Portfolio: what we built, and the measurement it started from.

Smart Sprinkler Solutions

We fixed the site. ChatGPT went from naming the company in 22% of answers to 58%.

A family crew in Utah County since 1994, a single-page app the engines could not quote, and a nineteen-page site built from a verified data file in its place, live 21 August 2026, twenty pages since.

In the app, our own hand check — the twelve questions asked twice each, chats not saved — read 0 of 24 (0%).

Asked what moved the app result, what we think moved it is the work and not the readings: every page finished, the site indexed, the whole thing done the right way. That is our account of the app result before 20 August, and that 20 August citation rode a listing older than the site.

On the API layer, a different instrument and never joined to the check above: on 20 August, the day before the rebuilt site went live, ChatGPT named the company in 13 of 60 answers (22%), twelve questions at five samples each. On 17 September, 28 days later at the same depth, it read 35 of 60 (58%). The report reads: ChatGPT: Change detected: hit-rate 22% → 58% across 12 paired question(s) (exact paired test on hit-rates, p = 0.033). Gemini: No detectable change: hit-rate 0% → 13% across 12 paired question(s) (exact paired test on hit-rates, p = 0.125). What moved ChatGPT: we fixed the site. The other 20 August five-sample run was compared too, but its scoring cannot be confirmed, so it carries no verdict. The 31 July run at one sample, a single pass on 20 August and a third five-sample run where ChatGPT errored stay uncompared. The full record is on the case study.

Read the case study

Stone Ranch Roofing

Not named in the ChatGPT app on 20 August, by our own check. On 2 September the app named it — reading the site we had already replaced.

Single answers in our own logged-in account — evidence, not a rate. On 20 August the app was asked six questions across South Jordan, Herriman and Draper, and named five or six other firms in each, never this one. On 2 September, asked for the best roofing company in Murray — its home market — the app named it, as the second of two alternatives behind the top pick. Different questions on the two dates, so this is what each one observed, not one question moving.

The first citation this client ever earned came from a site we had already replaced. ChatGPT’s write-up quoted details that appear nowhere on the live stoneranchroofing.com — we scanned all 23 live pages — and are still live on the retired domain, which still serves. The engine had read the old site, not the new one. That is the whole argument for watching what the engines actually read about you.

On the API layer — a different instrument, never joined to the readings above — twelve county-wide questions asked five times each on 4 September named it in 6 of 120 samples (5%): five of sixty on ChatGPT, one of sixty on Gemini, while the same answers handed out dozens of competitors. Perplexity is an add-on and was not part of this run; those answers are unknown, not zero. Separately, on 2 September and checked by hand, Google generated no AI Overview on any of the three questions asked — which reads as not generated, never as absent. We rebuilt the site around what buyers actually ask and added a roof pricing tool that returns a real range from the owner’s own job costs instead of a form.

Visit the live site

Steel Haven Fabrication

A one-off fabrication shop in Smith Valley, and a site being built the way the work is: the page opens with the shop’s own mark cut into plates that drop in and weld at the seams, then a pinned scrub walks a weldment from flat stock to finished part. Nothing is finished until the owner says it is, so every capability line is held blank until he confirms it — the process drawing on the page says in its own caption that it is a generic bracket, not one of his jobs.