Text-reading vs visual AI agents: how each books a hotel
Updated By Akkrev
Some AI agents read a page's structure as text, others look at screenshots and click. They fail in different places on hotel booking sites.
“An AI agent visited my website” can mean two quite different things. Some agents read the page as structured text. Others look at the screen and click, like a person. Many do both. Each type finds different parts of a hotel booking journey hard.
Text-reading agents
A text-reading agent loads the page and reads its structure: headings, links, buttons, form fields, prices. It often uses the accessibility tree, the same view a screen reader uses.
Strengths: fast, cheap and precise. It can read every price on a room list in one go and spot the cancellation terms under each rate.
Weak points:
- Buttons that only work through scripts, with no visible destination.
- Calendars where days aren’t labelled with full dates.
- Prices drawn as images, or inside widgets it can’t read.
- Content that loads only after scrolling or clicking.
Our pilot of 28 UK hotels used a text-reading scripted browser. Its main problems were hidden booking links and date pickers.
Visual agents
A visual agent, sometimes called computer use, takes screenshots and decides where to click or type. Anthropic, OpenAI and Google all offer versions of this.
Strengths: it sees what a person sees, so it can cope with pop-ups, odd layouts and calendars a text agent can’t read.
Weak points:
- Slow. Each step needs a screenshot and a decision. A long booking journey takes minutes.
- Can misread small text, similar-looking buttons or prices next to each other.
- Can be confused by animations, carousels and overlays that move.
- More expensive to run, so products may limit how long it tries.
Most products mix the two
In practice, consumer agents often read text first and switch to visual steps when they hit something they can’t handle. That means a hotel page that’s readable as text gets the fast path, and a page that isn’t gets the slow, error-prone path, or no path if the agent gives up.
What this means for a hotel
You can’t choose which agent a guest uses. What you can know is whether your booking journey works for both:
- Booking links with real web addresses.
- Prices, room names and policies as text.
- Dates that can be set without a fragile calendar.
- No overlays that block the page.
These also tend to help people using screen readers, so the work is rarely wasted.
What this means for AI recommendations
Text-reading assistants are also how most AI recommendations are made. What your site says in plain text, not just shows in images, decides what AI can repeat about you.
How to check
Try two agents on the same task: a chat assistant that browses, and a browser with an agent built in, such as Gemini in Chrome. Ask each to find a price for a specific date on your own site. Compare where each one slows down. Our free agent check can do this for you.
Questions
Which type of agent do consumers use?
Many products combine both. They read text where they can, because it's faster and cheaper, and fall back to screenshots when a page needs clicking. The exact mix varies by product and changes often.
Which type did Akkrev's pilot use?
A scripted browser that reads pages as text. Visual agents might do better on calendars and worse on speed. We plan to test consumer agents next.
Should a hotel design for one type?
No. A page that is readable as text and works with ordinary clicks suits both.
Are visual agents better at booking?
They can handle pages that text agents can't, but they're slower and can misread screens. Neither type is better everywhere.
Do agents identify themselves?
Some do through their user agent string or newer signing standards. Many browser-based agents look much like ordinary visitors.