Observability is the ability to work out what is happening inside a software system by examining the data it emits. In practice that data comes in three main forms: logs, which record discrete events; metrics, which count and measure over time; and traces, which follow a single request as it passes from one service to the next. The term exists because modern systems are too distributed for anyone to predict every way they can break. A dashboard built for known failures says little about a failure nobody anticipated. Observability is the property that lets an engineer ask a new question of a running system without shipping new code to answer it.

For advertising, the question is rarely academic. A programmatic auction, a conversion tag or an AI buying agent can fail silently, and the first symptom is often lost revenue.

From control theory to distributed software

The word comes from engineering mathematics. Rudolf Kalman introduced observability for linear dynamical systems in his 1960 paper "On the general theory of control systems", pairing it with its mathematical dual, controllability. A system is observable, in that sense, when its internal states can be inferred from knowledge of its external outputs.

Software borrowed the term much later. On September 9, 2013, Twitter's engineering blog published "Observability at Twitter", describing a team whose platform collected roughly 170 million individual metrics per minute from tens of thousands of endpoints, and consumed about 1.5% of the company's data centre machines. At Twitter the word was close to a synonym for telemetry.

The sharper definition came from Charity Majors, who co-founded Honeycomb with Christine Yen. According to a 2021 retrospective in The New Stack, Majors looked up the control-theory definition in July 2015 and found it matched what the pair were building: tooling for "unknown-unknowns" rather than the known failures that monitoring tools were designed to catch. In February 2017 Peter Bourgon, writing after a tracing summit, proposed the three pillars framing of metrics, logs and traces. Vendors adopted it quickly. Majors did not, arguing that "the world doesn't need another classification word for existing data types."

How it works

An observable system starts with instrumentation: code inside an application, or an agent beside it, that emits telemetry. Each piece carries context. A log line records a timestamp, a severity and a message. A metric records a value, such as requests per second or 95th percentile latency, with labels attached. A trace is built from spans, each representing one unit of work, linked by a shared trace identifier so that a request touching 12 services can be reassembled as one tree.

The telemetry is then shipped, usually through a collector that batches, filters and samples it, to a storage backend where it can be queried. Alerting sits on top. So does exploration: the ability to slice data by an attribute nobody thought to chart in advance, such as a single campaign ID, browser version or supply partner.

OpenTelemetry (OTel) is the open standard that now governs most of this plumbing. It was formed in May 2019 by merging two earlier projects, OpenTracing and OpenCensus, under the Cloud Native Computing Foundation (CNCF), the Linux Foundation body that also hosts Kubernetes. It entered the CNCF sandbox shortly after the merger and was accepted as an incubating project, according to a CNCF post dated August 26, 2021. OTel defines vendor-neutral APIs, software development kits, a collector and semantic conventions - agreed names for attributes so that one tool's "http.request.method" means the same thing as another's.

Where it sits in ad tech

Programmatic advertising is a distributed system with hard deadlines, which makes it an obvious candidate. In OpenRTB, the tmax field sets how many milliseconds an exchange will wait for bids. The PPC Land explainer on latency notes that Google Authorized Buyers deadlines range from 80 to 1,000 milliseconds, that Google throttles callouts when a bidder's error rate exceeds 15%, and that Prebid recommends an auction timeout of 1,000 milliseconds or less. Knowing which bidder breached which deadline, and why, is an observability problem. AdPushup's header bidding analytics, launched in February 2021, used bidder response times to set timeouts per segment.

Transaction identifiers play a role similar to trace IDs. In June 2022 an OpenRTB extension introduced a per-impression transaction ID that let buyers match the same impression across exchanges. When Prebid.org disabled that cross-exchange behaviour on August 27, 2025, buyers lost a tool they had used to detect duplicated requests; 59% of bid requests carried the IDs before the change. IAB Tech Lab said the change "materially violates the OpenRTB specification", and the standards clash pitted buyers' need for visibility against publishers' concerns about data leakage.

Measurement pipelines carry the same risk. Google's Tag Diagnostics, announced in June 2024, grades tag health from Excellent to Urgent and flags missing conversion linkers or untagged pages. In October 2024 Google Tag Manager added an alert for a Google tag that has stopped transmitting data within 48 hours. Those checks have limits. One advertiser's Google Ads conversions fell 90% overnight because a consent banner collected choices but never passed them to Google's tags, and the consent diagnostics carry a detection lag of 48 to 72 hours.

AI agents and LLM observability

The newest use is tracing what large language model (LLM) applications and autonomous agents actually do. OTel's GenAI semantic conventions model an agent run as a tree: an invoke_agent span as the parent, chat spans for each model call and execute_tool spans for tool calls, with attributes such as gen_ai.usage.input_tokens recording token counts. As of mid-2026 every gen_ai attribute remained at "Development" status rather than stable, and in June 2026 the conventions moved to a dedicated repository.

Vendors of agent platforms now treat this as a governance feature. In April 2026 Salesforce added Observability and Session Tracing designed to show "the reasoning behind" agent actions. Google Cloud said in October 2026 that its Gemini agents log every action to an audit trail attributed to the agent rather than a person. The Agentic Real Time Framework from IAB Tech Lab, finalised in August 2026, makes OpenTelemetry metrics and distributed tracing mandatory for containers running inside an auction.

