Methodology
How Visdenda produces each number
The console answers one question per page: what does this URL need, and did the fix work? This page is the method behind that answer — the key, the grain, the window and the failure mode of each source that feeds the console, including the things the product deliberately cannot see. It is a method page, not a feature page.
Last reviewed 18 November 2026
One row per page
Every block is aligned to one normalised URL. The console does not join a query row to a session row directly; it aggregates each source to the page first, so twenty queries do not multiply one session count twenty times.
Drop the query string and the fragment.
Collapse duplicate slashes, add a leading slash, and drop a trailing slash except at the root.
Do not decode percent-escapes and do not change case: HTTP paths are case-sensitive, and %2F is not /.
Key rule: pageKeyOf in packages/contracts/src/page-report.ts. The backend applies lowercase and a 512-character cap only at ingest.
| Block | What lands on the row | Grain | Window / freshness |
|---|---|---|---|
| Search | Search Console page totals and up to 10 top queries by clicks. The row is built for the page, not for each query. | Query × page at Google; page totals and query rows on the row. | 7, 28 or 30 days, or a custom window up to 60. GSC finalises 1–3 days behind; retention is 16 months. |
| Traffic | Sessions, engaged sessions, pageviews, conversions and average time. The page beacon is read first; the GA4 page snapshot fills the block when the beacon has no rows. | Session × day for the beacon; landing page × snapshot window for GA4. | Beacon: the selected window. GA4: its own 7, 28 or 30-day top-20 organic landing-page snapshot, disclosed as the block's window; conversions come from that snapshot. |
| AI fetches | Hits, 404s, blocked paths and last-seen, grouped by agent. Edge rows are used only when the probe has no rows for the window. | Agent × path × day. | Selected window; telemetry is retained 60 days. Needs a verified site; probe rows first, with edge rows as a fallback. |
| AI referrals | Beacon sessions whose referrer host is an AI surface, plus the count with a prior fetch on the same page. | Session × landing page × day; the prior-fetch link covers the session day and the two UTC days before it. | Selected window. Referral sessions need the page beacon; the prior-fetch count needs the crawler probe too. |
| Index | Verdict, coverage state, last crawl and canonical drift for the URL. | URL × most recent URL Inspection. | Point-in-time GSC URL Inspection; not a live index check, and no indexing guarantee. |
| Audit | Open and failed findings for the URL, up to five rule rows, and field LCP and CLS p75 with the sample count. | Rule × URL × run; CWV sample × page × day. | Latest completed run. CWV p75 needs 20 samples per metric. |
Every block carries its own na reason. A missing source reads na, not 0: zero is a measurement; na means the source was not connected or had no row for that page.
Tracked keywords do not land on the page row
A tracked term is stored per site, term, locale and device. The page row has no keyword column; keyword tracking is its own site-level surface.
Competitor changes do not land on the page row
A competitor event belongs to the competitor's URL. It is shown beside your work, not joined onto your page. There is no competitor traffic, backlink or AI-citation estimate to join.
The per-page join is documented in the console guide.
Each source, and what it cannot tell you
A source is only useful if its failure mode is visible. These are the sources this page covers: the blocks behind the page row and the two site-level surfaces, with the window each one reads and the question it cannot answer.
| Source | What it gives | Window / retention | What it cannot tell you |
|---|---|---|---|
| Search Console | Clicks, impressions, CTR and average position with the queries and pages behind them. URL Inspection adds index verdict, coverage, last crawl and canonical drift. | 1–3 day lag; 16-month retention. Position is the average for the selected window. | Not sessions, conversions or revenue. Large properties may be sampled. Search is search-side only; URL Inspection is not a live index check. |
| GA4 and the page beacon | GA4 sessions, engaged sessions and conversions when it is bound. The first-party beacon adds pageviews, referrers, duration and field vitals, and fills the Traffic block when GA4 has no rows. | GA4 page detail is a 7, 28 or 30-day top-20 organic landing-page snapshot, and conversions come from that snapshot. The beacon covers the selected window. | Pages outside GA4's top 20 read na. Consent and thresholding can block or hide rows. No cross-device identity, ad attribution or full-funnel revenue attribution. |
| Audit crawler | What a crawler meets: status, redirects, canonical, robots directives, titles, metas, H1, structured data, links and images. Rules turn those values into findings. | The page set of one run. The base crawl reads the no-JS response; a render pass is compared where one is available. | Cannot see Google's index status. It does not ingest server or CDN logs, does not crawl the open web, and honors robots.txt: a blocked URL is reported, not fetched. A 200 with an error page body is still a 200. |
| AI crawler telemetry | Fetches by agent, path and day, with 404s, blocked paths and machine-file reads. It records requests a page beacon cannot see because AI crawlers do not run JavaScript. | Selected window, up to 60 days; retained 60 days. Needs a verified site; crawler rows come from the server probe, with edge rows as a fallback. | A fetch is not a citation. A user-agent string is evidence, not identity: it can be spoofed, and an unknown agent is a lead. It sees only requests that reached your site. No per-answer attribution. |
| Competitor monitoring | Public sitemap changes (added, removed, lastmod), key-page content changes, structural coverage tiers and on-demand tracked-term SERP snapshots. | Free polls every 168 hours, Pro every 24 hours. Events are retained 90 days; 50 per response by default, 100 maximum. | No competitor traffic, sessions, revenue, backlinks or keyword gap. No competitor AI citations or crawler comparison. Public sitemaps only: a missing entry is not proof that nothing changed. The first poll is a silent baseline. |
| AI visibility sampling | Whether Google shows an AI Overview for a tracked keyword and whether the answer references your domain, with the cited URLs. | Paid sites rotate five tracked terms a day from the top 25 by 28-day GSC impressions; manual checks run on demand. US/English locale. | Not a census of AI answers, not a rank and not a share-of-model score. It does not query ChatGPT, Claude or Perplexity, and it does not compare competitor citations. |
A missing source is not a zero. The console returns na with a reason — no source, no probe, GSC unbound, no data, not indexed yet, or outside the GA4 top 20 — so “not connected” and “measured zero” never collapse into the same number.
