2026·01
The research program
We map how answer engines select and absorb sources: peer-reviewed literature, the pipeline detail in two court records, and our own citation probes. Every build runs on that playbook.
Edge
No. 001
Engines retrieve. Then they cite.
When you ask an AI assistant a commercial question, it does not consult a memory. It fans your question out into machine-generated sub-queries, retrieves passages from a search index at query time, and constructs an answer from what it finds. The pipeline is peer-reviewed and described identically in two independent court records.
The consequence is structural. Whoever publishes the best-structured, most extractable answer to a question becomes part of the answer. Retrieval happens at passage granularity, per question, every time. There is no fixed set of slots to compete for.
That is the whole mechanism. Everything Sternhaven does follows from taking it seriously.
No. 002
Billion-dollar verticals. Nobody keeping the record.
The research literature has never measured how many commercial verticals lack an authoritative source. So we built the instrument and measured it ourselves: a screen of roughly 5,400 commercial niches, each candidate checked against the named businesses actually operating in it.
Verified voids are real, checkable, and cheap to find. They are also rare: roughly a dozen buildable openings survived the screen. Rarity is the point. A maintained registry of verified gaps, including every dead end, is proprietary deal flow no competitor can shortcut.
What the screen kept finding instead of authority:
- Found
Directories listing big-box chains as "verified" local specialists.
- Found
A national brand's dealer locator serving zero dealer data in its page source.
- Found
Trade-association domains sitting parked.
- Found
A 282,000-searches-per-month category with almost no authoritative coverage.
What we blast and what we bless
Blast
The data rot machines are fed today
Parked authority
Trade bodies and "official" sources that stopped maintaining anything years ago.
"Verified" badges nobody verified
Trust language with no process behind it. We run a moratorium on the word.
Slop at scale
Roughly 16% of sources cited by major engines are themselves AI-generated. A slop site can win a citation. It cannot keep one.
Pay-to-play posing as ranking
Rankings that sell their order without saying so. Placement is disclosable. Merit is unbuyable.
Answer-shaped bait
Pages engineered for citation that read as engineered. Engines already name this as spam.
Bless
What a source of record owes both audiences
Provenance on every fact
Sourced, dated, corrected on request, with dead listings pruned. Maintenance is the moat.
Disclosed method
Editorial rankings carry their methodology. Opinion is labeled opinion.
Owned distribution
Email captured from day one on every property. The one channel no engine can reprice.
Structure machines can absorb
Definitions, numbers, comparisons, one buyer question per page. The shape that measurably wins citations.
Pages a person would use
Judged as if search engines did not exist. A page that fails that test does not ship.
No. 003
The window is open. We are moving while it is.
52.7→32.6
Top-1,000 domains' share of the top citation
From classic search results to generative answers. The redistribution is measured, present-tense, and not permanent.
23.8%
Of citations go to the 25 biggest domains
The incumbent tier holds under a quarter of generative citations, and 59.1% of cited domains are cited exactly once.
22–41%
Visibility gain from structured, cited content
Peer-reviewed. Statistics, definitions, comparisons, and sources measurably win answers, and small sources gain most.
No. 004
The record. January to now.
Eight months from thesis to operating portfolio. Every entry below is measured on our own instruments, not projected.
2026·03
The mechanic, proven on our own property
Our most mature property holds roughly 3,900 verified entities across roughly 9,300 indexed pages, and sustains a 63 to 71% AI-citation band on buyer queries, 30 points above its own internal control.
Proven
2026·06
The deal-flow instrument
5,400 niches screened for a few dollars in compute, every verified opening and every dead end recorded in a proprietary registry. Entry decisions run on evidence, never enthusiasm.
Built
2026·07
The fleet
Five more properties live on one shared build system, each with search indexing, instant-index submission, and email capture wired on day one.
Live
2026·08
The portfolio
Seven verticals run as reference sources, more than 27,000 pages in print, and a ranked target list feeding the next builds.
Scaling
The curve.
Total Google search impressions, all properties combined, complete weeks, and the share of AI answers citing our mature property. April to September 2026. One build system, every property counted.