Why it matters for marketers

Most advertising failures surface first as money. In January 2026, publishers reported revenue falls of 50% to 90% across AdSense and Ad Manager; detection came from their own dashboards and forums rather than any Google alert. Platform status pages are not a substitute. PPC Land's Search Status Dashboard explainer notes that a start time on Google's page marks when Google confirmed a problem, not when it began.

Automated buying raises the stakes. In a Nylas survey of 1,026 practitioners, monitoring and observability was cited as a barrier by 30.8%. At IAB Europe, a speaker listed a lack of observability at the inference layer among four problems with single agents.

Limitations and disputes

Definition creep is the oldest complaint. O'Reilly's 2021 trends report found interest in observability grew 128% while monitoring grew 9%, and warned that "observability risks becoming the new name for monitoring." Whether three pillars describe observability or simply rebrand existing data types remains contested.

Cost scales with data. Every span and high-cardinality label must be stored, so teams sample, discarding most traces and sometimes the one that mattered.

Privacy is the sharpest issue for advertising. Telemetry captures identifiers by design. An IMDEA study covered by PPC Land on October 5, 2026 found that ChatGPT and Claude sent conversation IDs to Datadog on the web. UC Berkeley's agentic AI risk profile recommends comprehensive agent logging while acknowledging it may function as continuous surveillance.

Sealed systems cut the other way. PPC Land's ARTF coverage argues that a container "neither counterparty inspects by design" weakens observability even when telemetry is mandated, and rejection reporting remains optional.

Not the same as

  • Monitoring watches predefined signals against thresholds. Observability is the broader capacity to investigate questions not anticipated in advance; monitoring is one activity it supports.
  • Social listening, sometimes called brand monitoring, tracks public conversation about a brand. It concerns audiences, not system internals.
  • AI visibility measures how often a brand appears in AI-generated answers. Despite similar vocabulary, it is a marketing metric, not telemetry about software.
  • Analytics describes user behaviour for business decisions. Observability describes system behaviour for diagnosis, though both can draw on the same events.

Recent developments

The CNCF announced OpenTelemetry's graduation on May 21, 2026, at the Observability Summit in Minneapolis, its highest maturity level. The foundation counted more than 12,000 contributors from over 2,800 companies, and more than 1.36 billion downloads of the JavaScript API package in the preceding 12 months. Profiles, a fourth signal, remain in alpha. Chris Aniszczyk, CNCF chief technology officer, called it "the essential, unified observability standard."

Ad tech is converging on the same tooling. ARTF version 1.0 was declared final on August 12, 2026, with OpenTelemetry written into its requirements. Meanwhile AI agent platforms from Salesforce, Google Cloud and others now sell tracing and audit logs as conditions for handing agents budgets. Whether buyers will see those traces, or only vendors' summaries of them, is the question the industry has not settled.

Timeline

  • 1960 - Rudolf Kalman introduces observability for linear dynamical systems in "On the general theory of control systems".
  • September 9, 2013 - Twitter publishes "Observability at Twitter", describing 170 million metrics collected per minute.
  • July 2015 - Charity Majors adopts the control-theory definition while building Honeycomb.
  • February 2017 - Peter Bourgon proposes the three pillars of metrics, logs and traces.
  • May 2019 - OpenTracing and OpenCensus merge to form OpenTelemetry under the CNCF; it enters the sandbox shortly after.
  • August 26, 2021 - CNCF announces OpenTelemetry as an incubating project.
  • June 2022 - OpenRTB extension for per-impression transaction IDs approved.
  • June 2024 - Google announces Tag Diagnostics.
  • October 2024 - Google Tag Manager adds an alert for tags that stop transmitting for 48 hours.
  • August 27, 2025 - Prebid.org disables cross-exchange transaction IDs.
  • January 13-14, 2026 - Google ad platform failures; publishers report revenue falls of 50% to 90%.
  • April 15, 2026 - Salesforce announces Observability and Session Tracing for agents.
  • May 21, 2026 - OpenTelemetry graduates from the CNCF.
  • June 12, 2026 - OTel GenAI conventions move to a dedicated repository (release 1.42.0).
  • August 12, 2026 - IAB Tech Lab declares ARTF 1.0 final, with mandatory OpenTelemetry.
  • October 8, 2026 - Google Cloud details audit trails and spend caps for Gemini agents.

Summary

Who: Software and site reliability engineers instrument systems; standards bodies such as the CNCF (OpenTelemetry) and IAB Tech Lab (ARTF) define the formats; ad tech platforms, publishers, advertisers and AI agent vendors rely on the resulting data.

What: The ability to understand a system's internal state from the telemetry it emits, mainly logs, metrics and traces, including for questions not anticipated in advance.

When: Defined in control theory by Rudolf Kalman in 1960, adopted in software from 2013, standardised through OpenTelemetry from May 2019, which graduated from the CNCF on May 21, 2026.

Where: Across distributed software, including programmatic auctions, header bidding wrappers, tag and measurement pipelines, and LLM-based agents.

Why: Distributed systems fail in ways nobody predicts, and in advertising those failures show up as lost bids, broken conversion data and unaccountable automated spend.