The vocabulary around agentic web monitoring formed faster than the products did. A handful of terms now mean something specific that the press and most buyers are still using loosely. This glossary fixes the definitions while the category is still young enough to fix.

Each entry is the shortest accurate definition I can write, followed by a paragraph of context. Cross-references point to related terms in this list and to the longer pieces elsewhere in the series.

Fixing the vocabulary before the press waters it down.

Agentic Search

Agentic search is search that runs in the background on a recurring schedule, takes action without re-prompting, and returns results when something matches. The user states a topic or URL once; the system samples or polls continuously; results arrive when conditions are met instead of when the user asks.

Agentic search names the umbrella that contains both open-web agents and page monitors. The term entered consumer vocabulary at Google I/O 2026, when Google announced its "information agents" feature inside Search. Before that, the same behavior had been shipping for years under names like "saved searches," "alerts," and "monitoring."

Not to be confused with AI search, which is a single on-demand query answered by a language model. Agentic search is recurring; AI search is one-shot.

Information Agent

Information agent is the consumer-facing name for an AI agent that continuously samples the open web for a topic and reports back. The input is a topic. The output is a synthesized briefing. The freshness model is wide and shallow.

Google introduced the term at I/O 2026 for the feature rolling out summer 2026 to AI Pro and Ultra subscribers ($20 to $200 per month). ChatGPT Pulse, live since September 2025, fits the same architectural definition though OpenAI uses different branding. The term is a category name now; expect competitors to adopt it.

See also: the two architectures of agentic search for why information agents and page monitors are different systems.

Open-Web Agent

Open-web agent is the architectural term for any agent that samples the broad web (rather than watching one specific URL) on behalf of a user. The named consumer instances are Google Information Agents, ChatGPT Pulse, Exa, and Perplexity.

The architecture inherits from search and retrieval. An open-web agent crawls, indexes, ranks, and synthesizes across sources. It does not produce timestamped evidence about a specific page. The freshness model is "what is happening in this topic this week," not "what is on this page right now."

Not to be confused with a web crawler. A crawler indexes pages. An open-web agent uses a crawler (usually somebody else's) to answer a recurring question.

Page Monitor

Page monitor is the architectural term for a system that watches one specific URL on a schedule, detects changes, classifies them, and routes an alert. The input is a URL. The output is a change event. The freshness model is narrow and deep.

Visualping has been the category leader in this layer since 2017, with a sample of 1,874,911 active monitors on a single day in May 2026 and 58.8 percent of those owned by a team or workspace. Page monitors inherit from testing and observability, not search. The two architectures are complementary, not competitive.

See also: the precision-layer explainer.

Monitoring Primitive

Monitoring primitive is the lowest-level building block that any agent or workflow needs when it has to know that web state has changed. A monitoring primitive exposes a small interface: create a monitor on a URL, read recent change events, receive a webhook when something changes.

Higher-level agents call monitoring primitives the same way browsers call TLS or databases call an index. The Visualping MCP and REST API are shaped this way. The term is useful because it separates "the thing that watches" (a primitive) from "the thing that decides what to do with the watch" (an information agent or coding agent).

See also: how Claude Code and Codex use Visualping as a monitoring primitive.

Discovery Layer

Discovery layer is the part of an agentic search stack that finds sources you do not yet know about. It samples the open web, surfaces candidate URLs, and synthesizes summaries across many pages.

The discovery layer is wide and shallow. It is the right layer when the user has a topic but no specific URL in mind. Google Information Agents, ChatGPT Pulse, and Exa all sit in this layer. The discovery layer's output is usually a list of candidate URLs (or a summary across them) that gets handed downstream to the precision layer for tight tracking.

Precision Layer

Precision layer is the part of an agentic search stack that verifies specific URLs on a defensible schedule with audit-grade evidence. It does not synthesize across sources. It tells you, with a timestamp, that one specific page changed.

The precision layer is narrow and deep. It is the right layer when the user already knows the URL and needs to be notified the instant it changes. Visualping is the category leader in this layer. Most real workflows use both layers together: discovery feeds precision, precision feeds action.

Change Event

Change event is the unit of output a page monitor produces when a watched page changes. A well-formed change event contains the URL, a timestamp, a visual or structural diff, an optional AI-generated summary of what changed, and (if the platform is doing its job) a binary "important" flag.

Change events are typed objects, not strings. They get routed to email, Slack, webhook, REST API consumers, or directly into an agent stack via MCP. The format Visualping uses, with sample fields shown in the developer cornerstone, is a reasonable industry default for any monitoring primitive.

Diff Fidelity

Diff fidelity is the accuracy of a page monitor's change detection. High-fidelity diffs catch meaningful changes (a price moved, a sentence was reworded, a button was renamed) while suppressing noise (a randomized hero image, a relative-time stamp, an analytics widget that re-renders on every fetch).

