"Second measurement cycle, same pipeline: the leading indicators kept compounding — the June foundations at Boss Hog added another 5.5 points of AI share of voice on their own. Position, once earned, calcifies."
Cartwheel (retail CPG) · Boss Hog Moving (local services) · Joybody Physio (clinic) — second cycle (Jul 13–Aug 24) spliced onto the same measured series, marketing spikes excluded from every number, every claim traceable to a dated window
Share of AI answers
+26%
Relative growth in AI answers naming Boss Hog — 21.0 → 26.5% (+5.5pt), now 1 in 4 answers — self-sustaining: no new work required in the window. National franchises measured in the same runs: 34.5–39.5% (Jul 14 reads). The June foundation compounding on its own.
AI citations
+107%
Cartwheel AI citations 1.62 → 3.36/day. Monthly arc 21 → 24 → 146, now consolidating at ~2× the spring floor. Conservative full-window basis — the masked-excluded read is +222%.
AI-customer velocity
+50%
Named-ChatGPT-customer velocity at Boss Hog: 2 per ~6 weeks → 3 per 6 weeks (5 cumulative) — plus its first AI referral sessions ever: 4 chatgpt.com sessions after zero all prior period. The zero-click wall showing its first crack. Self-reported intake, n=5 — instrumentation proof, not a traffic claim.
Branded lift, control case
+58%
Joybody branded clicks across the schema-v4 deploy gate (1.12 → 1.78/day). Caveat carried with the number: +9% vs the June baseline (the gate window held a lull) — the sturdier read is the engagement-long climb 1.29 → 1.43 → 1.67 → 1.78, the best sustained rate yet.
The instrument panel
ADI component growth
Each client's component pairs — engagement-long series where measured, this cycle's relative move as the lead number — reported as pairs, never composited into one ADI number. Notes ¹–⁴ sit under the component table below.
not aggregable — instrumented at 1 of 3rollout: 1 proven · 1 queued · 1 off-site by design
Derived: unweighted mean across 3 clients, engagement-long; mixed verticals and engagement lengths — direction-only. Components stay separate — never aggregated across components into one number.
+59% clicks across the v4 gate · −2% position (flat)clicks 1.29 → 1.43 → 1.67 → 1.78/day (engagement-long climb) · position 2.07 → 2.61 → 2.39
CD
n/a — Jane handoff broken; proven intake fix queued³
The pull quotes, case by case
Three engagements, second cycle
Boss Hog Moving · local services · 11 weeks in
The June foundations kept compounding — +5.5 points of AI share of voice in six weeks, on their own
"The June foundations kept working on their own — grounded SoV rose to 26.5% with no new work required."
No new deploys in the window (a natural persistence test — last change Jun 29); identical 50-prompt set × 4 platforms, 0 failures both runs; Anthropic 2% → 20% — first run with all four platforms ≥20%; now above B2 Moving's July read (25.5%), closing on the franchises (34.5–39.5%, Jul 14 reads)
"First chatgpt.com referral sessions ever — and three more customers typed 'ChatGPT' on the quote form."
4 sessions (Aug 1, 5, 15, 23) after 0 across the entire prior dataset; 5 named AI leads cumulative ≈ 4% of named leads; Aug 15 = first same-day co-occurrence of a named lead and a referral session — noted as observation, the channel remains structurally zero-click
26.5%
of AI answers name Boss Hog (21.0% in July; +26% rel) — on the June foundations alone
6 wks
self-sustaining — no new work required in the window (a natural persistence test; last change Jun 29)
5 named
cumulative ChatGPT leads on the quote form; velocity up (2 → 3 per ~6 weeks)
4 sessions
first AI referral sessions ever (GA4 ai-assistant channel, all in August)
Honestly: branded clicks are flat, not up — masking the newsletter pulses out of the baseline flips the delta sign (−2.6% → +3.1% = noise); quotes held at peak level with the predicted seasonal taper starting ~Aug 10. Open front: non-branded position 18.7 → 14.1 — better, still below the local pack; 34% of non-branded demand is commercial with zero commercial pages on the site.
