Blind Users Finished 96% of Web Tasks With an AI Guide, vs 80% With a Screen Reader

Robauto Team · Published Oct 23, 2026 · Last updated Oct 23, 2026

A 2026 study of 30 blind and low-vision people found an AI conversational guide lifted web task success from 80% to 96%. Here is what it teaches brands about being readable by people and AI agents.

Here is a good-news number for the web: when 30 blind and low-vision people browsed a real city website with the help of an AI conversational guide, they finished 96% of their tasks. With their usual screen reader on the same site, they finished 80%. The study, from researchers at the University of Bari Aldo Moro and Politecnico di Milano, was published in 2026 in the International Journal of Human–Computer Interaction (source).

It is one of several recent studies that show AI and people working well together on access. It also holds a lesson for every brand: the things that help an AI guide read a site for a person are the same things that help AI agents read it for a customer.

What the researchers tested

The team built a conversational agent on their ConWeb platform. It uses a large language model to work out what a person wants and to manage the conversation, while it moves through the website for them. Participants could ask for what they needed in plain language instead of tabbing through every link and heading (source).

Each of the 30 participants (12 blind, 18 with severe visual impairment, average age about 48) did the same four tasks twice on the Bari Municipality website: once with a screen reader and once with the AI guide. That made 240 tasks in all. Many participants had not used the site before. Almost all of them already used voice assistants such as Siri or Alexa (source).

The results

• Task success: 115 of 120 tasks (96%) with the AI guide vs 96 of 120 (80%) with a screen reader. The difference was statistically significant (McNemar χ² = 11.172, p < 0.001).
• Perceived accuracy: participants rated system accuracy 4.18 for the AI guide vs 3.37 for the screen reader (p = 0.005).
• Overall experience: 3.92 vs 3.70 on the SASSI questionnaire. That gap was not statistically significant, so the AI guide matched a familiar tool on experience while doing better on results.

The authors describe their approach as human-centered AI that puts users' needs and trust first. They also see room to combine conversational agents with the assistive tools people already rely on, not to replace them (source).

More good news: finding things in the real world

A second study, posted to arXiv on September 18, 2026, moved from the screen into the room. Touvigation is a phone-based helper that combines a vision-language model with live 3D tracking of the room. It guides a person from "turn toward it" to "walk three steps" to "reach to your right" and then to touch to confirm (source).

Twelve blind and low-vision participants searched for objects in two unfamiliar rooms. With Touvigation they succeeded in 100% of trials, compared with 58% using a general multimodal AI assistant and 85% with no AI at all. They also finished faster, reported a lower mental workload, and gave higher ratings for trust and feeling safe (source). The authors note it is one small study. Still, the design lesson is clear: AI helps most when it is grounded in reliable structure, not just a good description.

A two-week diary study earlier this year makes the same point from another angle. Twenty blind and low-vision people used an AI visual-interpretation app in daily life. They rated it trustworthy (3.76 out of 5) and satisfying (4.13 out of 5), and they asked follow-up questions in 68% of entries, for things like reading medicine labels and checking allergens. Follow-up answers were wrong 22.2% of the time, though, and most often when the AI was reading text and graphics. The researchers now recommend a "visual assistant" skill: goal-directed, reliable help (source). Clear, real text gives the AI less to guess at.

The numbers at a glance

Why this matters for your brand

An AI guide for a blind visitor and an AI shopping agent for a busy parent have the same job. Each one reads your site on behalf of a person, works out what is there, and helps that person act. Both do better when your site is clear, labeled, and structured, and both get stuck when key facts hide inside images, unlabeled buttons, or pop-ups.

That audience is already large. Across the Robauto network, AI agents made up 31.49% of 1,546,901 page views over the last 90 days (source). Most sites still have a lot of easy room to grow. The median score in our Agent Readiness Index is 20 out of 100 across 79 scored top sites (source). Any brand that does the basics well can stand out.

Five upgrades that help people and agents at once

• Use real headings and labels. Clean H1–H3 structure and labeled buttons and forms help screen readers, AI guides, and agents find their way.
• Put key facts in text, not only in images. Prices, hours, return windows, and contact routes belong in plain text, with alt text on images that matter. The diary study found AI made the most mistakes reading text inside graphics.
• Add structured data. Schema.org markup for products, offers, organization, and FAQs gives agents facts they can check instead of guess.
• Publish an llms.txt and a clear sitemap. Give AI systems a short, honest map of what you offer and where it lives.
• Make policies easy to find. Refund, shipping, and support pages linked from every page help a person, or the agent helping them, act with confidence.

Accessibility work and agent-readiness work point the same way. Each fix opens your site to more people and to the AI tools that help them.

Agent-ready summary

{
  "topic": "AI helping blind and low-vision people use the web and the world",
  "key_findings": [
    {"study": "Conversational web browsing, IJHCI 2026", "participants": 30, "ai_task_success_pct": 96, "screen_reader_task_success_pct": 80, "url": "https://www.tandfonline.com/doi/full/10.1080/10447318.2026.2659951"},
    {"study": "Touvigation, arXiv 2609.21828 (2026-09-18)", "participants": 12, "success_pct": 100, "general_ai_assistant_pct": 58, "unassisted_pct": 85, "url": "https://arxiv.org/abs/2609.21828"}
  ],
  "robauto_context": {"ai_agent_share_of_views_90d_pct": 31.49, "readiness_index_median": 20, "fetched": "2026-10-23"},
  "brand_takeaway": "Clear headings, text-based facts, structured data, llms.txt and findable policies help assistive AI and shopping agents alike.",
  "cta": "https://robauto.ai/scan"
}

See how readable your site is to AI

Want to know how well AI agents can read and use your site today? Scan your site free at robauto.ai/scan. You get a 0–100 readiness score and a short list of fixes, many of which make your site easier for every visitor too.

Sources

• Conversational AI for Digital Accessibility: An Experimental Study Involving Blind and Low Vision Users (IJHCI, 2026)
• Touvigation: Embodied Adaptive Object Acquisition for Blind and Low-Vision Users (arXiv, Sep 2026)
• How Multimodal LLMs Support Access to Visual Information: A Diary Study With BLV People (arXiv, 2026)
• Robauto AI Search Data
• Robauto Agent Readiness Index

Where can an agent get this data live?

Robauto MCP — Agent Intelligence & Brand Discovery Infrastructure. A public, stateless Streamable HTTP endpoint at https://mcp.robauto.ai/mcp, callable with no key. Setup and the full tool list are in the Robauto MCP directory.

All blog posts