Exp-008 Planned D Activates 2026-11-15
Does author consistency improve entity confidence?
Research question Does maintaining consistent author identity signals (same photo, bio, titles, and company) across multiple domains increase AI systems' confidence and correctness when answering questions about the author-organization relationship?
Motivation
Author consistency is widely recommended in quality/E-E-A-T guidance for human search systems. Its effect on LLM entity confidence is empirically under-documented in public experiment form.
Facts
- The same person can appear with divergent titles, bios, and images across domains.
sameAsand identicalPersonproperties are common consistency mechanisms.- Primary public surfaces for this subject: barbu.es/about, LinkedIn, GitHub, InfoWebPlus.
Assumptions
- “Entity confidence” can be proxied via: answer decisiveness, inter-model agreement, contradiction rate, and willingness to state the affiliation without hedging - while treating hedging as informative, not purely negative.
Unknowns
- Whether photo consistency matters for text-only LLM probes (may matter more for multimodal or KG pipelines).
- How many domains are required for a detectable effect.
Identity kit under test (intended)
| Signal | Consistency target |
|---|---|
| Photo | Same canonical headshot URL or identical file hash lineage |
| Bio | Same short bio string (±minor locale tweaks logged) |
| Titles | Controlled vocabulary (e.g. Founder) |
| Company | Canonical organization string (InfoWebPlus) |
Subject: George Barbu, Founder of InfoWebPlus, with related property barbu.es.
Hypothesis
H₀: After aligning photo/bio/title/company signals across domains, AI confidence proxies and affiliation accuracy do not improve vs baseline inconsistency.
H₁: Consistency alignment increases affiliation accuracy and reduces cross-domain identity contradictions in AI answers.
Methodology
Design type
Before/after consistency remediation across a declared domain set.
Baseline deliberately documents inconsistencies as Facts (do not “clean” before capturing baseline).
Environment
| Dimension | Value |
|---|---|
| Primary domains | Multi-domain set including barbu.es and InfoWebPlus properties |
| Surfaces under test | Author boxes, About pages, schema Person, social profiles |
| AI systems probed | ChatGPT, Claude, Gemini, Perplexity |
| Measurement window | Baseline inventory; post-alignment ≥ 30 days |
| Geographic / language scope | English (+ note other locales) |
| Tools used | Identity inventory sheet, image hash log, probe rubric |
Variables
Independent variables
- Cross-domain consistency level (low → high) across photo, bio, titles, company
Dependent variables
- Affiliation accuracy (“works at / founded”)
- Contradiction rate across platforms in a single answer
- Hedging frequency (coded, not automatically “bad”)
- Inter-model agreement on title + company
Controlled / held constant
- No new major PR campaigns during the window if avoidable
- No change to unrelated entities’ names
Confounds to monitor
- Schema deployment (Exp-002) shipping simultaneously
- Social network graph changes
- Homonym risk
Procedure
- Inventory all public author appearances; score consistency 0-1 per signal.
- Run baseline probes emphasizing identity questions.
- Remediate to the identity kit; log each URL changed.
- Re-probe; compare DVs; report per-signal ablations if resources allow (bio-only vs photo+bio, etc.).
Prompt protocol
Who is George Barbu? Where does he work?
What is George Barbu’s role at InfoWebPlus?
Are there conflicting identities for George Barbu online? What conflicts do you see?
Provide a short author biography for George Barbu suitable for a research byline.Results
Status: Planned. No results claimed.
| Metric | Baseline | Post-alignment | Notes |
|---|---|---|---|
| Affiliation accuracy | n/a | n/a | Not collected |
| Contradiction rate | n/a | n/a | Not collected |
| Inter-model agreement | n/a | n/a | Not collected |
Observations
- Pending identity inventory.
Limitations
- Confidence is latent; proxies are imperfect.
- Photos may be irrelevant to text-only models.
- Self-authored bios across domains are not independent evidence.
Future Work
- Multimodal probes with portrait images.
- Controlled inconsistency injection on a sandbox persona (ethics review).
References
- schema.org/Person properties: image, jobTitle, worksFor, sameAs
- Canonical identity kit: about, LinkedIn, GitHub, resume, portfolio, InfoWebPlus
- JSON-LD and Google structured data intro
- Lab Exp-002, Exp-003, Exp-006
Replication Notes
Store before/after screenshots. Do not erase historical inconsistencies from archives - measurement needs them.
Status: Planned
Author: George Barbu, Founder of InfoWebPlus