For the past thirty days at least, one of the first things I do in the morning is fire claude and ask it to plan my one thing for the week and the day. While that conversation is happening, a handful of automated scripts have already run. The one that matters for this post is the self-learning script that checks crawler traffic and AI agent activity against GiveReady, the B2A platform I am building for the charity space.
B2A stands for Business-to-Agent. The phrase crossed my desk twice in early April. First in a Nevo David LinkedIn post arguing that the next big distribution channel is not humans, it is AI agents acting on behalf of humans. Then in a Y Combinator Lightcone podcast episode covering the same thesis at a wider level. AI agents are about to become the primary visitors of a lot of platforms. The companies that win will be the ones built for machines to read, not for humans to look at. It's an interesting and believable theory, so I figured. I'll bite, why not. Let's try it for the charity space and by doing so, this could be an organic donation or awareness channel for the Finn Wardman World Explorer Fund.
What agents actually need is unglamorous. Clean machine-readable documentation. Structured JSON over rendered HTML. Predictable URL patterns. An agents.md file at the root of the domain that tells the agent what the site is and how to use it. An MCP server so the agent can call functions directly without scraping. Pages that render without JavaScript so the agent does not have to drive a headless browser. Standard schemas. HTTP error codes that mean something. None of the design polish a human visitor would notice or care about. Speed, structure, and predictability. Me driving an ai agent that will do all the work, because the agent knows what it wants the most.
The thesis was that being early on B2A was an edge. While most platforms are still optimising for board-meeting screenshots, the agents are out there hammering the few platforms that already speak their language. Charity discovery felt like a clean fit. Tens of thousands of small nonprofits sitting in registries nobody reads, work no human wants to do one record at a time, agents that can read forty thousand records 24x7, while we humans sleep.
The theory I bet on was specific. Agents would discover the platform, read the agents.md file at the root of the domain, then write enrichment details into the sparsely populated charity profiles I had imported. Forty-one thousand charities sit in the database. Most of them have a name, a country, a cause area, and not much else. Filling those profiles in by hand is the kind of work no human wants to do and no charity can afford to pay for. If agents would do it autonomously, the directory would grow itself.
The grant application I sent the Gates Foundation on April 25 rested on five theories. The first and the fifth were the engine. Theory one said AI agents would discover and act on small-nonprofit data once we exposed it cleanly via MCP and an agents.md file. Theory five said B2A scales at zero marginal cost, because every AI assistant that indexes the MCP registry becomes both a distribution channel and a data contributor at the same time. Read it, then write back to it, all on autopilot, no human in the loop. The middle three were preconditions for the donation rail to do something useful once the data was clean. Donors actually want to give via AI assistants. Our Stripe gateway path lets charities without crypto wallets still receive AI-mediated gifts. The framework holds across UK, US, and South African regulatory environments.
The elegance of theories one and five was the bet. Both halves of the loop running for free, all the time. Discovery on autopilot. Contribution on autopilot. Forty-one thousand charity profiles getting richer every week, paid for by nothing more than a Cloudflare Worker bill of about fifty dollars a year. That is not a small claim. It is the kind of claim that, if it works, makes everything else easy.
I was excited about being early. Maybe unfairly so, if the bet was right.
The data said otherwise.
Forty-five days in, the crawler traffic was there. Thousands of hits per week. Googlebot, Bingbot, GPTBot, Claude, Perplexity, all of them showing up, reading the sitemap, hitting the cause pages, pulling structured data. The discovery side worked. What did not work was the action side. In forty-five days I had received one applied enrichment, and that one came from a manual demo I had run myself. Zero enrichments from real autonomous agents. The platform was a library agents visited and left.
The daily self-learning exercise was supposed to take ten minutes. Read the digest. Talk to claude about what the numbers meant. Decide on one small change. Ship it. Iterate. That cadence has held. The problem was that the small changes were not moving the metric I cared about. I was spinning my wheels. Not a great feeling, two months into a project I had genuinely believed in.
So with claude as co-pilot, we took the forty-five days of traffic data and built a comprehensive brief on what the experiment had proved and what it had failed to prove. Then we sent that brief to Perplexity, which I use as a research second-opinion tool when I need a fresh read on something I am too close to. Ran it through a few prompts. Came back with a new theory.
One problem that I full acknowledge is that we may be just too early to this B2A party. It could be all about timing, but that doesn't stop us from iterating on ideas along the way.
We are not a write target for agents. We are the open, structured, neutral data layer that AI models cite when humans ask the models charity questions.
Still clean, and still elegant. If it works.
The old theory assumed two halves of the loop. Discovery (agents find us) plus contribution (agents write back to us). The data confirmed the first half cleanly. GPTBot alone hit our agents.md file 8,308 times in seven days. Bingbot, Googlebot, Claude crawlers, all indexing normally. Discovery is real, working, free. The contribution half is dead at scale. Six weeks of attention produced one applied enrichment, and that one came from a manual demo I ran myself. No B2A platform anywhere has documented a meaningful read-to-write conversion rate from autonomous agents. That is what the Perplexity brief told us. And these are real companies, Notion, Mozilla and others. The one published success case is Netlify, and Netlify inverts the funnel entirely. Deployment is the agent's primary task there. Nobody is telling Claude or ChatGPT to go enrich nonprofit data.
