Agent View
The site as an AI agent gets it, beside the site as a person sees it. What it can read, what it has to cut away, and what it never sees.
HomeAboutTeamFounder's LetterJoin our mailing listDonate
Latest: AI Rejected Your Job Application. Should You Know Why?
No Hype. Just Proof.
Proof is investigating the social impacts of AI. Join our mailing list for updates.
Explore AI and Labor, Video, AI and Health, AI Training Data, The AI Democracy Projects, Notebook, or see what's new.
What's new
AI Rejected Your Job Application. Should You Know Why?
Companies use AI tools to vet applications, but job seekers say they have little insight into how the tech works. A lawsuit could offer a glimpse.
By Varsha Bansal Oct 15, 2026
Image: AI Rejected Your Job Application. Should You Know Why?
The rest of the page is in full screen.
[ The Roboguard Revolution is Short-Circuiting Knightscope and other robotics companies are rethinking automated security amid canceled contracts. One pivot? Human guards. By Rebecca Plevin, Varsha Bansal Aug 10, 2026](https://www.proofnews.org/the-roboguard-revolution-is-short-circuiting/) [ Meet the Robot That Nurses Unplugged Spinning in circles and crying for help, Moxi robots became another charge to care for, nurses said. By Varsha Bansal Jun 9, 2026](https://www.proofnews.org/moxi/) [ More Stress, Fewer Breaks: Hotel Housekeepers Reveal What It's Like Working for an App Promising productivity, HotSOS Housekeeping tells workers when and where to clean and is now used by most global hotel brands. By Rebecca Plevin May 19, 2026](https://www.proofnews.org/hotel-housekeepers/)
Thin: Too few words of their own are in the HTML for a page of their kind. Render the rest on the server. No h1: Give each one h1 that says what the page is.
The figures
100 pages, set against the 168,376 sites with figures.
- Readablehigher is better
- 77%77 of 100 pages
- All 168,376 sitesBetter than 23%
Pages whose own text is in the HTML the server sends, enough of it for their kind.
- Cleanhigher is better
- 72%middle page
- All 168,376 sitesBetter than 52%
Of the words before script, the share that are the page and not menus, footer and forms. 2 pages are not measured, with no <main> or no words.
- Costshown, never scored
- 145,725tokens
- About 1,457 a page
To read every page once. A long page that answers well is not worse for it. Tokens are estimated at about 4 characters each.
How its pages spread on Clean
98 pages measured. Not measured: 2 shells with no words before script.
Against its rivals
The sites nearest it by what they do, with figures. Close means similar work, not better.
| Readinghigher is better | Costnot scored | ||
|---|---|---|---|
| Site | Readable | Clean | Tokens a page |
| proofnews.orgThis site, 100 pages | 77% | 72% | 1,457 |
| platformer.news100 pages | 61% | 90% | 3,500 |
| themarkup.org100 pages | 99% | 77% | 3,096 |
| 404media.co98 pages | 90% | 82% | 2,149 |
Third of 4 on Readable, fourth on Clean.
AI crawler access
What robots.txt says to each, as read on October 24, 2026. 8 of 8 may read the site.
| Crawler | robots.txt says | Used for | The line that decides |
|---|---|---|---|
| GPTBotOpenAI | Not named | Training models | Falls to User-agent: * |
| OAI-SearchBotOpenAI | Not named | ChatGPT search | Falls to User-agent: * |
| ChatGPT-UserOpenAI | Not named | A page a person asks about | Falls to User-agent: * |
| ClaudeBotAnthropic | Not named | Training models | Falls to User-agent: * |
| Claude-SearchBotAnthropic | Not named | Claude’s search | Falls to User-agent: * |
| PerplexityBotPerplexity | Not named | Perplexity search | Falls to User-agent: * |
| Google-ExtendedGoogle | Not named | Training Gemini | Falls to User-agent: * |
| bingbotMicrosoft | Not named | Bing and Copilot | Falls to User-agent: * |
Every AI crawler mttr checks may read it. Not named means the rules for every crawler apply.
Every page
Worst first. Each opens in Agent View at the top.
26 pages of 100 lose something. Pick a loss to see what to do about it.Open the action plan
Show where it stands
A card, a badge and a line, each with the figures.
A badge for the site
<a href="https://mttr.si/agent-view/proofnews.org"><img src="https://mttr.si/agent-view/proofnews.org/badge.svg" alt="Agent View: 77% readable, 72% clean" height="20"></a>
The figures in one line
proofnews.org in Agent View: 77% of pages readable, better than 23% of sites; 72% clean; 145,725 tokens to read all 100 pages. mttr.si/agent-view/proofnews.org
What to do next
Fix the pages, let an agent do it, and read the site again to see the figures move.
Fix what agents miss
- Render the rest of 21 pages on the serverReadable
- Give 4 pages a main headingClearer
- Put the text of 2 pages in the HTMLReadable
Connect your agent
Claude Code, Codex or Cursor reads the site’s fixes over MCP, fixes a page in your code, has mttr read it again, and checks the fix landed.
https://mttr.si/mcpSet up an agent
Claim the site
Prove it is yours with a DNS record or a file. Your agent can then reach its reads and jobs, and set each change to wait for your approval.
Claim the siteDrawn from the HTML mttr fetched on its read of October 24, 2026, not a screenshot, and only from the 100 pages read, where the read stopped at its page limit. An assistant’s own reader may differ. Tokens are estimated at about 4 characters each. A site an agent reads well is not a site an assistant cites.