View original page

Read the evidence behind Beacon’s metrics

Beacon

Check the answers, calculations and evidence behind each measurement.

Inspect the worked exampleExplore team workflows →

Illustrative question pair

Discovery: Which tools help a small team create reusable prompts from its product documents? Recognition: What does the named business and website do? Save both answers when useful. Only the first can measure whether the model names the business without being given its name.

Read the full example →

Replay the calculation and the action plan

Inspect four labeled observations, then follow an empty HTML document into its AEM actions. The downloadable replay runs the production calculations and check-to-action mapping locally. Every observation is a synthetic test input.

Reproducible example

Turn recorded findings into an accountable action plan

Synthetic answer labels and an authored empty HTML document on example.test. All numbers and actions are actual outputs of the production measurement and plan functions; none describes Prompeteer’s live visibility.

Checked 19 October 2026 · product-workflows-v2 · no model calls

  1. Measure Three usable answers out of four planned. One failure remains a failure, rather than becoming a negative answer.
  2. Separate One of two unbranded answers names the business. The branded answer uses a separate denominator.
  3. Act The empty document produces specific actions with source evidence, owners and acceptance checks.

Inspect inputs and recorded output

{
  "input": {
    "plannedSlots": 4,
    "observations": [
      {
        "slotId": "unbranded-a",
        "questionId": "U01",
        "layer": "web_grounded",
        "status": "answered",
        "mention": "named",
        "citation": "own_domain"
      },
      {
        "slotId": "unbranded-b",
        "questionId": "U01",
        "layer": "web_grounded",
        "status": "answered",
        "mention": "absent",
        "citation": "other_source"
      },
      {
        "slotId": "unbranded-c",
        "questionId": "U01",
        "layer": "web_grounded",
        "status": "failed",
        "mention": "unassessed",
        "citation": "unassessed"
      },
      {
        "slotId": "branded-a",
        "questionId": "B01",
        "layer": "web_grounded",
        "status": "answered",
        "mention": "named",
        "citation": "own_domain"
      }
    ],
    "url": "https://example.test/",
    "html": "<!doctype html><html lang=\"en\"><head><title>Fixture export tool</title></head><body><div id=\"root\"></div></body></html>",
    "rubricVersion": "2.0.0"
  },
  "output": {
    "completed": "3 / 4",
    "searchMentions": "1 / 2",
    "searchMentionRate": 50,
    "failed": 1,
    "brandedMentions": "1 / 1",
    "actions": [
      {
        "id": "P1",
        "title": "Content present without JavaScript",
        "evidence": "0 words in <body>",
        "owner": "engineering",
        "steps": [
          "Server-render meaningful body text into the initial HTML.",
          "Compare the HTML response with the rendered page.",
          "Include sufficient useful context; avoid word-count filler."
        ],
        "acceptance": "P1 passes on a re-run"
      },
      {
        "id": "P2",
        "title": "Not an empty SPA shell",
        "evidence": "<div id=\"root\"> is empty and the document holds only 0 words",
        "owner": "engineering",
        "steps": [
          "Render the meaningful public content into the initial HTML.",
          "Check the page with JavaScript disabled.",
          "Keep the visible content consistent across supported clients."
        ],
        "acceptance": "P2 passes on a re-run"
      },
      {
        "id": "A4",
        "title": "Structured content present",
        "evidence": "no list or table",
        "owner": "content",
        "steps": [
          "Use lists for genuinely sequential or parallel information.",
          "Use tables for comparisons with consistent criteria.",
          "Keep qualifications and units beside the facts they describe."
        ],
        "acceptance": "A4 passes on a re-run"
      },
      {
        "id": "A7",
        "title": "Sentences are readable",
        "evidence": "mean 0 words",
        "owner": "content",
        "steps": [
          "Express one clear idea at a time.",
          "Use concrete language suited to the intended audience.",
          "Keep essential qualifications attached to their claims."
        ],
        "acceptance": "A7 passes on a re-run"
      },
      {
        "id": "A8",
        "title": "Summary near the top",
        "evidence": "no heading followed by prose",
        "owner": "content",
        "steps": [
          "Place the main proposition near the start of the content.",
          "Give enough context to understand the offer.",
          "Review the first screen on mobile and desktop."
        ],
        "acceptance": "A8 passes on a re-run"
      },
      {
        "id": "E3",
        "title": "sameAs links",
        "evidence": "No public sameAs URL was declared in the structured data.",
        "owner": "content",
        "steps": [
          "Link only profiles that represent the same business.",
          "Check ownership, identity and destination accuracy.",
          "Remove stale or unrelated links rather than meeting a quota."
        ],
        "acceptance": "E3 passes on a re-run"
      },
      {
        "id": "E8",
        "title": "Social card complete",
        "evidence": "og:title no, og:description no, og:image no",
        "owner": "content",
        "steps": [
          "Match preview metadata to the visible page.",
          "Use an accessible image with suitable dimensions for target platforms.",
          "Test the resulting previews and correct stale information."
        ],
        "acceptance": "E8 passes on a re-run"
      },
      {
        "id": "R11",
        "title": "Self-referential canonical",
        "evidence": "no canonical link",
        "owner": "engineering",
        "steps": [
          "Confirm the preferred URL; keep intentional duplicates pointing there.",
          "Have your developer update the canonical in page settings or the HTML <head>.",
          "Publish, inspect the canonical URL in page source, and rerun the website check."
        ],
        "acceptance": "R11 passes on a re-run"
      },
      {
        "id": "P17",
        "title": "Mobile viewport",
        "evidence": "The homepage has no viewport meta tag.",
        "owner": "engineering",
        "steps": [
          "Apply the fix.",
          "Re-run the audit."
        ],
        "acceptance": "P17 passes on a re-run"
      },
      {
        "id": "P6",
        "title": "Content sits in a landmark",
        "evidence": "no main or article element",
        "owner": "engineering",
        "steps": [
          "Have your developer wrap the main copy in <main> in the page template.",
          "Keep site navigation and the site footer outside this container.",
          "Publish, confirm <main> contains the copy, and rerun the website check."
        ],
        "acceptance": "P6 passes on a re-run"
      },
      {
        "id": "P7",
        "title": "Content outweighs boilerplate",
        "evidence": "0% of body text",
        "owner": "content",
        "steps": [
          "Review the extracted main content alongside the rendered page.",
          "Reduce unnecessary repeated text without harming navigation.",
          "Add information only where it answers a real reader need."
        ],
        "acceptance": "P7 passes on a re-run"
      },
      {
        "id": "R10",
        "title": "Sitemap reachable",
        "evidence": "no sitemap.xml",
        "owner": "engineering",
        "steps": [
          "List the preferred public URLs in a valid sitemap.",
          "Reference the sitemap from robots.txt.",
          "Use actual modification dates and remove obsolete entries."
        ],
        "acceptance": "R10 passes on a re-run"
      },
      {
        "id": "R15",
        "title": "robots.txt names the sitemap",
        "evidence": "There is no robots.txt, so nothing names the sitemap.",
        "owner": "engineering",
        "steps": [
          "Apply the fix.",
          "Re-run the audit."
        ],
        "acceptance": "R15 passes on a re-run"
      },
      {
        "id": "R9",
        "title": "Response is compressed",
        "evidence": "no content-encoding header",
        "owner": "engineering",
        "steps": [
          "Enable supported compression for suitable public text responses.",
          "Check Content-Encoding and compare transferred bytes.",
          "Confirm that the decoded page content remains complete."
        ],
        "acceptance": "R9 passes on a re-run"
      },
      {
        "id": "G1",
        "title": "Markdown twin available",
        "evidence": "neither .md nor content negotiation",
        "owner": "engineering",
        "steps": [
          "Confirm which readers or tools would use a Markdown version.",
          "If useful, publish equivalent content and keep it synchronised.",
          "Advertise the supported format and test its content."
        ],
        "acceptance": "G1 passes on a re-run"
      },
      {
        "id": "G5",
        "title": "Link: alternate header",
        "evidence": "no Link rel=alternate header",
        "owner": "engineering",
        "steps": [
          "Identify maintained alternate formats or supported API descriptions.",
          "Publish the appropriate link relationship for each resource.",
          "Verify access rules and content consistency across formats."
        ],
        "acceptance": "G5 passes on a re-run"
      }
    ]
  }
}