AI citations broke out +107% and consolidated at ~2× the spring floor — branded intent keeps routing to the retail shelf
"July ran 146 citations against 21 in May — and August is holding near 2× the spring floor, not reverting."
Monthly arc 21 → 24 → 146; record week Jul 13–19 = 57; August settled ~1.7/day vs the 0.70/day May floor. Consistent with the schema + expert-article supply meeting seasonal demand — the components are not separable this cycle
"Lice season lifts every series except branded clicks — the divergence is consistent with retail routing, not demand loss."
Total clicks +26%, non-branded clicks +49%, organic sessions +40% while branded clicks fell −22% (−14% masked) at better-than-baseline position with CTR down ~4pp — the click resolves at Walgreens/Amazon listings; unverified at POS
of AI answers (from 4%) — flat within noise; the citation→answer lag is running longer than the July model predicted
Aug 13
first AI-assistant-referred session of the engagement window — a milestone, not a trend
Honestly: the July peak overlaps client content changes (Jul 14–Aug 20, dates approximate) and peak season — "consistent with," never "caused by"; the Aug 21 Openridge deploys are explicitly out of attribution this cycle. Open front: whether citations hold ≥2× is next cycle's first question; retail sell-through data (Amazon Brand Analytics + Walgreens) still missing — third cycle without the conversion end of the funnel.
The discipline report: leading indicators moving, headline honestly flat — no claim before the pre-registered checkpoint
"Branded demand at its best sustained rate of the engagement — and we still claim no AI-answer gain."
+58% across the v4 gate (1.12 → 1.78 clicks/day) — but window A held a branded lull, so vs the June baseline it's +9%; the sturdier read is the monotonic climb 1.29 → 1.43 → 1.67 → 1.78. Grounded SoV 8.5 → 8.5%, flat — the claim gate is the pre-registered week-12 checkpoint (~late October)
"AI-assistant visits run ≥2× the pre-deploy baseline in both windows — clinics click through where product brands go zero-click."
0.31/day (A) and 0.19/day (B) vs 0.10/day baseline; 10 chatgpt.com sessions in 43 days vs 6 in the entire prior period; window-over-window it is down — both figures reported. 5 sessions per window = direction, not a trend
+58%
branded clicks across the v4 gate (caveats carried: +9% vs June baseline)
≥2×
AI referral rate vs pre-deploy baseline, in both sub-windows
grounded SoV, flat — no gain claimed; week-12 checkpoint vs the clean Jun 29 baseline
Honestly: citation go-live dates are approximate (window A is a ramp, not a clean control), paid search rose +54% in window B with keywords still unknown, and August seasonality can't be excluded on 27 days of a first-year property. Open front: booking instrument still broken — a second full cycle blind on the outcome metric; the Jane intake "how did you hear" field (the Boss Hog instrument) remains the fix.
Three components, three clients, one calendar. Dots are measurements — no smoothing, no interpolation; flats and declines plotted as-is. SoV points are the per-client Argus run series; BSL/CD points are cycle-window per-day rates plotted at window end. Vertical ticks mark client deploys (labeled on the first chart). Caveated readings keep their notes ¹–⁴ in the table below.
Boss Hog MovingCartwheelJoybody Physio┊ deploy marker · ● measurement
SoV — share of AI answers (grounded), % of panel
Argus grounded all-platform runs, identical prompt sets per client (degraded Jun 9 Boss Hog run excluded per the failure-sidecar rule).
BSL — branded search clicks/day
Cycle-window per-day rates plotted at window end — 3–4 points per series, not daily data. Boss Hog's August point: flat is the honest read¹.
CD — conversion proxy/day
Instrumented at 1 of 3 (rollout: 1 proven · 1 queued · 1 off-site by design) — the chart is mostly Boss Hog, and that is the honest picture. Boss Hog = quote requests/day; Cartwheel = retailer-handoff clicks/day.
The numbers
Cycle-over-cycle detail
SoV = share of AI answers · BSL = branded search · CD = conversion. Components reported, never composited. July rows in grey for cycle-over-cycle reading.
