How internet platforms work

1. What is an internet platform?

A platform provides rules and technology through which distinct groups interact. A marketplace connects buyers and sellers; a social service connects people, creators and advertisers; an app store connects device users and developers. The platform may supply identity, search, ranking, payments, fraud prevention, delivery, analytics and dispute resolution. It is therefore both an intermediary and a governor of the market it creates.

The term is easily stretched too far. A CDN delivers content but does not necessarily operate a multi-sided marketplace. A retailer sells its own inventory, while a marketplace lets third parties transact. An app may use a platform without itself being one. This guide uses platform for a service that facilitates interactions among distinct, interdependent groups, consistent with the OECD’s digital-platform framework.

A useful investment analysis separates four possible advantages. Economies of scale spread fixed costs over more use. Network effects make participation more valuable as relevant participants join. Data effects improve a service when additional interactions create useful learning signals. Switching costs make departure costly because identity, history, purchases, relationships or integrations are hard to recreate. A company may have all four, one, or none.

2. DNS, internet routing and content delivery networks

When a person opens a site, several systems cooperate before the page appears. The Domain Name System, or DNS, translates a human-readable name into resource records such as an IP address. The namespace is hierarchical: the root points towards top-level domains, which lead to authoritative name servers for particular zones. A recursive resolver performs or coordinates these lookups and caches answers for their stated time to live. This design is set out in RFC 1034.

An address is not a route. The internet is made of independently operated networks called autonomous systems. BGP, the Border Gateway Protocol, distributes which IP prefixes each network can reach and attaches attributes such as the autonomous-system path. Operators apply local business and engineering policy when selecting routes; the result is a policy-acceptable path, not necessarily the geographically shortest one. RFC 4271 defines BGP-4.

A content delivery network, or CDN, keeps reusable content and application services across distributed edge locations. Request routing selects a suitable edge using signals such as client network, geography, server load, capability, network conditions and business policy. DNS responses, HTTP redirection and BGP anycast can all help steer requests. The IETF’s CDN interconnection framework describes those mechanisms; Cloudflare’s CDN architecture shows how a current anycast implementation advertises the same addresses from many locations.

Proximity is only part of CDN value. Repeated cache hits reduce origin load and bandwidth; anycast can absorb attacks and move traffic away from failures; shared edge capacity can support security, application and network products. Scale may improve peering economics and spread fixed infrastructure cost. It also expands the blast radius of a bad global configuration. Cloudflare’s 2025 Form 10-K describes both the scale of its shared network and widespread 2025 outages caused by globally deployed changes.

Routing security remains incomplete. A route leak propagates routing information beyond its intended scope and can redirect, overload or black-hole traffic. RPKI origin validation can check whether an autonomous system is authorised to originate a prefix, but it does not validate the full path or prevent every interception. See RFC 7908 on route leaks and RFC 6811 on origin validation.

3. How search engines work

A general search engine performs expensive preparation before the user asks a question. Crawlers discover URLs through links and sitemaps, fetch accessible pages and render content where needed. Indexing systems parse text and media, identify duplicates, select canonical versions and store representations and signals. At query time, retrieval finds plausible matches and ranking systems order them for relevance and quality. Google’s current documentation describes these as crawling, indexing and serving.

Ranking is not one timeless formula. It combines several systems for lexical matching, semantic understanding, links, freshness, location, language, deduplication, original content and specialised safety or policy needs. Google’s published ranking-systems guide lists PageRank alongside BERT, neural matching, RankBrain and other systems. PageRank remains historically important, but it is not a complete description of modern search.

Scale can reinforce search quality. More queries and clicks provide evidence for spelling, language understanding, evaluation and rare or fresh searches; advertising can finance the crawling, indexing and serving infrastructure that attracts users. The UK’s Competition and Markets Authority documented this feedback loop in its 2025 strategic-market-status decision. Returns are not uniform—common queries may offer less incremental learning—but a rival still needs broad coverage, distribution and sustained investment.

Distribution is therefore part of the product. Defaults in browsers, devices and access points reduce the effort needed to reach a service. In 2025 the US court’s Google-search remedy prohibited specified exclusive distribution arrangements and required certain data and syndication access; the final judgment and compliance material are collected in the Department of Justice case record. The lesson is broader than litigation: technical quality, habitual access and commercial distribution jointly determine share.

4. How recommendation systems work

Search begins with an explicit query. A recommender often begins with a user, context and inventory, then predicts which items will be valuable. At platform scale it is too slow to run the most expensive model over every video, post or product. Systems use a funnel: retrieve a few hundred or thousand candidates, apply progressively richer ranking models, then rerank for policy, diversity, freshness and other constraints.