Diff fidelity is the hardest engineering problem in page monitoring. Pure visual diffing has high recall but low precision on dynamic pages. Pure text diffing misses layout changes. The frontier is AI-classified diffing, where an importance prompt filters changes to a specific concern ("alert me when pricing changes, not when the hero image rotates"). The Visualping AI importance system shipped in 2024 and remains the differentiator most enterprise buyers cite.

AI Importance Classification

AI importance classification is the layer that decides whether a detected change is worth alerting on, based on a user-supplied prompt or a default importance model. It runs after diff detection and before notification routing.

The binary output ("important" or "not important") lets downstream systems behave differently. A noisy channel like Slack should fire only on important changes; a quieter audit log should record every change. Every Visualping change event includes this binary flag plus a plain-English summary. The classification depends on the prompt; a well-written importance prompt can cut false positives by an order of magnitude on a noisy page.

Audit-Grade Evidence

Audit-grade evidence is a change record that holds up in a compliance review, a legal proceeding, or an internal audit. The minimum bar: a timestamped screenshot of the page state before the change, a timestamped screenshot of the page state after, and a tamper-evident link between them.

Open-web agents do not produce audit-grade evidence. Their architecture (sampling many sources broadly) does not generate the artifact. Page monitors do, because the architecture (polling one URL on a schedule) naturally produces the record. This is one of the load-bearing reasons compliance teams, legal departments, and arbitration firms have been building on page monitors for nearly a decade.

Model Context Protocol (MCP)

Model Context Protocol (MCP) is an open standard that lets AI agents call external tools through a unified interface. An MCP server exposes a small set of named tools; an MCP client (Claude, ChatGPT, OpenAI Codex, Google Antigravity) discovers and invokes those tools by name during a conversation.

The protocol was published by Anthropic in late 2024 and adopted across the AI client space in 2025-2026. MCP is the integration surface that lets a monitoring primitive like Visualping show up inside an AI agent's tool list without custom plumbing. The Visualping MCP endpoint, in public beta as of May 2026, is https://visualping.io/mcp/sse.

Webhook

Webhook is a user-configured HTTP POST that a service sends to a target URL when an event happens. In the monitoring context, the webhook fires when a change event is produced, carrying the event payload to a downstream system (a Slack channel, a CRM, a ticketing system, an AI agent).

Webhooks are the canonical "push" alternative to polling for monitoring data. A well-built monitoring primitive retries failed webhook deliveries with exponential backoff and expects a fast HTTP 200 from the inbound endpoint. The Visualping webhook contract and recommended downstream patterns are documented in the developer cornerstone.

Cadence

Cadence (or check frequency) is how often a page monitor fetches its target URL. Cadence is set per monitor. Tighter cadences catch changes faster; looser cadences are cheaper.

Common cadences for serious use cases: one minute (deploys, status pages), five minutes (pricing pages, regulatory dockets), fifteen minutes (job boards, news pages), one hour (SEC filings, agency rulemaking), twenty-four hours (slow content). In a same-day snapshot of active Visualping monitors on May 24, 2026, 52,452 ran on a five-minute-or-faster cadence. Sub-minute monitoring exists but is volumetrically expensive and usually reserved for enterprise plans.

Frequently asked

What is the difference between an information agent and a page monitor?

An information agent samples the open web for a topic on a recurring schedule and synthesizes results. A page monitor watches one specific URL on a schedule and reports the diff. The two products live in different architectural layers and serve different jobs; most serious workflows use both, with discovery feeding precision.

When did Google announce "information agents"?

Google announced information agents at Google I/O 2026 on May 19, 2026, with a rollout planned for summer 2026 to AI Pro and Ultra subscribers ($20 to $200 per month). The name gave the broader agentic-search category its first consumer-facing label.

Is "agentic search" the same as "AI search"?

No. AI search answers a single on-demand query using a language model. Agentic search runs in the background on a recurring schedule, takes action without re-prompting, and returns results when conditions are met. The same vendor often ships both surfaces, but the user-facing behavior is different.

What is the simplest way for an AI agent to monitor a URL?

Add the Visualping MCP to Claude or ChatGPT as a custom connector at https://visualping.io/mcp/sse, then ask the agent to "monitor this URL." The full walkthrough lives in the developer cornerstone for this series. Setup takes four clicks and zero code.

Why is "monitoring primitive" a useful term?

It separates the work of watching a URL (a small, stable interface any agent can call) from the work of deciding what to do with the watch (which lives in the calling agent or workflow). The separation lets one tool stay small and focused while the agent stack above it grows. The pattern is the same one that made TLS, OAuth, and webhooks load-bearing infrastructure.

Closing

A category gets a name once. It gets a vocabulary over the next two years. The terms above are the working set as of mid-2026. Some will calcify; some will be replaced by sharper words; some will turn out to have been wrong about the underlying architecture. The point of writing them down now is to give the next round of work a stable place to start from.

Eric Do Couto is Head of Marketing at Visualping. He writes about the architecture of monitoring, AI agents, and the slow business of category formation.