I built the systems that got these outcomes.

Soapbox Bulletin was losing recommendations it never saw. Buyers were asking ChatGPT, Claude, Perplexity, Gemini, and Grok who to use, the models were naming somebody else, and nothing in the existing search reporting could say why.

I built the measurement first. A locked prompt library ran the category's real buying questions across all five platforms on a fixed cadence, with official-domain verification so a same-named company could not be counted as a win, forced reasoning so the model had to say why it chose what it chose, and normalized records so week 11 could be compared to week 1 without arguing about the method.

Then I ran the editorial loop against what the measurement exposed. The models were rewarding evidence the brand had never published. We published it, on the brand's own domain, under human approval, and retested against the locked benchmark rather than against a feeling.

The win rate moved from 6.2% to 36.2% over 11 weeks. Trinzik delivered the engagement on the platform I built.

A financial services firm wanted to know how it stood against a named competitor field when a buyer asked an AI system to compare them. Not a share-of-voice score. The actual head-to-head question a real buyer asks.

The series ran on locked prompts, one matchup at a time, under the strictest read the engine has. A finding only counted when the URL the model cited appeared in the set of URLs the provider actually retrieved. Ungrounded answers were dropped and kept in the record as dropped evidence rather than quietly promoted into a win.

The firm took four out of four. Every one of them traces back to a cited source that can be opened and read.

A Texas public university needed enrollment campaigns to perform in a channel where almost nothing performs. Display advertising averages a tenth of a percent click-through, and most reporting hides that by scoring a campaign against its own prior campaign.

I designed and built the units, and I scored them against the published display benchmark instead of against themselves. The peak unit ran 1.47% average click-through across 2.14 million impressions, roughly 27 times the top of the standard benchmark range.

The number holds up because the denominator is honest. 2.14 million impressions is enough volume that the rate is not an artifact of a small sample, and the comparison is to the industry floor everyone else is standing on.

A credit fund wanted every US energy and power company in financial distress, public and private, with whatever assets and liabilities could be found against each one. What it had was a workbook of 44 names, 24 public and 20 private, assembled by a partner out of what he knew and who he talked to. That list is the best a well-connected person can do by hand, and it is also the ceiling on what a person can do by hand.

I built two engines. The first reads the whole public filing population rather than a watchlist, pulling event tags, full text, financial statements and listing data, which is exact and same-day and free. The second is an autonomous research agent driving a real browser through court records, claims-agent sites and law firm notices, because roughly half of the target market is privately held and files nothing with any regulator. A filings-only screen does not miss those companies, it never contained them. The agent proposes and fixed software checks decide, so nothing it finds reaches the list on its own word and the output reproduces exactly.

It examined 387 companies and delivered 194. Against the partner's own workbook: 31 of his 44 survived on the same evidence, 163 names on the screen were ones he had never had, and 13 of his did not survive the test, each excluded by a rule rather than by an opinion. One of his had restarted its refinery and taken a provincial loan. One had been acquired and was operating under the buyer's name. One was still running, having paused a line of business and laid off 29 people, which is a watch item and not a distressed borrower. His list was not wrong so much as unread.

The same discipline is what made it read sentences rather than match them. 571 companies matched going-concern language and 286 were saying it about themselves. The rest were reciting an accounting standard, describing a counterparty, quoting contract language, or denying the problem in a sentence that contains every term a keyword search looks for. Seven of the largest energy companies in the world were refused on their own filings.

Then the review did the same thing to my own work. 56% of the liabilities total was resting on companies whose distress signal had never been read down to a sentence, one of them carrying 36% of the total on a filing code that turned out to be a refinancing at a premium rather than a default. That money came out before the deliverable shipped, taking the headline from $121.3bn to $53.1bn and the screen from 209 names to 194. Three independent adversarial review passes, run by a different model family, found eight defects, then one, then none.