The intervention shipped at 22:33 UTC on 23 April 2026. First measurement is five weeks away. This post exists so the hypothesis cannot be quietly rewritten to fit the data.
Hypothesis: A bundled entity-graph intervention — ORCID identifier, two Zenodo DOIs referenced as ScholarlyArticle and Dataset nodes in the About page JSON-LD, a public research code repository with cross-referenced DOIs, and (in a later follow-up intervention) three Wikidata items — raises Tier 2 Perplexity citation rate above the 22.0pp interpretability threshold established by E016.
Null prediction: T2 Perplexity citation rate stays at 0% on the next two E014 monthly measurements (M1 late May, M2 late June). A below-threshold lift is noise.
Intervention timestamp: 2026-04-23T22:33:39Z. Measurement window opens at M1.
Why This Post Exists Before the Data
The usual GEO Lab post order is: hypothesis, intervention, measurement, writeup. This post inverts it. The intervention deployed yesterday. The first usable measurement is five weeks away. The writeup is happening now.
The reason is not enthusiasm. It is a hygiene problem specific to bundled interventions: once a result comes in, there is always a tempting rewrite where the hypothesis becomes whatever the result happened to support. The safest defence against that is to write down the hypothesis in public before the result arrives. If the data comes back positive and the post’s prediction was null, the site reports a positive finding against a pre-registered null. If the data comes back null and the prediction was null, the post reports a null against a pre-registered null — which is also a legitimate finding.
Either way, the interpretation is constrained by what this page says today, and not by what looks good to say in June.
How the Intervention Was Designed
The starting point was not a controlled hypothesis. It was an audit. In reviewing the site’s entity footprint, the gap was straightforward: no Wikidata item, no sameAs cross-references between the identifiers that did exist, no unified JSON-LD on the About page, no DOI links into the Zenodo records, no public research repositories. The standard framing for closing those gaps is aspirational: do all of this and the site becomes “Knowledge Panel eligible.”
The framing did not survive contact with the Lab’s existing measurement record. Three pieces of that record mattered.
First, the EDX null (Zenodo 10.5281/zenodo.19450361) established that on Tier 2 category queries, the site achieves 0% Perplexity citation rate across ten pages spanning 12.9 to 29.5 unique named entities per 1,000 words. The finding was interpreted as a domain authority threshold effect: below a competitive citation gate, page-level variables do not operate. If the gate is real, an on-page entity density intervention cannot move T2 citation rate. The only interventions that can move the gate are external authority signals — not more entities on the page.
Second, the E016 noise floor measurement established that the interpretability threshold for any citation rate change is 22.0 percentage points combined. Anything below that is indistinguishable from platform variance. This is the threshold any claimed lift has to clear.
Third, E016 also established zero within-day variance on Perplexity (40% across all checks of a Tier 1 query set) and 0% across 50 checks on Google AI Overviews. Those platform behaviours constrain what “positive result” can mean on each platform.
With those three findings in hand, the entity-footprint audit was re-scoped as a narrow, pre-registered, measurement-first experiment. The “Knowledge Panel eligibility” framing was dropped — it overclaims what any Wikidata item can produce on its own. The question became smaller and testable: does adding external, cross-referenced entity identifiers move T2 Perplexity citation rate above the E016 interpretability threshold, or does the authority gate hold?
That is E025.
Hypothesis and Null Prediction
EXPERIMENT_ENTITY_REINFORCEMENT_E025:
hypothesis: "External, cross-referenced entity identifiers (ORCID, Zenodo
DOIs as JSON-LD nodes, public research repos, Wikidata items)
raise T2 Perplexity citation rate above the 22.0pp
interpretability threshold from E016."
null_prediction: "T2 Perplexity citation rate remains at 0% on M1 and M2
post-intervention measurements. A below-threshold lift
is noise, not signal."
interpretability_threshold: "22.0pp combined (from E016 noise floor)"
intervention_date: "2026-04-23T22:33:39Z"
measurement_cadence: "E014 monthly, M1 late May 2026, M2 late June 2026"
design: "bundled — positive result does not identify which component
produced the lift; decomposition is a follow-up experiment"The hypothesis is framed against the authority gate finding from EDX. If the gate is real and on-page variables do not operate below it, the intervention’s only path to a T2 lift is by partially clearing the gate itself — through external identifiers that systems outside the site recognise as authority signals. Not through anything the site says about itself.
The null prediction is the stronger position. If the gate holds against this intervention, that is a meaningful Layer 4 result: it says ORCID plus Wikidata plus cross-referenced DOIs plus public repos are not enough, on their own, to partially clear the citation gate for a site at this authority tier. A companion null to EDX, at a different layer.