YouTube’s production research separates candidate generation from ranking. Instagram Explore uses retrieval, first-stage ranking, second-stage ranking and final reranking; cached item embeddings and approximate nearest-neighbour search make a vast catalogue tractable. Meta’s engineering account also shows why cheap early models and heavier late models coexist.

The system learns representations of users and items from features and behaviour. A two-tower model can separately encode a user and an item, then retrieve nearby item embeddings. Later models can add candidate-specific interaction features and estimate several outcomes: a click, long watch, save, share, purchase, report or “see less” action. A weighted objective converts those predictions into a score.

Those weights are product policy expressed mathematically. Optimising clicks is different from optimising completed purchases, long-term satisfaction, meaningful interactions, creator diversity or low complaint rates. Meta describes combining predictions and applying integrity and diversity passes in its explanation of Feed ranking. The exact trade-offs and current weights remain proprietary.

Feedback is not neutral. A person can click only what the earlier system exposed, and prominent items tend to collect more reactions. This creates exposure bias: logged behaviour reflects both preference and the platform’s previous choices. Exploration traffic, randomised experiments and counterfactual methods help separate the two. Google research on off-policy correction explains the problem and one approach.

The economic effect is powerful. A ranking change can redirect attention, merchant sales, creator income and future training data without changing the underlying supply. Better recommendations can enlarge consumption by reducing search cost; poorly chosen objectives can concentrate exposure, reward sensational material or trap the model in its own feedback. In the EU, very large platforms face recommendation transparency, risk-assessment and non-profiled-option duties under the Digital Services Act.

5. Social graphs and interest graphs

A social graph represents entities as nodes and their relationships as edges: a person follows another person, belongs to a group, comments on a post or tags a photograph. Attributes describe the entities and connections. Meta’s distributed graph store TAO was designed around these read-heavy traversals; its engineering overview calls this the power of the graph.

The graph provides explicit context. If close friends post, the platform already has a candidate source. It also creates direct network effects: a communication service is more useful when the people a user wants to reach are present. Accumulated connections, shared history and identity create switching costs because moving an account does not move the community.

Modern discovery relies on more than declared connections. An interest graph is a useful analytical label for inferred relationships between users and topics, creators, formats or products based on viewing and interaction patterns. Recommendations can therefore surface material from accounts a user never followed. Meta’s 2025 filing says AI powers a discovery engine across its services; that suggests the durable asset is the combination of graph data, behavioural feedback, content supply and ranking infrastructure, not simply the friend list. Meta 2025 Form 10-K.

Graphs can also create negative effects. More creators mean more competition for attention. More sellers can increase selection for buyers while making discovery harder for each seller. Spam and coordinated abuse become easier as a network grows. The platform’s trust, ranking and moderation systems determine whether participation improves the experience or merely adds congestion.

6. How advertising auctions work

Read the dedicated digital advertising infrastructure deep dive → for the complete transaction path, header bidding, DSP and SSP economics, supply-path optimisation, fraud controls and competition analysis.

Digital advertising joins three problems: choosing an eligible ad, predicting its value and setting the price. A bid is important but rarely acts alone. Search-ad systems also consider relevance, expected response, landing-page quality, policy, minimum thresholds and the context of the query. Google’s advertiser documentation lists six factors in its ad auction and explains why a more relevant ad can outrank a higher nominal bid.

Automated bidding moves the optimisation target beyond a fixed cost per click. An advertiser may ask a system to maximise conversions or conversion value within a budget or return constraint. The platform predicts how valuable each opportunity is and submits an auction-specific bid. Better prediction can improve advertiser return, but it also makes auditing the division of value between advertiser and platform more difficult.

Open-web display advertising commonly uses real-time bidding. A publisher or its supply-side platform describes an available impression in a bid request. Exchanges send it to demand-side bidders, which respond within a tight millisecond deadline. The exchange applies eligibility, floors and auction rules, selects a winner and sends notices for settlement and measurement. The OpenRTB 2.6 specification standardises this communication, not every participant’s business logic.

OpenRTB supports first-price and second-price-plus auction types. In a first-price auction the winner pays its submitted price, encouraging bidders to estimate the minimum winning bid rather than simply reveal full value. “Second-price-plus” should not be treated as a pure textbook Vickrey auction: floors, increments, fees, bid shading and sequential auctions change incentives.

Vertical integration can reduce latency and coordination cost, but also creates conflicts when one company represents buyers, represents sellers and operates the exchange. A US federal court found in April 2025 that Google unlawfully monopolised defined open-web publisher-ad-server and ad-exchange markets; the Department of Justice summary should be read for the exact legal scope. It is not a finding about every digital-advertising product.