Production functions: summarizeGrowthObservations, runChecks, actionPlan.

Download the input and result record, or replay all four examples locally with Node.js 20 or later.

Inputs and results (JSON)Offline replay (.mjs)File checksums

node replay-product-workflows-v2.mjs

The replay contains the relevant production functions. It needs no account, credentials, package install or network connection. Inspect the downloaded script before running it.

These four synthetic observations are a calculation test, not a visibility benchmark. The authored page tests check-to-action coverage. A plan does not prove a fix was implemented or that mentions, search position, conversions or traffic will improve.

Source: Downloadable inputs, outputs and checks

Check what completed before interpreting a rate

An AI visibility sample has a question panel version, company identity, market, language, model route, retrieval method and collection date. Transport success alone does not establish that a model answered the question. Preserve refusals, empty responses and failed retrieval as specific states.

Answer completion measures usable answers collected out of planned observations. Report completeness measures evidence available for each section. A report can collect every answer while still lacking selected competitors, sentiment assessments or comparable history.

Result How it enters the report
Substantive answer, brand absent Eligible for the mention denominator; the recorded mention value is zero.
Refusal or non-answer Captured response, but no substantive answer. Excluded from measured mention and recommendation rates.
Timeout, failed or missing Unavailable evidence. Show the reason and retry status; never count it as a negative answer.
Citation extraction unavailable The answer may be usable for mentions while its citation status remains unknown.
No denominator Not measured. Do not display 0%, infer neutrality or draw a chart point.

Keep the unit and denominator beside the number

Use one eligible observation per question, model and sample attempt selected for the report. Retries do not create extra votes. Display the numerator, denominator and excluded observations alongside percentages. Keep web-grounded and model-memory results separate.