Intervention Components
The intervention is bundled. It has eight components across three waves. The first wave deployed on 23 April. The second and third waves follow over the coming weeks with their own intervention timestamps.
| Component | Description | Wave | Status |
|---|---|---|---|
| ORCID | Researcher identifier 0009-0004-4072-9741 attached to Person schema node and both Zenodo records |
1 | Live 23 Apr |
| Public code repo | arturseo-geo/geo-citation-index flipped public with Zenodo DOI badge and BibTeX in README |
1 | Live 23 Apr |
| Zenodo related identifiers | Record 19253920 linked to GitHub repository; ORCID attached to creator on both records | 1 | Live 23 Apr |
Unified @graph on /about/ |
Single JSON-LD block emitting Person, Organization as ResearchProject, DefinedTerm for the GEO Stack, Dataset for Zenodo 19253920, ScholarlyArticle for Zenodo 19450361. Cross-referenced by @id |
1 | Live 23 Apr |
| Wikidata Person item | Three statements minimum: instance of human, notable work The GEO Lab, notable work The GEO Stack. ORCID and official website as external identifiers | 2 | Pending |
| Wikidata Lab item | The GEO Lab as research project, with founder link to Person item, and part — The GEO Stack | 2 | Pending |
| Wikidata Stack item | The GEO Stack as conceptual model, creator link to Person item | 2 | Pending |
| Cross-platform profile alignment | LinkedIn headline and featured section + GitHub profile README aligned with the About page entity description | 3 | Pending |
The bundled design is a deliberate trade-off. A single-variable test across eight components would take two years at a one-per-month cadence and the earliest components would decay before the last ones deployed. A bundled test produces an interpretable yes or no on the overall intervention in two months. Decomposition is a follow-up experiment.
Methodology
The measurement design reuses E014 Month 1 Citation Rate Baseline‘s monthly cadence and query set. No new measurement infrastructure is introduced. This is intentional — it keeps the measurement comparable to the baseline and to every other monthly run.
The E014 combined query set of Tier 1 proprietary concept queries and Tier 2 category queries. Same ten queries per tier, same phrasing, same order.
Perplexity sonar-pro with return_citations=True, Google AIO via DataForSEO, ChatGPT web search. Five iterations per query per platform, matching E016.
Tier 2 Perplexity citation rate post-intervention minus pre-intervention. Positive result: lift above 22.0pp combined threshold. Null: lift at or below threshold.
T1 Perplexity rate change (should stay at the E016 baseline of 80%), T2 AIO and ChatGPT rates (expected 0% floor from E016), and Tier 2 mention rate across all platforms.
No content changes to thegeolab.net between intervention and M1 measurement. No new posts at all on the query-matched pages. Posts unrelated to the query set are permitted and expected. The intervention window is the JSON-LD graph, the Wikidata items once accepted, and the external identifiers — nothing else.
What Would Falsify This
Four outcomes would constitute clear falsification or partial falsification of the hypothesis:
- T2 Perplexity stays at 0% on both M1 and M2. Strong falsification. The authority gate holds against the full bundled intervention. Publishable as a Layer 4 null, companion to EDX.
- T2 Perplexity lifts but stays below 22.0pp combined on both measurements. The intervention produced a sub-threshold shift indistinguishable from platform variance. Reported as null with the raw numbers shown. Not claimed as positive.
- T1 Perplexity drops below the E016 baseline of 80%. An unexpected secondary effect. Either a confound or an indication the intervention hurt an existing signal. Triggers investigation before M2.
- Wikidata items are all deleted under notability criteria. A valid outcome on its own — the intervention was not deployable at this stage. Reported as a deployment-failure null, distinct from a citation-rate null.
Known Confounds and Honest Limits
Three limits apply to whatever this experiment produces. They are flagged here rather than after the data because their presence shapes what a positive or negative result can and cannot support.
The intervention is bundled. A positive result cannot attribute the lift to any single component — ORCID, Wikidata, public repos, JSON-LD graph. The follow-up experiment (E026, not yet designed) would need to hold some components constant and add others sequentially. The current design answers the yes-or-no question, not the which-component question.
Indexing latency is unknown. Wikidata items that are accepted still take time to propagate into the downstream knowledge graphs that AI platforms draw on. If M1 shows no lift and M2 shows a lift, the finding is positive but the timing tells us nothing clean about when the signal reached the platform. Two months of measurement is the minimum honest window, not a precise one.
The authority gate is measured, not modelled. EDX established the gate as an empirical regularity on this site at this point in time. It does not provide a model of how the gate clears. The intervention is a bet on the gate being partially clearable by external identifiers — a bet grounded in existing entity-disambiguation literature, but still a bet. A null result does not rule out the gate clearing for a different combination of signals. It rules out this one.
What Happens Next
Three things are scheduled. None of them involves changes to this page or the measurement query set.
Wave 2 deploys the three Wikidata items in sequence: Person first, then The GEO Lab, then The GEO Stack. Each submission is timestamped separately in the E025 Notion row so that, if a lift shows up, the measurement data can at least be compared against the Wikidata acceptance timeline — even though the design does not formally isolate it. Wave 3 aligns LinkedIn and GitHub profile metadata.
M1 measurement runs as part of E014’s late-May cycle. M2 runs as part of E014’s late-June cycle. The results will update this post to version 2.0 with the primary outcome, all secondary outcomes, and the interpretation under the four falsification criteria above. If the result triggers E026 (component decomposition), the design will publish separately before running.
If the prediction turns out to be wrong — if T2 Perplexity lifts above 22.0pp — the post updates and the site publishes a positive Layer 4 finding. If the prediction is right, the site publishes a null, and E025 joins EDX in the failure registry. Both outcomes are the same amount of work to report. Only one of them is comfortable to write.
Key Takeaway
Pre-registering a hypothesis before the data arrives is the only defence against quietly rewriting it to fit the measurement. This post fixes E025’s null prediction — T2 Perplexity citation rate stays at 0% on M1 and M2 — before five weeks of waiting begins.