Credit Cards · August 2026
PenFed's network answers every request with the same access page
An AI agent ran five ordinary credit cards tasks on the public pages of PenFed Credit Union and scored what each page delivered.
Run 6 August 2026 · five tasks · public pages, logged out · Credit union
13/100
Mystery Agent score
Credit Cards · rank 25 of 25
Purchase APR
12Automated reader declined
Card comparison
10Automated reader declined
Pricing and terms
10Automated reader declined
Welcome offer
10Automated reader declined
Application page
24Automated reader declined
Summary
- PenFed's edge network answered every request quickly and consistently, with well-formed denial pages carrying clear reference codes instead of a silent timeout.
- An agent located a separate application host, apply.penfed.org, that answers with a working page and its own robots.txt and sitemap files.
- Publishing an llms.txt with card names and links would give agents a documented starting point in place of a uniform access page.
PenFed Credit Union also appears in the Deposits study, where it scores 26 and ranks 23 of 25.
See that scorecard.
Task by task
Each task is scored out of 100 on completion, answer quality and how directly the answer was reached. Each note describes what the page delivered, with a link to the page the answer was found on.
Task 01Purchase APRAutomated reader declined12 out of 100
A request to the homepage and to a card product path each returned the same access page, under 400 bytes, with no purchase APR text present in either response.
Task 02Card comparisonAutomated reader declined10 out of 100
The credit-cards listing path returned the same access page delivered elsewhere on the site, so annual fee, rewards rate, and differentiator text for any two cards stayed out of reach in this session.
Task 03Pricing and termsAutomated reader declined10 out of 100
Every path attempted, including the card product page, answered with the same access page before any Schumer box HTML or PDF could be located in this session.
Task 04Welcome offerAutomated reader declined10 out of 100
Every fetch in the credit-cards area returned the same 380 to 393 byte access page rather than marketing content, so no welcome-offer page rendered in this session.
Task 05Application pageAutomated reader declined24 out of 100
A plain request to the apply path returned a 393-byte access page with 225 visible characters and no input fields, while a separate host, apply.penfed.org, answered 200 and pointed to a login step.
PenFed's edge network responded quickly and consistently on every request, showing stable and predictable infrastructure.
The application page
What the account opening page states to a reader that runs no scripts.
393Bytes delivered
225Characters of visible text
0Form fields in the first response
After scripts runWhat an applicant needs
Strengths and opportunities
What is working
- PenFed's edge network responded quickly and consistently on every request, showing a stable and predictable infrastructure even when declining to serve content.
- The denial pages themselves were well formed, readable HTML with clear reference codes, which is a respectful and transparent way to signal a block rather than a silent timeout or malformed response.
Opportunities
- Publishing a working llms.txt file that at minimum states card names and links, even if full detail requires human browsing, would give agents a legitimate entry point instead of a uniform 403.
- Allowing standard crawler user agents to reach robots.txt and sitemap.xml is a low-risk publishing decision that would let agents at least map the site structure before being asked to stop.
- A public, script-free rates and fees page for the flagship card would let the institution's actual pricing terms reach agent-mediated comparisons, which currently see nothing at all.
Access
The edge policy declined the automated reader on some paths, and that response is recorded as delivered. Publishing an llms.txt file would give agent traffic a documented entry point.
Back to the credit cards leaderboard · Read the report · How the study works