7. Identity, attribution and measurement

An auction can optimise only what the system can observe or estimate. Advertising measurement tries to connect exposure to an outcome: a click, signup, store visit or purchase. The link may use a logged-in account, first-party cookie, mobile advertising identifier, third-party cookie, conversion tag, server-to-server event or a privacy-preserving aggregate report.

Attribution assigns credit; it does not prove causation. A customer who clicked an ad may have purchased anyway. Incrementality experiments compare exposed and control populations to estimate what the advertising changed. Marketing-mix models use aggregated time-series data; conversion models infer missing outcomes; clean rooms allow constrained matching or aggregation across parties. Each trades granularity, timeliness, coverage and privacy differently.

Closed platforms such as Google Search, Meta’s apps and Amazon’s marketplace observe activity within a logged-in or first-party environment. That can make their measurement more resilient than cross-site advertising when browser identifiers disappear. It does not make measurement perfect: conversions can happen elsewhere, platform-reported attribution can overlap, and the platform marking its own homework creates an incentive for independent testing.

8. Third-party cookies: what they are and what changed

Read the dedicated third-party cookies deep dive → for the full browser timeline, privacy and competition debate, replacement stack and company implications.

A cookie is a small piece of browser-held state exchanged through HTTP headers. It is first-party when its site matches the top-level site the person is visiting and third-party when it belongs to a different embedded site. The modern cookie specification defines attributes such as SameSite, Secure and HttpOnly; see the IETF’s current RFC6265bis draft.

Suppose many publishers embed an advertising resource from the same domain. Historically, that domain could read the same unpartitioned cookie across those sites. A stable pseudonymous identifier helped the ad-tech provider connect impressions, build audience segments, control frequency and attribute later conversions. It did not need to contain a person’s name to enable cross-site recognition.

Third-party cookies are also used for embedded sign-in, payments, maps, chat and fraud prevention. Browsers therefore use several approaches: blocking, partitioning, explicit storage access and user controls. A partitioned cookie is additionally keyed by the top-level site, so the embedded provider cannot use one value to link activity across unrelated sites. MDN’s CHIPS guide explains this double-keying.

The phrase “cookies are dead” is wrong. Safari blocks third-party cookies by default; Firefox’s Total Cookie Protection partitions them; Chrome retained its user-choice approach in April 2025 rather than completing a universal phase-out. Availability therefore depends on browser, mode, settings and context. The IETF recommends restrictive policy as compatibility permits and warns that sites cannot rely on consistent third-party-cookie treatment.

Cookies are not the whole privacy question. Login identifiers, pixels, link decoration, IP addresses, SDKs and device fingerprinting can also connect behaviour. The UK’s Information Commissioner’s Office states that its 2026 storage and access guidance covers cookies, pixels, fingerprinting and similar technologies. Legal requirements vary by jurisdiction, but renaming or moving an identifier to a server does not remove the underlying privacy and governance risk.

9. Privacy-preserving advertising

Privacy-preserving advertising is not one product. It is a family of designs that try to retain useful selection or measurement while reducing exposure of individual behaviour. Common techniques include processing on the device, partitioning state, releasing only aggregated results, delaying reports, limiting the number of bits disclosed, applying minimum crowd thresholds, adding statistical noise and using cryptography so parties can compute an intersection or total without revealing their raw inputs.

Apple’s Private Click Measurement, for example, limits attribution information and adds a random reporting delay; AdAttributionKit releases more detail only when privacy thresholds are met. NIST describes secure multiparty computation, private set intersection, homomorphic encryption and differential privacy as complementary privacy-enhancing techniques. None automatically supplies consent, eliminates every re-identification risk or preserves the precision of unrestricted user-level logs.

Chrome’s Privacy Sandbox is a caution against treating proposals as settled infrastructure. In April 2025 Google retained third-party-cookie choice. In October 2025 it announced the retirement of Topics, Protected Audience, Attribution Reporting, Private Aggregation and several related advertising APIs after low adoption. Google’s status page, updated 14 August 2026, lists those technologies for deprecation or removal while continuing CHIPS, FedCM, Storage Access and other privacy or security features.

Cross-browser private attribution remains an active standards problem. The W3C’s Private Advertising Technology Working Group is working on APIs with stronger technical privacy assurances, but a charter or proposal is not a deployed interoperable standard. No single replacement currently gives the open web old-cookie levels of cross-browser targeting, frequency control, fraud defence and attribution.

The durable response is a portfolio: consented first-party relationships; contextual signals about the page or query; platform-specific measurement; incrementality tests; modelled outcomes; and aggregate or cryptographic collaboration where it genuinely reduces disclosure. This tends to advantage companies with direct user relationships and large first-party environments, but can also create an opening for independent measurement that advertisers trust.

