exmxc · AI Retrieval Observatory · Baseline 001
What does generative AI actually retrieve from the open web?
A longitudinal first-party research program measuring how Google generative-AI features retrieve evidence, frameworks, and knowledge objects across publisher-controlled web properties.
Research premise
AI visibility is not the same thing as search traffic. The Observatory measures a different layer: when a search system chooses a site's knowledge as material for a generative answer. Baseline #001 begins with two structurally different sites and 92 days of first-party Search Console data.
Two published checkpoints are available: Baseline #001, exported August 31, 2026, covers May 30–August 29; Checkpoint #002, exported September 30, covers June 29–September 28. Each section retains its own frozen results. The 92-day windows overlap by 62 days, so changes between them are observations rather than independent experimental effects.
01
Baseline #001
The two sites differ in audience and subject matter, so raw counts are not directly comparable. The useful measures normalize AI retrieval against each site's own Search exposure and page universe.
TrailGenic
First-hand field evidence · consumer / longevityWeb Search impressions15,010
Generative-AI impressions2,011
AI Retrieval Share13.4%
AI Retrieval Coverage50.5%
Multi-URL Retrieval Factor1.087×
Dominant AI deviceMobile · 65.0%
exmxc
Frameworks · defined concepts · institutional intelligenceWeb Search impressions7,085
Generative-AI impressions241
AI Retrieval Share3.4%
AI Retrieval Coverage30.0%
Multi-URL Retrieval Factor1.058×
Dominant AI deviceDesktop · 72.6%
AI Retrieval Share = Generative-AI impressions ÷ Web Search impressionsGoogle states that the Generative AI report is part of the Web Search performance dataset. The ratio therefore estimates the share of a property's reported Web impressions occurring inside supported generative-AI features; it does not add AI impressions on top of Web totals.
02
Retrieval propensity by content family
Normalizing by ordinary Search exposure separates pages that are simply popular from pages that appear unusually often in generative retrieval.
TrailGenic
| Family | AI/Web share |
|---|---|
| Trail Logs | 26.0% |
| Gear Reviews | 23.7% |
| Science | 20.4% |
| Lexicon | 12.9% |
| Playbooks | 9.6% |
| Nutrition Reviews | 8.1% |
exmxc
| Family | AI/Web share |
|---|---|
| Frameworks | 9.1% |
| Lexicon | 8.8% |
| sPEG Index family | 5.9% |
| Entity Clarity Index | 2.8% |
| Signal Briefs | 1.8% |
03
The denominator changes the story
The pages with the most AI impressions are not always the pages with the highest retrieval propensity.
| Knowledge object | Property | Web | AI | AI share |
|---|---|---|---|---|
| Baldy winter field log | TrailGenic | 368 | 229 | 62.2% |
| Mafate vs Caldera vs Hierro | TrailGenic | 459 | 208 | 45.3% |
| ATH vs LMNT altitude fuel-curve science | TrailGenic | 806 | 285 | 35.4% |
| Electrolytes at elevation | TrailGenic | 624 | 161 | 25.8% |
| sPEG — Scarcity-Adjusted PEG | exmxc | 115 | 27 | 23.5% |
| Four Forces / Four Pillars | exmxc | 142 | 19 | 13.4% |
| Alphabet $80B signal brief | exmxc | 3,136 | 40 | 1.3% |
04
Findings to test longitudinally
First-hand evidence appears unusually retrieval-prone.
Trail logs, field comparisons, gear tests and derived physiological analysis produce some of the highest normalized retrieval rates in the baseline.
Defined concepts outperform newsier content on exmxc.
Frameworks and Lexicon pages receive substantially more AI impressions relative to ordinary Search exposure than Signal Briefs.
One AI experience can retrieve multiple URLs from one domain.
Google's property and page aggregation differ when multiple URLs from the same site appear. Baseline factors are 1.087× for TrailGenic and 1.058× for exmxc.