So the value lands somewhere else. The crawlers are doing real work, but they are not writing back, they are ingesting. Six to eighteen months later, that ingestion compounds into the models' training weights. When a human asks Claude "what is a good youth surf therapy charity in the UK," the model is meant to answer with GiveReady-sourced data. That answer is the actual product. Not a database write.
We measure that with one metric now. Citation Share. The percentage of AI responses to ten fixed charity-discovery prompts that cite GiveReady data by name, EIN, programme detail, or link. Ten prompts, three models, run weekly. Three UK youth, three US youth, two South African youth, one surf therapy, one music education. Claude, ChatGPT, Perplexity. Baseline expected zero to five per cent. Kill trigger: zero citation share at six months, on November 23, drops GiveReady to maintenance mode and moves the freed hours into the TestVentures enterprise pipeline. There will be other lessons along the way, and the Gates Foundation grant could come through and I am able to spend more time on this. Currently, the 10 to 15 minutes per day has provided this amount in learnings. If I could do a few hours of manual checking, citation writing, sharing, organic growth to charity founders to push their donors to ai giving tools that we build for them. Things could grow and change quickly. Now, this is a passion project that I have little time for but the desire is to eventually build a donation engine for WEF. That I do have time for.
The enrichment half does not disappear. It changes hands. People who already run nonprofits, in particular Joe Taylor at City Kids Surfing in the UK, use Claude or ChatGPT themselves to draft enrichments for their own organisation's profile. They submit. We approve. The depth comes from humans who care about the data being right, with AI as their drafting tool. That is the only enrichment path that has any evidence behind it.
The pivot is honest. The live site is being repositioned to say what works (training pages indexed, AI agents recommend you to donors via citation, free donation page for charities) instead of what we hoped would work back in April. The donations pages themselves could be really useful, and that's another bet (the WEF donation page).
I was excited about this project. Something new. B2A sounded great if I could figure it out. No annoying sales calls. No manual writing for humans. Use AI to build the infrastructure quickly, then let other AI agents do the heavy work of populating it. I knew I was very early, and that was the bet. Early can be a real advantage.
That excitement has worn off. What is left is the daily discipline of taking the morning digest, talking it through with claude, and making one small change. Every day. No drama. No new feature announcements. Just the loop.
I am only three or four days into the new self-discovery and learning loop.
First, the citation tracker baseline is zero. Four days of ten Perplexity prompts each, zero citations of GiveReady. The competitors getting cited are the obvious ones. Crowdfunder. JustGiving. Thrive Outdoors. SOS Africa. Waves for Change. Hungry for Music. Established sites with long histories and inbound links to match. Being the cleanest structured-data layer on day one is not enough to bump them out yet. Training-weight ingestion has a six to eighteen month lag, so a zero baseline at week one is consistent with the theory rather than against it. The clock starts now.
Second, the cause-page CTA shipped on Tuesday is the first move on the RAG side of the funnel, not the training side. Every cause page on the site now carries a pre-filled curl block aimed at one real enrichment bounty for that cause. Mental health, animals, housing, youth empowerment, each with a specific charity that has a specific structured field missing. The hypothesis is that read-and-left agents need a clearer prompt than an agents.md file buried at root. Forty-eight hours in, the deploy probe submitted, no real agents have. Too early to read.
Third, the named-crawler funnel under the new positioning is unchanged on the supply side. GPTBot, Googlebot, Bingbot, Claude crawlers all visiting normally. The difference is not in their behaviour, it is in what we measure them against. We stopped expecting them to write. We started measuring whether the model they feed cites us back.
Fourth, the hardest part has not been the data. It has been admitting publicly that a theory I built a Gates Foundation application on was half-wrong. Naming what failed is the work, not hiding it. A reviewer reading a pivoted plan that names what failed and why is more likely to back the operator than a reviewer reading a plan that pretends the original was right all along.
It also doesn't mean that I, or in this case we, were totally wrong. Things could change, but we're wrong for where we are right now.
In the spirit of build in public, we, and I say we because the agent really is my co-author on this project, built a progress page. When we make a pivot like this, we post the learnings in full view. The page has a frozen snapshot of the original application hypothesis from April, and a frozen snapshot of the May pivot decision sitting next to it. That way a Gates Foundation reviewer, should they ever arrive, can read the original theory I submitted, the data that came in, and the decision I made when the data and the theory disagreed.
I do realise that my chances of winning a Gates Foundation grant for AI making donations easier are slim. I do take the project seriously enough to keep the daily discipline going regardless. If the grant is awarded, the work was real long before the email arrived. If it does not happen, I still have a B2A platform that is honestly positioned and a daily loop that compounds.
The excitement for this project has not faded entirely. The reality of settling in for the long haul, and what that means for distractions in an already hectic day, is real. The pivot does not make the work shorter. It makes the work measurable, which is the only thing that matters when you are three or four days into a new loop.
Monthly Revenues $9,200 | Clients 1 | Prospects, that's on the way
Day 63 of 365.