10. Marketplaces, network effects and platform economics

A two-sided marketplace must solve a coordination problem. Buyers want selection and reliable fulfilment; sellers want reachable demand. More sellers can make the service more useful to buyers, and more buyers can attract sellers. These are cross-side network effects. The platform may subsidise one side—possibly with a zero monetary price—and earn revenue from commissions, ads, payments, logistics or subscriptions on another. The foundational economics are developed in Rochet and Tirole’s platform-competition research.

Amazon illustrates how layers reinforce one another. It combines first-party retail with third-party marketplace services, fulfilment, Prime, discovery and sponsored advertising. Its 2025 filing reported $172.2 billion of third-party seller-services revenue and $68.6 billion of advertising-services revenue, alongside $269.3 billion of online-store sales. Seller services include commissions, fulfilment and shipping, so the figure is not a pure take rate. Amazon 2025 Form 10-K.

The flywheel is more specific than “more users equals a moat.” More selection must be relevant and trustworthy. More sellers can create congestion, counterfeit risk and advertising pressure. Participants may use several marketplaces, weakening lock-in. Logistics density and fulfilment quality are economies of scale, while buyer–seller attraction is a network effect. Reviews and transaction history can create data advantages; Prime membership and operational integration can create switching costs.

Governance is inseparable from economics. A marketplace can rank offers, set fees, define data access, sell advertisements, operate fulfilment and compete with participants. Integration can reduce transaction cost; it can also create incentives for self-preferencing or fee extraction. Allegations should not be reported as adjudicated facts, but the FTC’s Amazon marketplace complaint and the EU’s Digital Markets Act designations show why these control points attract scrutiny.

11. App stores and the economics of access

An app store bundles distribution, discovery, signing and identity, review, updates, payments, refunds, analytics and access to an installed device base. Users value a broad, trusted catalogue; developers value reachable users. A new store faces a chicken-and-egg problem: without users it struggles to attract apps, and without apps it struggles to attract users. The UK’s mobile-ecosystems market study documents these indirect network effects and the role of defaults and operating-system control.

Store rules convert technical control into governance. Apple reviews apps, updates, in-app purchases and events under its App Review Guidelines. Google describes release tools, promotion, experimentation, analytics and safety as part of Google Play’s service. These are real services, but the platform owners’ descriptions are not independent evidence that every fee equals cost.

There is no universal 30% commission. Fees differ by developer size, subscription status, transaction, programme, jurisdiction and whether the platform’s payment or distribution channel is used. Google says most fee-paying developers qualify for 15% or less; Apple’s small-business programme uses a 15% rate for qualifying developers. Exact schedules are volatile, so current terms matter more than a timeless headline.

The strategic question is who controls discovery, installation and payment. That control can generate attractive toll economics and protect users from malicious software, while constraining alternative stores, billing and business models. Rules are diverging by jurisdiction. The EU’s DMA app-distribution portal describes obligations for designated gatekeepers covering alternative distribution, steering and purchasing. In the US, the Ninth Circuit upheld anti-steering relief in Epic Games v Apple while preserving the rejection of Epic’s broader federal monopolisation claim. Those are different legal conclusions.

12. Mapping the stack to Alphabet, Meta, Amazon and Cloudflare

Alphabet’s reporting groups Search, YouTube, network advertising, subscriptions, platforms and devices inside Google Services. Its financial glossary also identifies traffic-acquisition costs paid to distribution and network partners. This is a reminder that even a strong network effect may be purchased and maintained through commercial agreements. Alphabet financial-statements glossary.

Meta reported 3.58 billion average Family daily active people in December 2025; ad impressions and average price per ad both rose during 2025. Those metrics separate inventory growth from monetisation per impression, but do not reveal whether value came from more time, better ranking, improved conversion prediction or auction demand. Meta 2025 results.

The companies increasingly overlap. Amazon competes for commerce discovery and ad budgets; Meta and YouTube compete for creators and attention; AI assistants alter how people reach information and products; Cloudflare sells developer, connectivity and security services at the delivery layer. The right competitive boundary depends on the user job and revenue pool, not on a single label such as “big tech.”

13. What to watch

The central analytical question is where control and feedback sit. A platform becomes economically powerful when it repeatedly improves a valuable interaction, controls scarce access to it and can charge without degrading participation. Its durability depends on whether users and counterparties remain because the service is genuinely better, because leaving is difficult, or both.

Technical architecture sets the possibilities; product objectives allocate attention; commercial rules divide the value; privacy and competition law bound the design. Following all four is the best way to understand how internet platforms create—and sometimes lose—power.