Optimizing for AI agents (AVO): what is known and what is not
More and more people do not visit websites: they ask an agent to search, compare and choose. It is real and growing fast. But most of what is sold under the AVO name cannot be measured yet, and it is worth knowing before paying for it.
What AVO is
AVO, Agentic Visibility Optimization, is the work of getting an AI agent that researches on behalf of the user to find you, understand your facts and choose you when it is given a task, without the user ever visiting your site.
The agent does two different things. First it asks a model, and the usual rules apply: being cited, your brand being understood, third-party sources talking about you. That is GEO. Then it visits your site to find a specific fact: the price, the opening hours, how to hire you. That second part is the new one, and it is where a website wins or loses the agent.
Watch the name: "AVO" is also used for AI Visibility Optimization, which is GEO under another acronym.
What is known, with sources
- It exists and is growing. Traffic from agents browsing like a user was 1.7% of AI traffic in December 2025, growing close to 80 times in a year (HUMAN Security, monthly agentic traffic reports).
- AI browsers drive it. Browsers with a built-in agent make up most of that traffic; in August 2026 OpenAI closed its browser and folded it into the ChatGPT agent (HUMAN Security, August 2026).
- In-chat checkout only works in the United States. OpenAI's protocol (Agentic Commerce Protocol, with Stripe) and Google's (Universal Commerce Protocol) are live with large US merchants. For a European business, the agent paying for the user is not an option yet.
- Agents fail a lot, and the website matters. On the Web Bench benchmark, agents complete the task 100% of the time on apple.com and 35% on booking.com. Bot checks, forms and pages that only work with JavaScript are the most common causes.
What cannot be measured yet
- How many customers reach you through an agent. There is no reliable referrer: the agent leaves no trace in your analytics when it rules you out, and when it picks you the enquiry usually arrives as a normal call or email.
- Whether technical "agent" signals do anything. Serving Markdown, declaring an MCP server, WebMCP in the browser… No major agent has confirmed using them to choose. We audit them as a dated inventory, without scoring: across 23 client sites there was not a single one (finding 08).
- An "AVO score". Any 0-100 number sold to you today mixes the above with assumptions. An indicator that fails almost every website does not tell them apart.
What can be measured today
Three things, all already in the dashboard: whether AI states your operational facts correctly, whether you are chosen when it is given the task, and whether an agent finds on your site what it is looking for.
| What | How it is measured | Why it matters |
|---|---|---|
| Operational facts | The Accuracy Score separates price, opening hours, area, lead time and contact from the rest of your brand facts | An agent acting on an old phone number or a wrong price is worse than not appearing |
| Delegated questions | We measure questions like "find me three, compare them and tell me which one to call", not only "which is the best" | It is how people who hand over the decision ask |
| Task test | An AI browser looks for your price, opening hours and contact on your site, three times each, and reports where it gets stuck. Read-only: it never fills in or submits anything | It separates what is the fault of your site (bot wall, price in a PDF) from the noise of the agent itself |
What to do on your website
- Put the price, or a range, as text on the page. Not only in a PDF or an image.
- Write down your opening hours, service area and how to hire you, not only inside a widget.
- Check that your firewall does not block the search crawlers of AI assistants, and that the bot check does not fire on your contact page.
- Keep those same facts identical on your third-party listings. The agent reads them too.
What this page does not solve It does not say how much this will matter next year. It is the fastest-changing layer: if a major agent starts using a technical signal, or in-chat checkout reaches Europe, the page changes that same month.
Frequently asked questions
What is AVO (Agentic Visibility Optimization)?
It is the work of getting an AI agent that researches on behalf of the user to find you, understand your facts and choose you when given a task, without the user visiting your site. It has two parts: the answer from the model the agent consults, worked like GEO, and the visit of the agent to your site to find a specific fact such as the price or contact route.
Is AVO the same as GEO?
They overlap. The part where the agent asks a model is GEO. What is new is the agent visiting your site: if it cannot find the price or hits a bot check, it rules you out without you knowing. "AVO" is also used for AI Visibility Optimization, which is simply GEO under another acronym.
Can an AI agent buy or book on my website?
In-chat checkout with the OpenAI and Google protocols works today only in the United States and with large merchants. Elsewhere the agent researches and recommends, and a person makes the contact afterwards.
How do I know whether an AI agent finds my price and contact details?
With a task test: an AI browser tries to find your price, opening hours and contact route on your site, several times, and notes where it gets stuck. It must be read-only and separate the failures of your site from those of the agent, because agents also fail by themselves.
Is it worth implementing WebMCP or serving Markdown to agents?
Today there is no confirmation that major agents use those signals to choose. Keep them inventoried and dated, not as a priority: none of the 23 client sites we audited had any.
See what AI says about you today
The free audit measures your presence in buying questions for your sector, delegated ones included, and tells you who appears in your place. No signup, no card.
Run my free audit →Home · Plans & pricing · Methodology · Research · Perplexity · Metrics · How to choose · Use cases · Changelog · Agencies · Alternatives · Diagnosis · Check if AI cites you · Noise and margin of error · Four-step method · X · Product Hunt