AI retrieval can occur without observable visits.
The Baldy field log recorded 229 AI impressions while the Web Search export recorded zero clicks. Retrieval measures an exposure layer traffic analytics alone misses.
05
Preregistered hypotheses
Written before the September checkpoint. Future reports will record support, contradiction or ambiguity rather than rewrite the premise after seeing the outcome.
Original evidence increases retrieval magnitude.
First-hand tests, original measurements, proprietary datasets and direct observations will receive disproportionate generative-AI retrieval versus commodity summaries.
Comparative pages outperform broad explainers.
Pages that help a user choose, distinguish or evaluate will show higher normalized retrieval than broad informational pages.
Knowledge architecture increases breadth.
Entity, lexicon and framework architecture will correlate more with retrieval coverage and multi-URL retrieval than with raw volume alone.
Retrieval context differs by knowledge market.
Consumer experiential knowledge will remain more mobile-skewed; institutional and strategic knowledge will remain more desktop-skewed.
06
TrailGenic momentum
Same-length windows reduce distortion from partial calendar months.
| Window | Web impressions | AI impressions | AI Retrieval Share |
|---|---|---|---|
| May 30–June 28 | 3,878 | 348 | 9.0% |
| July 31–August 29 | 5,547 | 950 | 17.1% |
CP2
Checkpoint #002 — September 30, 2026
A second 92-day rolling window, June 29–September 28, compared with Baseline #001, May 30–August 29. The windows overlap by 62 days; deltas describe rolling composition, not independent month-over-month samples.
| Property | Metric | Baseline #001 | Checkpoint #002 | Change |
|---|---|---|---|---|
| exmxc | Web impressions | 7,085 | 6,406 | −9.6% |
| exmxc | Generative AI impressions | 241 | 271 | +12.4% |
| exmxc | AI Retrieval Share | 3.40% | 4.23% | +0.83 pp |
| exmxc | AI Retrieval Coverage | 30.0% | 33.5% | +3.5 pp |
| exmxc | Multi-URL Retrieval Factor | 1.058× | 1.063× | stable |
| exmxc | Top-five concentration | 50.2% | 40.6% | −9.6 pp |
| TrailGenic | Web impressions | 15,010 | 16,190 | +7.9% |
| TrailGenic | Generative AI impressions | 2,011 | 2,352 | +17.0% |
| TrailGenic | AI Retrieval Share | 13.4% | 14.53% | +1.13 pp |
| TrailGenic | AI Retrieval Coverage | 50.5% | 50.5% | flat |
| TrailGenic | Multi-URL Retrieval Factor | 1.087× | 1.080× | stable |
| TrailGenic | Top-five concentration | 64.5% | 65.9% | +1.4 pp |
Retrieval broadened while a transient event page faded.
The Alphabet $80B Signal Brief fell from 40 AI impressions to 9, yet total exmxc AI impressions rose to 271, coverage expanded from 30.0% to 33.5%, and top-five concentration fell from 50.2% to 40.6%. The current leaders are Narrative Authority (31), AI Territorial Claims (27), sPEG (23), Four Forces / Four Pillars (18), and the Entity Clarity Index (18). This is consistent with a broader retrieval footprint centered on reusable concepts and frameworks rather than one event-driven object.
AI retrieval is not simply following ordinary Search popularity.
Narrative Authority recorded 31 AI impressions against 60 Web impressions, while the Mike Ye bio recorded 5 AI impressions against 2,770 Web impressions. The contrast supports the Observatory's use of normalized retrieval measures: high Search exposure alone does not explain which exmxc knowledge objects appear in generative-AI features.
Comparison and field-evidence objects strengthened.
The Mafate 5 vs Caldera 8 vs Hierro v9 comparison increased from 208 to 399 AI impressions (+91.8%), Fuel Curve Science rose from 285 to 339 (+18.9%), and Protein After Fasted Hiking rose from 79 to 106 (+34.2%). The ATH Lytes vs LMNT comparison remained the largest object at 558. The pattern remains consistent with retrieval value in specific comparisons, first-party evidence, and reusable conclusions.