Weekly impressions
AI citation share share of AI answers citing our mature property on a fixed panel of buyer queries, four engines
| Week | Impressions | Clicks |
|---|---|---|
| Week of 2026-04-06 | 8,764 | 51 |
| Week of 2026-04-13 | 10,017 | 44 |
| Week of 2026-04-20 | 9,468 | 44 |
| Week of 2026-04-27 | 19,944 | 48 |
| Week of 2026-05-04 | 22,448 | 49 |
| Week of 2026-05-11 | 28,157 | 75 |
| Week of 2026-05-18 | 28,159 | 98 |
| Week of 2026-05-25 | 38,467 | 122 |
| Week of 2026-06-01 | 36,848 | 124 |
| Week of 2026-06-08 | 45,771 | 122 |
| Week of 2026-06-15 | 47,584 | 126 |
| Week of 2026-06-22 | 44,166 | 137 |
| Week of 2026-06-29 | 53,533 | 162 |
| Week of 2026-07-06 | 82,741 | 174 |
| Week of 2026-07-13 | 70,489 | 186 |
| Week of 2026-07-20 | 135,967 | 278 |
| Week of 2026-07-27 | 215,451 | 343 |
| Week of 2026-08-03 | 178,082 | 365 |
| Week of 2026-08-10 | 151,277 | 390 |
| Week of 2026-08-17 | 146,417 | 392 |
| Week of 2026-08-24 | 132,560 | 331 |
| Week of 2026-08-31 | 142,922 | 349 |
| Run date | Answers citing the property | Share |
|---|---|---|
| 2026-07-08 | 15 of 24 | 62% |
| 2026-07-10 | 17 of 24 | 71% |
| 2026-07-12 | 15 of 24 | 62% |
| 2026-07-13 | 17 of 24 | 71% |
| 2026-07-14 | 16 of 24 | 67% |
| 2026-07-15 | 16 of 24 | 67% |
| 2026-07-16 | 17 of 24 | 71% |
| 2026-07-18 | 17 of 24 | 71% |
| 2026-07-21 | 16 of 24 | 67% |
| 2026-07-22 | 18 of 24 | 75% |
| 2026-07-23 | 17 of 24 | 71% |
| 2026-07-24 | 16 of 24 | 67% |
| 2026-07-25 | 18 of 24 | 75% |
| 2026-07-27 | 17 of 24 | 71% |
| 2026-07-29 | 16 of 24 | 67% |
| 2026-07-30 | 14 of 24 | 58% |
| 2026-07-31 | 15 of 24 | 62% |
| 2026-08-01 | 16 of 24 | 67% |
| 2026-08-02 | 16 of 24 | 67% |
| 2026-08-03 | 15 of 24 | 62% |
| 2026-08-10 | 20 of 24 | 83% |
| 2026-08-17 | 19 of 24 | 79% |
| 2026-08-31 | 18 of 24 | 75% |
| 2026-09-07 | 16 of 24 | 67% |
| 2026-09-14 | 18 of 24 | 75% |
The question nobody else can measure: how citation share behaves in verticals where competing sources approach zero. A portfolio of maintained sources in thin categories is the only instrument positioned to answer it, and every new property sharpens it. That research edge compounds.
No. 005
What we build. One vertical at a time.
More than 27,000 published reference pages across the portfolio.
Every fact provenance-logged, every dead listing pruned. Maintenance is the moat.
The most complete source in the category
Each property is the fullest, most accurate, most honestly maintained dataset and editorial layer in a vertical that never had one.
Structured for machines
Passage-coherent pages, one buyer question at a time, dense in the definitions, numbers, and comparisons engines measurably absorb.
Trustworthy for people
Disclosed methodology, labeled opinion, a corrections process, and a real legal entity standing behind every page.
Born ready to stand alone
Its own domain, dataset, audience, and books from day one. A property can graduate or sell as a transfer, never an untangling.
Know a vertical running on bad data?
If you work in a category where the "authoritative" sources are parked domains, stale listings, or chains dressed as specialists, we want to hear about it. Every nomination runs the same verification gauntlet as our own research. Most fail. The ones that pass get built properly.