Observe the site
Crawl pages and inspect recorded links, schema entities and unresolved identities.
SCHEMAGRAPH · SITE GRAPH METHODOLOGY
Evidence-Weighted Authority Simulation
EWAS connects observed website structure with a reviewable internal-link implementation decision.
By Jason Gibson · Holistic Growth Marketing LLC · Published October 26, 2026
Start with a public schema audit. The Internal Authority Model and EWAS workflow require Professional access.
Read the featured HGM methodology article: Evidence-Weighted Authority Simulation
Abstract
Internal-link analysis often stops at a score or a diagram. The implementation decision requires more context: whether a source page belongs in the same topic, which assumptions produced the score, how a proposed link changes the distribution across the recorded site, and whether the recommendation can be reviewed before publishing.
Evidence-Weighted Authority Simulation (EWAS) is the workflow in SchemaGraph that brings those questions together. It compares an observed internal-link network with a proposed-link scenario, optionally blends page-level historical evidence into the model’s restart distribution, and preserves the assumptions alongside the output. This paper documents the implementation, not a new Google ranking factor or a causal traffic forecast.
How this paper connects to SchemaGraph
SchemaGraph is the application implementing the workflow described here. Its Site Graph visualizes page and entity relationships in 2D/3D, while its audit evidence, identity review and schema-repair workflows provide the surrounding context. EWAS is one Professional capability inside that product, rather than a separate application or a replacement for the schema audit. See the complete feature inventory and plan limits for the surrounding platform scope.
The SchemaGraph product page on HGM explains the broader structured-data and entity-resolution use case. This paper owns the narrower technical question: how a proposed internal link is evaluated against a recorded network and explicit assumptions. The companion HGM article on Evidence-Weighted Authority Simulation is a related methodology destination. Each page has a distinct role: product scope, technical method and company context.
Compare a scenario
Select a target, review source candidates and model proposed links without editing the site.
Review the action
Export assumptions and proposed implementation work, then verify changes through a later audit.
In the application, open Site Graph → Internal Authority Model → Internal Link Simulator. Choose the internal-link layer when evaluating PageRank-inspired authority; a semantic relationship is not automatically an HTML hyperlink.
What EWAS means
Evidence-Weighted Authority Simulation is HGM’s name for the workflow in SchemaGraph that tests proposed internal links using a recorded link network, observed entity relationships and optional historical performance evidence. Authority here means relative share within an internal-link model. It is not attribution of conversions, Google PageRank, or a prediction of rankings.
EWAS is a SchemaGraph feature, published by Holistic Growth Marketing LLC. The name identifies this implementation and workflow; it does not claim invention or exclusive ownership of weighted or personalized PageRank. Trademark clearance and legal exclusivity have not been established.
Two graphs, connected through pages
The link graph represents directed hyperlinks between sampled pages. The semantic graph represents observed entities and relationships. Shared resolved entity IDs can suggest relevant source pages. Generic organizational and website nodes, unresolved identities and inferred nodes are excluded from candidate matching.
Semantic overlap is shared topical entity membership divided by the union of memberships. It is not embedding similarity or an editorial judgment. For each entity group, EWAS can sum the modeled shares of its member pages before and after a scenario. Groups may overlap; their shares must not be added together.
How historical evidence changes the scenario
Upload a page-level GSC Pages export or GA4 Landing page CSV/JSON. Select a supported metric: impressions, clicks, sessions or key events. The interface reports matched pages, exclusions, duplicate handling, available reporting dates and channel scope. Counts remain supplied evidence, not independent verification of tracking accuracy.
The restart prior is a blend: (1 − historical influence) × the selected model’s prior + historical influence × normalized historical counts. The selected model starts with a uniform prior or imported followed referring-domain counts. A slider sets historical influence from 0% to 100%; no coefficient is fitted from outcomes. The same prior redistributes dead-end mass.
Metric and period matter. High impressions may indicate demand; key events reflect configured business outcomes. Neither establishes backlink authority or causal search importance. Missing rows do not prove zero demand. GA4 without a channel dimension inherits the uploaded report’s scope. New pages may be disadvantaged by strong historical weighting.
A worked example: separating history from the link intervention
This is a hypothetical arithmetic example, not a customer result. Suppose a recorded three-page network contains a service page, a practical guide and a reference page. Their selected historical metric has counts of 60, 30 and 10 for the same reporting period and scope. Normalizing those counts gives historical shares of 0.60, 0.30 and 0.10.
With a uniform starting prior and historical influence of 40%, the service page receives a restart weight of (0.60 × 1/3) + (0.40 × 0.60) = 0.44. The guide receives 0.32, and the reference page receives 0.24. These weights sum to one. They are the restart distribution; they are not the final PageRank-inspired scores, traffic shares or measured ranking probabilities.
Now propose an editorially relevant link from the guide to the service page. EWAS calculates the original link network and the modified network using that same restart distribution. The difference between the two results isolates the modeled effect of the proposed edge under the selected assumptions. Comparing a historical model against a different nonhistorical model would mix two interventions and make the link’s effect harder to interpret.
Adding the link can also dilute the guide’s modeled contribution to its other outgoing destinations, because the guide’s outgoing weights are normalized. Review gains and losses across the graph instead of reading the target page’s gain alone. The actual resulting shares depend on every recorded link, placement weight, dead end and restart assumption; this paper does not invent numerical scenario results without that complete graph.
The entity layer answers a separate question: does the guide share resolved topical entity evidence with the target? Shared membership can justify reviewing a candidate, but the placement still needs to make sense to a reader. A mathematical improvement cannot substitute for a useful anchor, accurate destination and sensible editorial context.