Client · cycle
SoV — AI answers (live)
SoV — AI memory (cold)
BSL — brand position
BSL — brand clicks/day
CD — conversion proxy/day
Boss Hog · Jul
19 → 21% (+11%)
0.5 → 5.5% (×11)⁴
3.6 → 2.6 (+28%)
1.3 → 3.12 (+140%)
quotes 0.6 → 3.47 (+478%)
Boss Hog · Aug
21 → 26.5% (+26%)
5.5 → 6.5% (+18%)⁴
2.79 → 1.67 (+40%)
2.56 → 2.64 (+3%)¹
quotes 2.19 → 2.36 (+8%)¹
Cartwheel · Jul
0 → 4% (new)
0% (flat)
5.5 → 3.6 (+35%)
2.19 → 0.9 (−59%)²
retail handoff — 0.08 (too sparse to trend)
Cartwheel · Aug
0 → 5% (new)
0% (flat)
5.5 → 5 (+9%)
2.19 → 1.3 (−41%)²
retail handoff — 0 this window²
Joybody · Jul
8 → 8.5% (noise)
1 → 2.5% (noise)
2.07 → 2.61 (dilution)
1.43 → 1.67 (+17%)
booking handoff — instrument broken³
Joybody · Aug
8.5 → 8.5% (flat)
2.5 → 3% (noise band)
2.35 → 2.39 (flat)
1.12 → 1.78 (+59%)
booking handoff — still broken³
¹ Boss Hog branded and quotes: flat is the honest read. The recurring-newsletter pulse days sat inside the pre baseline; the strict mask flips the branded delta sign (−2.6% std → +3.1% strict) and does the same to quotes (−4.5% → +7.8%) — neither clears noise. Quotes held at peak-season level; seasonal taper from ~Aug 10, as predicted in July. ² Cartwheel clicks route to Walgreens/Amazon retail listings (~4–11 clicks/wk resolving at retail listings depending on CTR reference; July's read was ~10/wk) — the season lifts every other series while branded clicks fall; consistent with routing, not demand loss; unverified at POS, and true CD still requires Amazon Brand Analytics + Walgreens sell-through (requested, third cycle). ³ Joybody's Jane handoff tracking is literally broken (cross-domain suppression + tag not on Jane; root-caused Jul 9) — a second full cycle blind on the outcome metric. The fix is the instrument already proven at Boss Hog (the intake "how did you hear" field, which caught all 5 AI-attributed customers there): build item #1, queued. Portfolio rollout status: 1 proven · 1 queued · 1 awaiting client data access (Cartwheel — conversion is off-site by design). ⁴ The Boss Hog "cold" series is Perplexity retrieval, not model memory — see the correction below.
Continuity with the July page — nothing dropped, nothing hidden: July's hero +478% now reads as the series quotes 0.6 → 3.47 → 2.36/day — held at peak-season level, with the seasonal taper July predicted now in the data (note ¹). July's hero +140% branded reads as 1.3 → 3.12 → 2.64 clicks/day — the honest August continuation is flat (note ¹). July's cohort roll-ups (cohort SoV average 8.7 → 11.2%, cohort branded +53%, AI-sourced touches ×2 with 9 named portfolio-wide) are not restated — no fresh cohort roll-up was computed this cycle; the per-client series above are their continuation, and the touch counts continue per client (Boss Hog 5 named + 4 referral sessions; Joybody 16 lifetime AI referrals). July's 48h first-effect / median-12-days tile was a first-cycle onboarding metric, not re-measured; this cycle's analogue is the Joybody v4 gate — deploy Jul 29, leading indicators within 27 days, headline gated to week 12.
The causal architecture
Three natural experiments
Not an RCT — but a control, a persistence test, and a gated deploy across three unrelated businesses, all on one pre-registered measurement pipeline. Each deploy→response pairing in the client reports carries an evidence grade — Gated (locked pre/post window around a dated deploy) · Aligned (event-aligned with a deploy, confounds named) · Directional (small-n movement in the predicted direction) — and the per-client reports carry the full deploy→response ledgers.
A · Control · Joybody
Same pipeline, no treatment, no effect.
