organization",
"name": "MeetWho"
},
"publisher": {
"@type": "Organization",
"name": "MeetWho",
"url": "https://meetwho.app/"
},
"mainEntityOfPage": {
"@type": "WebPage",
"@id": "https://meetwho.app/blog/bias-in-matchmaking-algorithms-and-how-to-audit-it"
}
}
```

### FAQPage Schema

```
{
"@context": "https://schema.org",
"@type": "FAQPage",
"mainEntity": [
{
"@type": "Question",
"name": "What causes bias in matchmaking algorithms?",
"acceptedAnswer": {
"@type": "Answer",
"text": "Bias can originate in historical data, incomplete profiles, underrepresentation, proxy variables, inappropriate success metrics, ranking decisions, and feedback loops."
}
},
{
"@type": "Question",
"name": "How do you perform an algorithmic bias audit?",
"acceptedAnswer": {
"@type": "Answer",
"text": "Define fairness objectives, document and review data, test recommendations, compare outcomes, investigate disparities, implement corrective actions, and monitor the system continuously."
}
},
{
"@type": "Question",
"name": "Can AI matchmaking algorithms be completely unbiased?",
"acceptedAnswer": {
"@type": "Answer",
"text": "No system should be assumed to be permanently free from bias. The practical goal is to reduce unjustified disparities through regular testing, transparency, oversight, and remediation."
}
},
{
"@type": "Question",
"name": "Why is explainability important in AI recommendations?",
"acceptedAnswer": {
"@type": "Answer",
"text": "Explainability helps users assess recommendation relevance and enables product teams to investigate inappropriate or inconsistent outcomes."
}
},
{
"@type": "Question",
"name": "How does privacy affect AI matchmaking systems?",
"acceptedAnswer": {
"@type": "Answer",
"text": "Privacy determines which data may be used and whether users can be shown or contacted. Responsible systems prioritize consent, controlled visibility, and purpose-limited data."
}
}
]
}
```

### BreadcrumbList Schema

```
{
"@context": "https://schema.org",
"@type": "BreadcrumbList",
"itemListElement": [
{
"@type": "ListItem",
"position": 1,
"name": "Home",
"item": "https://meetwho.app/"
},
{
"@type": "ListItem",
"position": 2,
"name": "Guides",
"item": "https://meetwho.app/blog/guides"
},
{
"@type": "ListItem",
"position": 3,
"name": "Bias in Matchmaking Algorithms and How to Audit It",
"item": "https://meetwho.app/blog/bias-in-matchmaking-algorithms-and-how-to-audit-it"
}
]
}
```

---

Canonical HTML version: https://meetwho.app/blog/bias-in-matchmaking-algorithms-and-how-to-audit-it
Machine-readable site index: https://meetwho.app/llms.txt