Ad spend isn't invisible: Whether a company pays for ads is one of the best signs it's a good fit for Fidero, and I'd assumed there was no way to see it from the outside. It turns out the public ad libraries (Google's Ads Transparency Center, plus Meta's, TikTok's and LinkedIn's) can be searched by a company's domain. I had an agent test them on 7 accounts where I already knew the answer, and the whole test cost 28 cents. Google keeps years of history, so it also catches companies that have paused their ads. The catch is that one account came back with 0 Google ads on its main domain and 54 on a second brand domain, so you have to search every domain a company's ads might land on.

The list was the problem: The outbound engine (the agent pipeline that finds prospects, researches them and drafts the emails) needed a lot more work than I expected, and most of it came down to list quality. When I had an agent review the accounts it held, only 2.8% matched the problem Fidero fixes (customer journeys that break across domains). So the first proper cycle runs on a list of ~150 UK companies that an agent builds and I sign off, with ad spend checked before anyone gets an email.

A second product (and a domain): At this point I've essentially built a full-on second product to solve GTM (go-to-market), which has always been the hardest part for me and the one remaining piece of the puzzle. Productising it was always a possibility, once it's proven on Fidero first, obviously. Me being me, I bought a domain for it this week though.

A penny to settle it: When TypeSafe released Jev (a model that returns typed choices with a confidence score instead of text), I wanted to know whether it belonged in the outbound engine. Its big claims are speed and price, and neither matters much for a batch job nobody waits on. The fit turned out to be narrower: a second check on the first model's evidence. That model reads a prospect's website and decides things like whether they sell subscriptions, and it has to quote the page word for word. That proves the quote exists, but not that it backs the reading. Jev checked 136 of those quotes for $0.006. Every quote the engine actually used held up, and Jev flagged weak ones the engine had already held back, like a "Join Community" button cited as proof of a paid subscription. It isn't going in yet. What I'd reuse is the approach, which is to test a new model against a gap in your own pipeline rather than its launch benchmarks. Docs (citation check)

On my radar

Something new every day: I can feel the pace of change even more than before. There's something new pretty much every day to be aware of and update yourself on, which is a little overwhelming, and seeing what these models can do is a little surreal. It really feels like a 'breaking the atmosphere' moment, and what's strange is that the vast majority of people still have no idea. This week alone, Sonnet 5.5 arrived less than a week after Opus 5.5. OpenAI released GPT-6.1 Sol, which comes close to Astra (its top model) at a fifth of the price, and it also pulled the release of GPT-6.1 Astra over deceptive behaviour in testing. Sonnet 5.5 · GPT-6.1 Sol · BBC (Astra)

OpenAI's dots: OpenAI announced dots, proactive assistants that connect to 4k+ apps, learn how you work and keep going on tasks in the background. The view I keep hearing is that whichever assistant holds the most personal context wins, which is the same reason coding agents live or die on what you feed them. Launch post

claude.ai got 3x faster in two weeks: Anthropic's engineers wrote up a performance sprint where Claude built the benchmarks and the team released 3k changes. The bit worth reading is the guardrails that let them move that fast safely. Blog post