The model
Focus, outliers, and the 3 signals that drive drift
We explain drift using a simple framework:
1) Site Focus
How tightly your site stays centered on its core topic(s). Measured by actual semantic distance from your topical center.
2) Outlier Pages (Site "Radius")
Pages that drift far from your core topic and weaken overall topical coherence. Default review threshold: distance ≥ 0.700.
3) The "ABC" signals (what we triangulate)
A — Anchors
Internal link anchors + link context describe what your pages mean to search engines.
B — Body
Your headings + main content represent what the page is actually about.
C — Clicks & queries (GSC reality)
GSC shows what the page is earning traffic/impressions for (intent in the wild).
Drift shows up when these signals disagree (e.g., anchors imply Topic A, content is Topic B, and GSC queries trend Topic C).
Where “ABC” comes from (and how we use it):
“ABC” (Anchors, Body, Clicks) appears in public reporting and DOJ trial exhibits discussing Google’s topicality / base relevance components.
We borrow the high-level mnemonic to explain our framework, but we adapt C to mean your connected
Google Search Console queries + clicks (first-party performance data), not Google’s internal click-satisfaction signals.
Extraction
Main-content extraction (why it matters)
Raw HTML includes navigation, footers, "related posts," and template noise. If you embed all of that, your vectors
become a representation of your site template — not your topic.
We aim to keep
Primary article/body, headings (H1-H6), lists, key supporting text, main containers.
We aim to reduce
Nav/footer blocks, boilerplate, repeated CTAs, unrelated "widgets," template repetition.
Result: Embeddings reflect page meaning, not layout repetition. This makes clustering more accurate and drift detection more precise.
Embeddings
How embeddings are used
An embedding is a high-dimensional numeric vector that represents semantic meaning. We use embeddings to:
- Measure similarity between pages (semantic neighborhoods)
- Group pages into clusters (topic hubs via K-means)
- Project pages into 2D space for visualization (via UMAP)
- Compare internal-link context meaning vs destination page meaning
- Calculate actual semantic distance from your site's topical center
similarity(pageA, pageB) = cosine(embeddingA, embeddingB)
distance = 1 - similarity
Embedding model: We generate embeddings using a modern semantic embedding model (model choice may evolve; scoring remains cosine similarity in embedding space).
For the same input and model version, embeddings are generally consistent, and all similarity/distance scoring is computed in the original embedding space.
You don't need the math to use the output — the report translates it into actions.
Clustering
How clustering reveals topical structure
We use K-means clustering on page embeddings to group semantically similar pages.
This reveals:
Topical hubs
Your site's "neighborhoods" of related content (e.g., "Interior Painting," "Exterior Services").
Overlap / cannibalization
Multiple pages competing inside the same intent space.
Content gaps
Clusters with only 1-2 pages indicate under-developed topics.
Drift reference frame
Clusters define "normal" — pages far from their cluster centroid are drifting.
cluster_drift = 1 - cosine(page, cluster_centroid)
Automatic optimization: The tool can evaluate multiple cluster counts (e.g., 3–10) and select a strong option using separation metrics (such as silhouette score), depending on your configuration.
Drift score
How we calculate drift (conceptually)
Drift isn't one number from one signal. We triangulate multiple signals, then summarize them into a score and a "why."
Semantic drift (primary)
Distance from page → cluster centroid (or global centroid in actual mode).
drift_sem = 1 - cosine(page, centroid)
Intent drift (GSC)
Queries the page earns vs the cluster's dominant intent set.
drift_intent = divergence(queries, cluster_queries)
Link structure penalty
Pages with few internal links get penalized (orphan detection).
link_penalty = 1 - (total_links / max_links)
Engagement penalty
Low-traffic pages may indicate poor targeting or drift.
engagement_penalty = 1 - (clicks / max_clicks)
About SDI: SDI is a prioritization score. By default it blends semantic drift with structural signals (internal linking + engagement).
Intent drift from GSC is reported as a separate diagnostic (“why this page is drifting”) and can be optionally weighted in custom scoring.
SDI = (60% × semantic_drift) + (30% × link_penalty) + (10% × engagement_penalty)
The report always includes why the score is high (e.g., "ranking shifted toward X intent" or "link context pushes toward Y cluster" or "isolated from semantic neighbors").
Scan frequency
Unlimited scans (fair use) — scan when it makes sense
Drift is time-based and change-driven. You need flexibility to scan when you publish major updates,
test fixes, or track progress — not when an arbitrary calendar limit says so.
Common scanning patterns:
Monthly maintenance (most common)
Scan once per month to catch drift, monitor cluster health, and identify new issues.
Good for stable sites with regular content updates.
Weekly during projects
Scan weekly when doing major content updates, site redesigns, or fixing drift issues.
Track improvements in real-time with actual distance mode.
Before/after testing
Run baseline scan → implement fixes → rescan days later to measure impact.
See if content changes reduce semantic drift (lower actual distances).
Quarterly audits
For mature, stable sites, quarterly scans may be enough.
Good for tracking long-term trends and catching slow drift.
Track improvements over time:
With actual distance mode, you can measure whether your fixes reduce semantic drift:
- Run baseline scan → note distances for problematic pages (e.g., page at 0.680 distance)
- Implement fixes (content updates, link changes, heading adjustments)
- Rescan after changes are indexed (7-14 days)
- Compare distances → page now at 0.485 = 28% improvement in topical alignment
What "fair use" means:
Fair use includes
- Scanning whenever you publish major content
- Weekly scans during site updates/redesigns
- Before/after testing to measure fixes
- Ad-hoc scans when traffic drops unexpectedly
- Comparing different internal linking strategies
Not fair use
- Automated high-frequency scanning (hourly/daily)
- Reselling or redistributing scan access
- Abuse or attempts to overload the system
- Commercial scraping or data harvesting
- Running scans on sites you don't own/manage
Bottom line: If you're using the tool for its intended purpose (improving your site's topical focus),
scan as often as your workflow requires. We don't police normal usage — the fair use policy exists to prevent abuse,
not to restrict legitimate analysis.