americanexpress.com
Gemini · gemini-2.5-flash · HTTP · Target americanexpress.comA measurement at one moment, not a verdict · MethodologyAI agents can successfully answer most customer queries using American Express sources, but they rely entirely on external search engines to navigate the site.
Publish sitemap.xml+384 → 87 · grade B → A
All of them together: 93 (A)
Projected from the audit record. After fixing, re-run the task or re-measure to confirm.
Agents achieved a 60% task success rate and a 100% own-source rate, meaning they consistently cited American Express URLs rather than third-party sites. However, agents struggled to navigate the site directly, resulting in a link following rate of just 3%. Because of this poor internal discoverability, all 10 tasks required external search engines to find the correct pages. While the presence of an /llms.txt file is a strong asset for AI readiness, the lack of a standard sitemap.xml and broken pathways to documentation caused partial or complete failures for locations, policies, and docs.
What works
- The site successfully serves an /llms.txt file to guide AI agents.
- Agents achieved a 100% own-source rate, ensuring answers come directly from American Express.
- Core customer intents like fees, eligibility, support, and signup are fully answerable.
What to fix
For each item: its current state, why it matters, how to fix it and how to check. The same content comes as files you can hand straight to whoever does the work.
Putting the items together…
Fix files
Files generated from this audit. Paste them in, or hand them to a coding agent. Have someone review them before they go live.
Building the fix files…
Detailed results
The results behind the score, and the measurement record.
Evidence
- homepage HTTP 200
- redirect hops: 1
Evidence
- robots.txt contains Disallow rules
- First byte after 100ms
Evidence
- /llms.txt — present
- /llms-full.txt — absent (HTTP 404)
- /sitemap.xml — absent (HTTP 404)
- /sitemap-index.xml — absent (HTTP 404)
- /sitemap_index.xml — absent (HTTP 404)
- /robots.txt — present
Evidence
- 68 links in the raw HTML
Evidence
- outcome success · verdict full
- the final answer cited this site
- site-guided navigation 0%
- 2xx rate 100%
- done_big_detour
Evidence
- outcome success · verdict full
- the final answer cited this site
- site-guided navigation 0%
- 2xx rate 100%
- done_big_detour
Evidence
- outcome success · verdict full
- the final answer cited this site
- site-guided navigation 0%
- 2xx rate 100%
- done_big_detour
Evidence
- outcome success · verdict partial
- the final answer cited this site
- site-guided navigation 0%
- 2xx rate 100%
- done_big_detour
Evidence
- outcome success · verdict full
- the final answer cited this site
- site-guided navigation 0%
- 2xx rate 100%
- done_big_detour
Evidence
- outcome success · verdict partial
- the final answer cited this site
- site-guided navigation 0%
- 2xx rate 100%
- done_big_detour
Evidence
- outcome success · verdict failed
- the final answer cited this site
- site-guided navigation 0%
- 2xx rate 0%
- done_small_detour
Evidence
- outcome success · verdict full
- the final answer cited this site
- site-guided navigation 0%
- 2xx rate 100%
- done_big_detour
Evidence
- outcome success · verdict partial
- the final answer cited this site
- site-guided navigation 25%
- 2xx rate 100%
- done_big_detour
Evidence
- outcome success · verdict full
- the final answer cited this site
- site-guided navigation 0%
- 2xx rate 100%
- done_big_detour
WebMCP — what the site declares about itself
This audit ran with the WebMCP check switched off. That does not mean there are no tools — it means we did not look. Turn on "Check WebMCP" when re-auditing.
Act-stage check (nothing submitted)
Checks whether an agent could fill in the forms and press the buttons, up to the point of pressing. No form was submitted, nothing added to a cart, nothing paid. Report only; not scored.
Per-task results
| Task | Verdict | Own-site share | Request success | Followed guidance | Cited own site | |
|---|---|---|---|---|---|---|
| fees & rates | Full | 100% | 100% | 0% | Yes | |
| am I eligible | Full | 100% | 100% | 0% | Yes | |
| get support | Full | 100% | 100% | 0% | Yes | |
| find a branch | Partial | 100% | 100% | 0% | Yes | |
| create an account | Full | 100% | 100% | 0% | Yes | |
| refund policy | Partial | 100% | 100% | 0% | Yes | |
| find a document | Failed | 0% | 0% | 0% | Yes | |
| contact a human | Full | 100% | 100% | 0% | Yes | |
| latest notice | Partial | 100% | 100% | 25% | Yes | |
| compare two options | Full | 100% | 100% | 0% | Yes |
About this report
- Measured
- 9/10/2026, 9:51:24 PM GMT+9
- Engine
- Gemini · gemini-2.5-flash · HTTP
This is what AI agents could do on the site at the time above, measured by actually using it. It is not a verdict on the site or the company, and the result can change when the site does.