We found 3 issuescosting you truston emergentmind.com.
The top paper's summary has a nonsense bullet and a misnamed GRPO link.
Watch the inspection.
We navigate your site the way customers would. Idle time is skipped.
How the composite score works
The overall score is a weighted average of the categories: Getting the product 40%, Trust before buying 20%, Mobile 15%, Speed 15%, and Polish and audit 10%. It is never more than 10 points above the Getting the product score, because a product a visitor cannot get cannot score well on anything else, and never more than 25 points above the weakest category, so one badly broken area is not averaged away. Any issue in the main list caps it at 89, so a score above 90 means nothing worth fixing was found.
Each category starts at 100 and stays there when nothing was found in it, whether or not the inspection got that far. Each distinct verified finding deducts 60 points for a serious problem on the path from arriving to paying, 40 for another serious problem, 30, 15 or 8 by high, medium or low severity otherwise, and 4 for an audit finding. Each additional finding below the list takes 2 more points off the overall score, up to 10. Scores stop at zero and round to whole numbers. The formula is fixed and versioned, so the same saved evidence always receives the same score.
The top paper's summary has a nonsense bullet and a misnamed GRPO link
Researchers who judge the paid AI summaries by the free summary of the #1 trending paper read a sentence that makes no sense and a link to a made-up algorithm name.
On /papers/2610.13025 the third Summary bullet reads "Email the cost-effectiveness of harness-evolution experience and exposes deployed minimised cost." The "GRPO" link in the detailed summary opens /topics/grouped-rolie-ppo-grpo, titled "Grouped-ROLIE PPO (GRPO) Overview", and that same page says the method is "often under the name Group Relative Policy Optimization." Visitors trying the product meet these errors in the free summary they use to decide whether the paid summaries are accurate.
Try it: Open /papers/2610.13025 and read the third bullet in the Summary box. Then click "GRPO" in the detailed summary and read the title of the topic page it opens.
The topic-page "Papers" link opens the AI Product Lab home
Researchers who arrive on a topic page and click "Papers" for more research reach a project hub, not the paper search.
On topic pages such as /topics/grouped-rolie-ppo-grpo, the top nav still reads "Papers / Open Problems / API / Labs / Pricing". "Papers" links to the root URL, which now opens "Emergent Mind | AI Product Lab" with featured project cards and no paper list or search box. The trending papers and search live at /assistant, so a researcher who came in from search and clicks "Papers" has to find the AI Research Assistant card to continue.
Try it: Open https://www.emergentmind.com/topics/grouped-rolie-ppo-grpo and click "Papers" in the top nav. The AI Product Lab home opens, not the paper list.
The /assistant page shifts its content while it loads
New visitors trying search or the trending papers list on desktop see the page move under them before it settles.
On desktop, /assistant has a cumulative layout shift of 0.276, and Google classes anything above 0.25 as poor. The largest shift is in the homepage section that holds the trending papers list, below the search box, and it was measured before any scrolling or clicking. This is the page where a visitor first tries the product, so the list they want to read or click moves while it loads.
The top trending paper shows raw LaTeX and broken superscript markup
The first paper under Trending Papers is listed as "Evo$^2$: Evolving the Evolver through Weight-Harness Co-Evolution", with the LaTeX dollar signs and caret left raw. On /papers/2610.13025 the heading correctly shows "Evo²", but the abstract typesets a broken superscript tag as math three times and displays "Evo< sup > 2". Visitors trying the product meet this on the paper it promotes first, where they are deciding whether its summaries can be trusted.
Google's sign-in prompt covers focused paper links on phones
On /assistant at phone width, the fixed Google sign-in iframe (credential_picker_iframe) sits over the bottom of the screen. When the trending paper link "RSI-Anything: Unlocking Recursive Self-Improvement with World Models" has focus, every point tested inside it lands on that iframe instead. The prompt can be closed with its X button, and after that the links are reachable again.
The home page head has no twitter:card tag
The home page head at / declares no twitter:card meta tag. Its og:title, og:description and og:image are present, so X falls back to those values, but it may show a small summary card instead of a large image preview.
The home page lacks CSP, frame protection and CORP headers
The response from https://www.emergentmind.com/ is missing Content-Security-Policy, CSP frame-ancestors or X-Frame-Options, and Cross-Origin-Resource-Policy.
A button on /assistant has no accessible name
It appears to be the icon-only arrow button under the search box. Assistive technology announces it only as "button".
Section counts and card dates fall below the contrast minimum
On the home page and /assistant, the grey section count "6 items" measures 4.36:1 against its background, and card dates such as "since 2022" measure 4.04:1. WCAG AA requires 4.5:1 for text this size. The card titles and descriptions next to them stay readable.
Testimonial author links look the same as the quote text
On /assistant, the testimonial author links (for example "AI Explained" and "AI Educator") are #2700a8 inside #302a21 body text. That is a 1.11:1 contrast with the surrounding text, and the links have no underline. Visitors who cannot rely on colour cannot pick out these links from the surrounding testimonial text.
Probie fixes this for you.
Stop customers hitting the same friction. Connect your codebase and Probie reproduces these failures, tests the fix, and opens a pull request for your review.
Sign up and get this fixedFree inspection
What should Probie inspect next?
Enter a project URL. There is nothing to install.
Probie checks every issue twice, respects robots.txt, and backs off at bot challenges.
Enter an address and we will email the report card link.