Simulate a concrete implementation plan
- Choose a target page and review source candidates supported by shared entity evidence.
- Add up to five proposed links, choosing their intended HTML placement.
- Calculate the original graph and scenario with identical historical and backlink assumptions.
- Inspect modeled gains and losses across pages and entity groups.
- Preview hypothetical edges in Site Graph and export the implementation list with assumptions and evidence.
Candidate ordering combines entity overlap with modeled target-share change; at most twelve candidates are evaluated. This is a bounded heuristic, not a global optimization guarantee. Existing source-to-target links are not proposed again. Manual proposals require editorial review.
Graph Series evidence and handoff
Load same-domain SearchGraph and CopyGraph snapshots from the authenticated Import Center. Dated SearchGraph performance facts can supply the historical prior; crawl status, canonical differences and CopyGraph editorial findings remain attributed review context. Historical aggregation prefers page totals over query detail and consolidates duplicate facts before summing dates.
Export the scenario using graph-series/1.0, or save it to Import Center to use configured receiving links. The package includes scenario assumptions, proposed recommendations and evidence provenance. Delivery status is reported separately from saving. Review SchemaGraph’s data-handling disclosures before exporting or sharing historical performance evidence. Importing the package does not implement links or establish a new billing entitlement.
When to use EWAS, and what belongs elsewhere
Use EWAS when you have a defined target page and want to compare a small number of reviewable internal-link proposals. It can support a discussion about a new service page, a supporting guide that lacks relevant inbound links, or a cluster whose observed linking does not reflect its intended content relationships. Candidate quality depends on the crawl and resolved entities available for that website.
Use SchemaGraph’s page evidence and schema-repair tools for missing markup, conflicting IDs and schema-to-content discrepancies. Those are implementation and identity questions; a higher internal authority share does not correct an inaccurate schema claim. Use your search and analytics evidence to evaluate observed outcomes after publishing, and preserve the implementation date so later changes can be interpreted in context.
For the wider operating model, explore HGM’s Graph Series. Within this workflow, SearchGraph supplies compatible search-performance history and review signals, while CopyGraph contributes content and governance context. SchemaGraph owns the EWAS calculation. A connection transports evidence; it does not make the source measurements interchangeable or establish causation.
Math, weights and convergence
EWAS uses a PageRank-style random walk with damping 0.85. Each source distributes its share proportionally to its eligible outgoing weights. Default relative placement assumptions are content 1, navigation 0.25, header/footer 0.1, aside 0.5, unknown 1. Repeated links use the strongest eligible placement rather than accumulating votes.
Those weights are transparent assumptions, not measured Google values. Baseline mode gives links equal weights. The solver runs for at most 500 iterations and reports convergence using an L1 iteration-change tolerance of 1e−10. Unconverged scenarios are labeled and cannot be previewed as completed calculations.
Sampling, evidence and implementation limits
Calculations cover the recorded crawl. At most 200 unique internal targets per page and 10,000 edges per audit are retained. Compacted source HTML may omit links; this is not a full rendered-browser DOM analysis. Uncrawled targets, external links, canonical consolidation and redirect resolution are not modeled. Rescan older audits to collect link-placement evidence. The platform specification table brings these limits together with input and output requirements.
GSC/GA4 imports are browser-local and cleared on reload or audit change. CSV/JSON exports can retain scenario provenance; exporting does not publish links or change audited pages. Duplicate historical rows use maximum values rather than sums, so upload one aggregate row per page for the intended period. Referencing canonical entities in schema does not guarantee AI visibility.
Validation and prior work
Verification includes analytical stationary distributions, independent matrix-solution comparisons, probability conservation, non-convergence reporting, malformed-import rejection and checks that scenario changes preserve the original audit. These tests validate implementation behavior, not improvements in rankings, traffic, AI citations or revenue.
PageRank’s foundational work is credited to Lawrence Page, Sergey Brin, Rajeev Motwani and Terry Winograd. Topic-sensitive PageRank is associated with Taher Haveliwala; edge-weighted personalized PageRank research includes Wenlei Xie, David Bindel, Alan Demers and Johannes Gehrke. WebKnoGraph, by Emilija Gjorgjevska, Georgina Mirceva and Miroslav Mirchev, is related research on authority redistribution and semantic coherence. HGM’s contribution is the SchemaGraph implementation and workflow, not ownership of those foundations. Relevant references include Haveliwala’s topic-sensitive PageRank paper, edge-weighted personalized PageRank research, WebKnoGraph research on link interventions and semantic coherence, NetworkX’s PageRank reference, GSC reporting limitations, and GA4 metric definitions.
Real-world evaluation requires compatible reporting periods, an implementation record and consideration of content changes, seasonality and other confounders. No causal lift study is claimed here.
What to record before and after implementation
Before changing a link, retain the audit timestamp, crawl coverage, target and source URLs, proposed anchor and placement, model variant, historical metric and reporting period, influence setting, convergence result and export package. Confirm that relevant URLs and canonical identities represent the intended pages. A partial crawl or inconsistent URL identity can change the network enough to affect the comparison.
After editorial approval, implement the reviewed links through your normal publishing workflow. Recrawl to verify that the links exist in the observed HTML, then compare appropriately scoped performance reports over a suitable period. Content edits, indexing changes, seasonality, external links and marketing campaigns can all affect observed outcomes. A before-and-after difference alone cannot prove that the simulated link caused a change.
This evidence record makes the recommendation reproducible and open to challenge. It also makes it easier to reject a proposal when its modeled gain is too sensitive to a placement assumption, missing pages or an unusually strong historical prior.