Glossary · Beginner
What Is an AI Agent? Definition and Example
An AI agent is an AI system that can plan and carry out actions toward a goal, rather than only answering a question. Agents may browse websites, use tools, fill in forms, compare options or complete tasks across several steps on a user's behalf. For marketers, agents are a new kind of visitor that reads, evaluates and sometimes buys.
On this page
- Chatbot vs agent
- How do agents connect to websites and tools?
- Why do agents matter for GEO?
- How do you prepare?
- From MentionX data: why readable sources matter
- Worked example: Ledgerline prepares for agents
- What can agents do today?
- How do agents identify themselves?
- What are the risks?
- Common questions
- Related
Checked on 31 October 2026. "Our view" marks MentionX opinion. Ledgerline is a fictional company used for illustration.
Chatbot vs agent
| Chatbot / assistant | AI agent | |
|---|---|---|
| Main job | Answer questions | Achieve a goal |
| Actions | Generates text | Browses, clicks, uses tools, transacts |
| Steps | Usually one response | Multi-step plans |
| Example request | "What are good CRMs for nonprofits?" | "Find three CRMs under €50 a month, compare them and book demos" |
How do agents connect to websites and tools?
Some agents browse the web like a user. Others connect to tools and data through protocols such as the Model Context Protocol, or to commerce systems through specifications such as OpenAI's Agentic Commerce Protocol. Providers that fetch pages on a user's request often publish separate user agents; for example, OpenAI documents ChatGPT-User (OpenAI).
Why do agents matter for GEO?
When an agent shortlists products, it needs information it can read and trust: clear pricing, specifications, availability and policies. If that information is hidden behind scripts, images or sales calls, the agent may skip you or rely on third-party sources instead.
How do you prepare?
- Make key facts available in plain HTML text
- Keep pricing, specs and policies current and consistent across sites
- Use structured data where it applies
- Check that firewalls don't block legitimate user-triggered agents you want to allow
- Watch logs for agent user agents
From MentionX data: why readable sources matter
Method. Read-only aggregate queries over MentionX scans from 22 August to 29 October 2026 (UTC): 2,217 engine runs (2,154 completed) across 18 anonymised projects and six engines.
MentionX doesn't track agents directly, but its answer data shows how engines gather information. Between 22 Aug and 29 Oct 2026 (2,217 runs, 18 projects), engines cited these page types:
| Cited URL type | Share of 9,750 citations |
|---|---|
| Blog or article | 19.6% |
| Homepage | 10.0% |
| Review page | 2.7% |
| Docs | 2.6% |
| Pricing | 1.9% |
| Wikipedia | 1.6% |
Pricing pages were only 1.9% of cited URLs and docs 2.6%. Those are exactly the pages an agent needs to verify a product. Engines also differed in how many sources they used: Perplexity averaged 9.56 citations per answer, Gemini 4.77 and AI Mode 3.15.
Our view: if your pricing and specifications aren't readable, agents will rely on third-party sources or skip you. We have no data on agent task completion, so treat this as a reason to check your own pages, not a measured effect.
Limits. A convenience sample of mostly B2B software projects, run through engine APIs, which can behave differently from the consumer apps. Perplexity, Copilot and DeepSeek figures come from a single day (26 October 2026), so treat them as a snapshot, not a trend.
Worked example: Ledgerline prepares for agents
Ledgerline is a fictional accounts-payable tool. Figures are illustrative.
Ledgerline, a fictional accounts-payable tool, asks an agent to "find three AP tools under €500 a month that integrate with NetSuite and book a demo". The agent finds Ledgerline but can't confirm its price, because the pricing table loads from a script, and moves on.
The team publishes pricing in plain HTML, adds an integrations page listing what syncs with NetSuite, and makes the demo form usable without a pop-up. It checks its firewall logs to make sure user-triggered agents it wants to allow, such as ChatGPT-User, aren't blocked.
What can agents do today?
Capabilities vary by product. Some agents browse websites and fill in forms; some call tools through connectors; some can complete purchases with supported merchants. OpenAI's Agentic Commerce Protocol, for example, describes how agents and merchants exchange checkout information (OpenAI). Expect capabilities to change quickly and check each provider's documentation.
How do agents identify themselves?
Some providers publish user agents for fetches made on a user's behalf, such as ChatGPT-User (OpenAI). Others browse through a standard browser and are hard to tell apart from people. Look for documented user agents in your logs, and don't rely on blocking by user agent alone.
What are the risks?
Agents can misread pages, act on outdated information or be tricked by hidden instructions on untrusted sites. For brands, the main risks are being skipped because facts aren't readable and being misrepresented through stale third-party data. Keep facts consistent everywhere they appear.
Common questions
Should I block AI agents? Only if you have a clear reason. Blocking user-triggered agents can stop buyers' agents from considering you.
Do agents read structured data? Some may. It helps where it matches visible content, but plain readable text matters most.
Related
Questions
How is an AI agent different from a chatbot?
A chatbot answers questions. An agent can also take actions, such as browsing sites, using tools or completing tasks, often across several steps.
Why do AI agents matter for marketers?
Agents may research and shortlist products for people, so your site needs to be readable and your information accurate for machines as well as humans.
Do agents follow robots.txt?
It depends on the provider. Some publish user agents and controls; check each provider's documentation.