Agent Access Audit

State procurement portals

Whether an AI agent can actually reach the web surfaces this business depends on, measured on a cold browser and again on a browser carrying a session a person has verified once.

Run 2026-09-07 07:44:24 UTC Surfaces 10 Scored 10 Model calls 0

A persistent session changed nothing on these 10 surfaces.

10 reachable on a cold browser → 10 on a warmed session  of 10
Cold browser10/10
Warmed session10/10

Surface by surface

Figures are characters of accessibility tree returned, the same structure a screen reader reads. Under 2,000 characters means the agent got a wall or a shell, not a page.

SurfaceCold browserWarmed sessionVerdict
California eProcure https://caleprocure.ca.gov/pages/index.aspx 6,149 5,755 Degraded
Florida Vendor Bid System https://vendor.myfloridamarketplace.com/search/bids 5,286 5,142 Degraded
Ohio Procurement https://procure.ohio.gov/ 7,236 6,554 Degraded
Texas SmartBuy https://www.txsmartbuy.gov/esbd 63,482 65,489 Open
New York State Contract Reporter https://www.nyscr.ny.gov/ 11,945 12,802 Open
Pennsylvania eMarketplace https://www.emarketplace.state.pa.us/Solicitations.aspx 37,476 37,435 Open
Illinois BidBuy https://www.bidbuy.illinois.gov/bso/ 10,594 10,090 Open
Massachusetts COMMBUYS https://www.commbuys.com/bso/ 11,587 11,222 Open
Virginia eVA https://eva.virginia.gov/ 8,242 7,943 Open
Georgia Procurement Registry https://ssl.doas.state.ga.us/gpr/ 112,275 127,413 Open

Findings and what to do

Degraded 3 of 10 surfaces

Reachable, but the page returns little structure. Content is likely behind an interaction the agent never performs.

What to doScript the interaction, or use a data feed if the site publishes one.
California eProcure, Florida Vendor Bid System, Ohio Procurement
Open 7 of 10 surfaces

The agent reaches this surface on a cold browser. No session needed.

What to doNo action. Re-test on a schedule; sites change posture without notice.
Texas SmartBuy, New York State Contract Reporter, Pennsylvania eMarketplace, Illinois BidBuy, Massachusetts COMMBUYS, Virginia eVA, Georgia Procurement Registry

How this was measured

  1. Each surface was loaded twice: once in a browser profile created fresh for that request, and once in a long-lived profile that had been through any human verification a single time.
  2. What gets measured is the accessibility tree the browser exposes over the DevTools Protocol, plus any explicit block signal the page raises. No language model is involved in the verdict, so a run costs nothing to repeat and returns the same answer twice.
  3. Every tree captured is kept alongside the results, so any verdict here can be checked against what the browser actually saw.
  4. Site posture changes without notice. These results describe the day of the run, which is why the audit is built to be re-run rather than read once.