The May schema deploy silently failed — and every metric ran 8 weeks flat while it went undetected. The fix was verified live Jun 30, and branded appearances rose +43% within 12 days. The attribution the other two cases lean on.
B · Persistence · Boss Hog
Gains compound unattended.
No new deploys Jun 29 → Aug 25 — a natural persistence test: grounded SoV rose +5.5pt (21.0 → 26.5%) anyway, on the identical 50-prompt set with 0 failures both runs. The June foundation being re-crawled and re-weighted, not new work.
C · Gated deploy · Joybody
A dated gate, caveats carried with the number.
Schema v4 live Jul 29 (curl-verified). Branded clicks +58% across the gate window, +9% vs the June baseline — the window-A lull and the approximate citation-ramp dates are named confounds, and the read is robust to excluding the deploy day itself.
The dose–response read
Two inputs, three doses
Discoverability responds to two distinct inputs — a machine-readable foundation, and third-party corroboration. Our portfolio measured them separately.
Dose 1 · Foundation alone
Entry, then plateau.
Cartwheel entered the AI answer set from absolute zero on schema work, then flat ~5% since (series 0 → 8 → 5 → 4 → 5) — and its +107% citation lift is Bing retrieving their own site (retrieval supply, not third-party corroboration), which has not moved answer-share. Joybody's foundation moved branded demand +43% / +58% while answer-share held 8.5% flat. The foundation's effect is real, one-time, and floor-shaped.
Dose 2 · Foundation + corroboration events
Strong growth.
Boss Hog — the only client with both inputs (GEO sprint + KNWA TV + NWA Daily coverage) — is the only client with strong monotonic SoV growth: 16.1 → 26.5% across four runs (Jun 9 pre-deploy → Aug 25). Its corroboration was earned media, not a managed program — an accidental half-dose produced the portfolio's best result. Runs: 16.1% Jun 9 (pre-deploy baseline, n=193) → 19.0% Jun 11 (deploy day) → 21.0% Jul 14 → 26.5% Aug 25 (n=200 each). Source: argus-sov-series.csv; a degraded same-day Jun 9 run (25 failures) excluded per the failure-sidecar rule.
Dose 3 · Foundation + sustained corroboration
Never yet run.
Every corroboration input in this portfolio was one-shot or accidental, and every SoV curve plateaued or coasted accordingly. The third rung is the recommendation: discoverability is a maintained system, not a deploy.
Cross-client reading, n=3, grade: Aligned — consistent with the ladder, not proven dose–response.
How to read this honestly
What the portfolio shows: the same sequence — foundation → citations → AI answers → branded demand → conversion — holding through a second measurement cycle across three unrelated businesses, including a six-week natural persistence test at one client — no new deploys in the window — where AI share of voice rose 5.5 points anyway. Every number is clean-window, event-masked, and per-day normalized; every confound is named; cycle-over-cycle rows share one spliced series per client.
What it doesn't show yet: n=3 is a pattern, not a population — say "across our first three engagements." Conversion instrumentation is a rollout, not a defect: proven at one client (Boss Hog's intake question — it caught all 5 AI-attributed customers), queued at one (Joybody: the Jane intake field, build item #1), and off-site by design at one (Cartwheel is retail-first — conversion lives at Walgreens/Amazon; the measure is sell-through data, requested). No composite score is presented because the components aren't normalized to one scale. Every claim traces to a dated, reproducible measurement — ask for any receipt.
What this cycle corrected (flagging our own July inconsistencies): (1) Boss Hog's "cold memory ×11" was Perplexity retrieval all along — Perplexity retrieves even in cold mode, and every cold mention in both runs is Perplexity with domain citations attached. True non-retrieval cold is 0/150 → 0/150 both runs; from this cycle we report non-retrieval cold as the memory metric. Zero at month 3 is normal — it is the next frontier, not a regression. (2) Joybody's +58% carries a window-A-lull caveat in every appearance: window A held a branded lull that inflates the gate contrast; against the June baseline the lift is +9%, and the sturdier read is the engagement-long monotonic climb to the best sustained rate yet.