What we proposed in the Gates Foundation application, what the data has shown, and how the approach has evolved. No spin. The numbers are what they are. The snapshots below let any reviewer compare what we claimed against what we now claim, in our own words on both sides.
The Gates application proposed AI agents would autonomously enrich the nonprofit directory by writing data back through `/api/enrich`. Four to six weeks of live traffic showed otherwise. Training crawlers (GPTBot, ClaudeBot, Google AI) ingest the directory at high volume, but they cannot submit by design. Interactive agents (ChatGPT-User, Claude-User, Perplexity-User) rarely arrive at the site, and when they do, they read and leave. The autonomous-write thesis is dead at scale.
What does work, on today's evidence: training-corpus ingestion is real and high-volume, so the path to AI-mediated giving is being the structured-data source that models cite when users ask charity questions. The path to verified depth is human-in-loop enrichment via consortium partners (people who run the orgs use AI to draft their own profile updates, then submit). Both are measurable. Neither requires the autonomous-write magic.
So the approach has evolved. The directory is still live, the donate widget still ships zero-fee donations, the MCP server still indexes for every AI assistant that follows the standards. The metrics below now lead with what's measurable today, not what was aspirational at submission.
Each snapshot is a frozen point-in-time record. None has been edited since the date on its banner. New snapshots are added on three triggers: a pivot decision, a Gates milestone reached, or a quarterly review. Reverse chronological below.
Open structured data layer indexed by AI training crawlers and search crawlers. When models are trained on the web, GiveReady's nonprofit profiles are in the corpus. When users ask charity questions, GiveReady-sourced data appears in the answers. We measure this directly via the Citation Share metric below.
Human-in-loop. Consortium partners (Joe Taylor at City Kids Surfing, Bridges for Music, Wave Project, Finn WEF) use AI assistants to draft enrichments for their own profiles, then submit. The "AI does the research, the human signs the submission" pattern that every B2A platform with documented writes (Notion MCP, Linear MCP, Stripe Agent Toolkit) has converged on.
Weekly: 10 fixed prompts run against three models. Perplexity returns source URLs natively, so that arm is automated. Claude and ChatGPT return text without source attribution, so those arms are manual spot-checks logged into the tracker. Prompts cover three UK youth charities, three US youth charities, two SA youth charities, one surf therapy, one music education. Tracker output lives at 01-Projects/GiveReady/citation-tracking/YYYY-MM-DD.md in the project repo. Methodology stays public. If Perplexity changes how it returns source attribution, the tracker falls back to manual for all three models.
Three workstreams are on the shelf until the citation-share signal moves or the Gates decision lands: a sandbox /api/enrich endpoint with anonymous-write tier (Netlify-model attempt at the original autonomous-write thesis, time-boxed at 14 days if reactivated); a 50-org verticalised "deep on few" pilot in one Gates GH&D-aligned cause area; a Gates honest-update milestones doc to send when Gates emails for a check-in. Full list in the project repo's TODOS.md.