Field Records

A Cavalier program scored against the site that would outrank it

A small Cavalier King Charles Spaniel breeding program in Clackamas County, Oregon, on Wix. Real litters, real health testing, and invisible for every query a buyer actually types. Rather than hand over a list of opinions, the site was scored out of 100 against a single defined benchmark: the site that would hold first place for “cavalier king charles spaniel puppies oregon.” Then the work was done and it was scored again the same day.

Method. Every page in the sitemap (13) plus 7 blog posts, fetched live. Redirects, robots.txt, sitemaps, canonicals and response headers read directly. Structured data parsed. Page weight and load timing taken from real page loads in Chrome. Live Google result pages for three money queries and the brand query, plus site: coverage and the Bing index. PageSpeed Insights was unavailable that day — Google's free quota was exhausted — so speed was measured from actual page loads instead of estimated.

Score movement in one day

Morning score Evening score, same day Bars scale to each area's weight in the 100-point total.

What actually moved it

• A root cause, not a symptom. Page titles kept reverting after every edit. The cause was that the platform's SEO-tags API returns a stale read, so each change was being built on top of outdated state. The fix was to stop building on a read, apply one canonical tag set in full, and mirror the same values into the editor's own page settings so a publish no longer reverts them. Verified twice over.
• Thirty images given real alt text — 19 that were blank and 11 that were serving the filename — and 477 blog meta descriptions written across a 506-post archive that had none.
• Five stale “Available” puppy pages taken out of the index and out of the sitemap. A breeder site advertising puppies that sold two years ago is a trust problem before it is a ranking problem.
• A third-party tracking script removed from every page, plus FAQ and breadcrumb structured data added, and the reservation page linked from the header and from the body of three other pages. It had previously been reachable from nowhere on the homepage.
• 28 of 28 live signals verified by script after the work, not asserted. Where layout mattered, by screenshot as well.

The plan that overturned its own predecessor

A family Golden Retriever program near Portland, on WordPress. A first-pass SEO package had already been written for it. Before planning the second pass, the live site was re-fetched and checked against what the first plan had assumed — and five of those assumptions turned out to be wrong.

Method. Live-site re-fetch before any planning, on the standing rule that a written plan goes stale faster than the site it describes. Competitor page depth measured by word count on comparable pages. AI visibility measured separately across three named assistants — Claude, ChatGPT and Perplexity — in a 180-answer study run on 2026-09-02.

What the recheck found

• Two pages the first plan said needed building already existed. A full FAQ page was live with nine real questions and answers, and a Guides section was published. Both were unoptimised. That reordered the whole priority list: the fastest wins were structural, not editorial.
• The FAQ schema was then built verbatim from those nine published answers, with an explicit written instruction not to invent a single one. Structured data that says something the page does not say is a liability, not a rich result.
• A health-testing guide was written with its own gate attached — publish only after the actual clearances are confirmed. False health claims about a breeding program are the one mistake in this niche that cannot be walked back.
• The street address was recommended out of publication, not in. A home-based breeder listed as a service-area business protects the household without costing local ranking, and the phone number was locked as canonical before a single directory submission went out. Wrong data propagates to hundreds of sites and is sticky to fix.
• A competitor was measured out-depthing the site on every comparable page — 1,575, 1,300, 2,528 and 3,838 words against 205 to 443. That is recorded here because it was found, and because a portfolio that only contains flattering measurements is not a record of measurement.

Independently measured AI visibility

Building the instrument before selling the reading

The two records above rest on the same question: who do AI assistants actually name, and which sources do they read to decide? Nobody publishes that, so it was measured directly — first 180 answers across 12 states, then extended to 513 answers across 27 states, each run against three named assistants with every source they cited captured.

Method. Identical buyer-intent questions put to Claude, ChatGPT and Perplexity with web search enabled, one state at a time, every cited source recorded. Published openly with the raw data as downloadable CSVs so any claim on this page can be checked against the transcripts rather than taken on trust.

The finding that changes how you'd spend a budget

Each assistant has its own gatekeeper. Asked the same question on the same day, the three engines read substantially different sources — so “we optimised for AI search” is not a coherent sentence unless it names which engine.

• Attention concentrates hard. The ten most-cited domains carry 1,809 of 5,076 citations — 35.6% of everything quoted. One marketplace domain alone accounts for 571 citations across 84 of the 180 answers.
• And the tail is enormous. 552 of the 919 domains appear in exactly one answer. Most of what an assistant reads, it reads once.

First place for the query the business is actually bought on

North Star Tutoring, a reading-specialist practice in Gresham, Oregon. The one record here for a business the author does not own, and the only one naming a third-party client. It is named with that client's written permission, given on 2026-09-16 against the exact wording below. The permission is revocable; if it is withdrawn, this record comes down.

Method. Live Google result pages loaded on 2026-09-09 with Google's own “Gresham, OR” location set — not a rank-tracking tool's approximation of a local result, which is a different measurement and frequently a more flattering one. Business Profile read directly. Site structure taken from the published sitemap and the raw HTML of all nine pages.

Re-checked 2026-09-22. Same method, same “Gresham, OR” localisation, thirteen days later. Still first organic, and still the first business named in Google's AI Overview; the map-pack position had moved from #2 to #1. The figures above are left exactly as they were measured on 2026-09-09, because that is the measurement the client's permission was given against — a later, better number is not something to slide in under an older consent. Rankings move in both directions, which is why every figure here carries the date it was taken.

A detail worth keeping

The practice trades on a .info domain, and a differently-named company owns the matching .com doing unrelated work — test prep and college admissions. Any write-up reaching for the obvious domain would have credited a competitor's site and been caught by the first reader who clicked. Checking which domain you are actually looking at is not a sophisticated step. It is the one most often skipped.