Deterministic findings, model-written words
Findings come from rules. Words that a model writes have one narrow, visible place — an external agent's proposal — and no route into a verdict.
A finding is a rule result
Each finding is a rule id, URL and run, with a severity frozen at write time and a status that can later resolve. The rule carries the check, its version and an official source; explain_finding is a lookup of that rule, not generated prose.
The audit makes no AI calls
A round of the audit and its read surfaces call no AI model. The console does not generate AI advice or an AI-written explanation: findings give the rule and its source, and the review is yours.
A model can propose; it cannot decide
An external agent using MCP propose_fix can submit a fix proposal. The text is agent-authored, stored as pending for seven days, deduplicated, and never applied automatically. A human approves or rejects it; approval authorizes, it does not execute.
No model writes a verdict
Verdicts, scores, severities and statuses come from code paths and stored rule output. No language model sets a finding status, a health score or an effect verdict.
Rule reference: /docs/checkers. Machine surface and proposal rules: /docs/mcp.
Before and after
A fix is checked at three levels, and each level proves a different amount. None of them turns a page edit into a ranking guarantee.
28 d
Baseline
Days before the change
14 d
Verify
Starts on change day +3
≤5
Control pages
Same site, no change in window
>20%
CTR signal
Or >2 positions; exactly 20% is no-signal
1. Single-page re-check
After an edit, re-check the exact URL against the rule. A pass closes the open finding as fix-applied; a fail records the verdict and leaves it open. The re-check does not create an audit run or consume an audit quota.
2. Re-audit the page set
Re-crawl the same pages and compare run over run: added, fixed and regressed findings per URL. A fix only counts when the re-check reaches the URL again; one run is a snapshot, not a trend.
3. Measure the effect window
For a detected content change, the console compares GSC page windows: 28 days before the change against 14 days starting three days after it, then waits three more days for GSC to finalise. Fewer than 100 impressions in either window is insufficient.
Controls are same-site pages whose baseline average position is within 3 positions, whose baseline impressions are 0.5–2× the target's, and which had no change event in the window; at most five are used, chosen by closest impressions. Net Δ = target Δ − median control Δ. Thresholds are a relative CTR change above 20% or a position change above 2.0 positions.
Verifiable fields are title, meta, H1 and JSON-LD. A body-only change records an event but does not open a verification.
A re-audit proves the finding is gone from the page we fetched. It does not prove Google indexed or ranked it; index and ranking stay GSC reads with their own lag.
The control comparison is the only difference-in-differences estimate in the product. It is GSC page-level only, covers detected content changes, and does not extend to sessions, revenue, AI answers or a change the fingerprint did not see. Without a control page a positive or negative result is directional-improved or directional-degraded and is labelled as directional; a flat result stays no-signal. Even with controls it is a comparison, not proof of causation.
What this page does not measure
These omissions are deliberate. A product that claims all of them is either estimating or guessing.
- Search Console sessions, conversions or revenue. Search Console is search-side only.
- Google index status inferred from a crawl. Index status comes from URL Inspection, not from fetching a page.
- A census of AI answers, an AI rank or a share-of-model score. Sampling covers Google AI Overviews for tracked keywords in a US/English locale.
- ChatGPT, Claude or Perplexity citations. The sampler does not query those engines.
- Competitor traffic, revenue, backlinks, keyword gaps or AI citations. Competitor monitoring reads public sitemaps and page snapshots.
- Server-log or CDN-log completeness. Crawler telemetry sees probe rows, with same-path edge rows as a fallback, not every request.
- Cross-device identity, ad attribution or full-funnel revenue attribution.
- A daily rank grid. Average position is an average across the selected window.
- Causal proof for an arbitrary change. The effect estimate exists only for detected content changes on GSC page windows.
- Real-time data. GSC finalises 1–3 days behind, and GA4 page detail is a snapshot.
Verify us
The method is public where it can be. Check the outputs, not the claims.
Public sample report
We ran the public checks on ourselves
Eight sampled pages, all 206 sitemap URLs, robots.txt, llms.txt and JSON-LD, checked on 18 November 2026, with the fixes still owed.
Open /me-reportCrawler directory
The crawler directory method
A vendor source for every entry, network data only when published, and a 90-day re-review clock for each entry.
Open /crawlersMachine surface
The same evidence over MCP
Findings, rule definitions, page reports, Search Console reads and telemetry, with quota and evidence metadata on each result.
Open /docs/mcpRun the method
Check the console against a site you own
Free includes one verified site, the page dossier and single-page fix re-checks.