The device split remains property-specific.
exmxc became more desktop-skewed, from 72.6% to 78.6% of AI impressions. TrailGenic remains mobile-majority at 62.6%, versus 65.0% at baseline. The directional split anticipated in H4 persists: strategic and institutional retrieval is more desktop-heavy, while experiential TrailGenic retrieval remains more mobile-heavy.
Rolling improvement and near-term momentum are not the same signal.
exmxc's 92-day AI Retrieval Share rose to 4.23%, but its latest 30-day window (Aug. 30–Sep. 28) produced 124 AI impressions on 3,976 Web impressions, a 3.12% share versus 4.21% in Jul. 31–Aug. 29. The final nine days accelerated sharply to 69 AI impressions, but one short burst is not enough to establish a new regime. TrailGenic's latest 30-day share cooled from 17.1% to 13.6% after an unusually strong August even as its 92-day rolling share rose to 14.53%.
Checkpoint #002 is supportive, not dispositive.
H1 and H2 receive additional support from TrailGenic's evidence-rich and comparative objects. H3 receives support from exmxc's wider coverage and lower concentration, although the Multi-URL Retrieval Factor is essentially unchanged. H4 remains supported by the persistent device split. These are observational associations; the checkpoint does not isolate causality or reveal the prompts that triggered retrieval.
07
Methodology and boundaries
First-party Search Console exports
Baseline #001 uses Web Performance and Generative AI Performance exports downloaded August 31, 2026, covering May 30–August 29. Checkpoint #002 uses the same export types downloaded September 30, 2026, covering June 29–September 28. Raw property and page aggregation remain distinct.
Host variants normalized for coverage.
www / non-www variants are normalized when counting canonical knowledge objects. Raw property and page impression counts remain Google's reported values.
Observational, not causal.
The sites operate in different query markets, with different histories and addressable demand. A higher retrieval share does not prove one architecture is intrinsically superior.
The triggering questions are hidden.
The report exposes pages, countries, devices and dates, but not exact prompts. Content-level interpretation is inference, not prompt-level attribution.
08
Research log
Baseline #001 frozen. exmxc and TrailGenic exports preserved; metrics and four hypotheses preregistered.
Checkpoint #002 completed. Re-exported both properties and compared rolling 92-day retrieval share, canonical coverage, multi-URL retrieval, concentration, device mix, momentum, and knowledge-object composition.
Longitudinal report. Publish when accumulated observations support a meaningful update; monthly measurement does not require monthly narrative.
Treatment #003 deployed — ECI named-evidence exposure. Post-baseline intervention: the Entity Clarity Index hub now surfaces selected dated entity evidence, and Finance plus Technology report summaries expose company-specific examples with explicit interpretation boundaries. Baseline #001 metrics remain frozen and unchanged.
09
Primary measurement references
- Google Search Console — Generative AI performance report (Search)
- Google — How impressions, positions and clicks are counted, including AI Mode and AI Overviews
- Google Search Console — Search Performance report methodology
- Google Search Console — Data anomalies
Metric definitions. AI Retrieval Share = property-level Generative AI impressions ÷ property-level Web Search impressions. AI Retrieval Coverage = normalized canonical pages with at least one Generative AI impression ÷ normalized canonical pages with at least one Web Search impression. Multi-URL Retrieval Factor = sum of page-level Generative AI impressions ÷ property-level Generative AI impressions.
Interpretation boundary. “Retrieved” means a link to the property was reported as shown within a supported Google generative-AI Search feature. It does not prove that a specific sentence was quoted, that a model trained on the page, that the page caused the answer, or that the user visited the site.
Working model, not conclusion. Baseline #001 motivates a testable proposition: original evidence may create information gain, while knowledge architecture may make that evidence easier to identify, retrieve and combine.