Metric Calculation and evidence
Brand mention rate Substantive answers that name the verified business ÷ eligible substantive answers × 100.
Any-source citation rate Answers with at least one recorded source ÷ answers with assessable citation evidence × 100.
Own-site citation rate Answers citing an exact page on a verified company domain ÷ answers with assessable citation evidence × 100.
Source references All recorded citation references. Repeated references are not additional answers; distinct source pages are counted separately.
Recommendation position The explicit ordered recommendations in the answer. A mention or compatibility statement alone has no position.
Top-three among ranked mentions Named answers with explicit rank 1–3 ÷ named answers with an explicit rank × 100. Report unranked mentions separately.
Sentiment share Assessed brand mentions with the selected label ÷ all assessed brand mentions × 100. Show unassessed mentions beside the result.

Resolve the business and the cited page

Match the website, approved company aliases and the surrounding answer text. A short name shared by several businesses needs disambiguation. A technology provider, partner or cited website is not automatically a competitor. Competitive comparisons use the companies explicitly selected for the workspace.

Retain the original citation URL. Where the destination can be resolved, store the exact destination separately, remove known tracking parameters and fragments for comparison, and preserve query parameters that identify different content. An unresolved search redirect or retrieval placeholder remains unresolved. Do not substitute the company homepage for missing evidence.

Read the exact brand wording

Every assessed sentiment label needs an exact quote from a saved answer naming the business. Assess the wording about that business, not the tone of the whole answer or a competitor paragraph. Neutral means an assessed description without a clear positive or negative position; missing or ambiguous evidence stays unassessed.

Lead a small sample with its counts. “Two assessed mentions, both neutral” communicates more than “100% neutral” alone. These labels describe model answers, not customer opinion, market reputation or a probability of purchase.

Keep recognition out of discovery scoring

Stable unbranded questions test spontaneous discovery. Separate branded questions test whether the model can describe the named business accurately. Source-supplied product analysis tests a third condition: understanding the provided material. Never combine these into one discovery rate.

Illustrative question pair

Discovery: Which tools help a small team create reusable prompts from its product documents? Recognition: What does the named business and website do? Save both answers when useful. Only the first can measure whether the model names the business without being given its name.

Read website readiness as a versioned heuristic

The website audit applies a versioned rubric to recorded page checks. Passed and failed checks affect the applicable weighted calculation; unverified checks do not become failures. Informational declarations and not-applicable observations should not manufacture points. A report must retain its rubric version and check-level evidence so a later scoring change is distinguishable from a website change.

The public sample discloses a simplified set of demonstration weights and the complete arithmetic. Those teaching weights are not the production scoring rubric. Neither a readiness grade nor structured data establishes observed visibility or guarantees a citation.

Compare compatible snapshots and disclose gaps

Compare unchanged questions, market, language, model family and retrieval method before plotting a continuous trend. Retain the collection date separately from the export date. Mark a changed panel or model route; a freshly exported PDF does not mean new answers were collected.

A first report has no prior comparison. Missing company selections, unavailable storage, failed retrieval and unavailable trend data need their own availability reasons. Forecasts require enough comparable dated history and validation against held-out observations. Until then, show history required and the observed data, without inventing a future line.

The API route is a sample surface. It does not reproduce every consumer-app search experience, personalization or market. Small samples have limited precision; a change after a website edit does not establish that the edit caused the change.

Inspect the worked example and correction record

Prompeteer prepared this explanation from its report contracts and evidence rules. Content edited 10 September 2026. This is a product methodology explanation, not independent certification.

The on-page replay includes synthetic observations, derived measurements and proposed actions. Download its inputs, outputs and checks to inspect the calculation. The illustrative sample report is public. Private reports require an authorized Beacon workspace with active access and the appropriate role. Historical answers remain dated observations; corrections should identify what changed without silently rewriting them.

Common questions

Does a citation prove endorsement?

No. It establishes that an answer references a source. Read the surrounding text to determine whether it describes, criticizes, compares or recommends the business.

Can a zero mean a model request failed?

A measured zero belongs to a usable answer with no matching observation. A failed request, refusal or unavailable assessment needs a separate status and is excluded from that metric’s denominator.

Sources and review

Published by Prompeteer. See the people responsible for the product and contact the team with corrections. The linked workflow artifacts describe exactly what was checked; illustrative examples are not customer results.

Reviewed September 14, 2026

Continue your workflow

See how Beacon measures AI visibility →

Learn how Beacon separates website readiness, sampled AI answers and recommended actions. Private reports and workspace features require an authorized Beacon workspace with active access and the appropriate role.

Turn a Beacon finding into a checkable action →

Use Beacon findings to decide what to change on your site. Define the evidence, owner, action and verification before claiming improvement.

Decide what a good answer must pass →

Evaluate factual accuracy, task completion and format separately. Build representative cases, failure examples and human review criteria for AI workflows.

Put the workflow to work.

Inspect the worked exampleExplore team workflows →


Source: https://prompeteer.ai/guides/ai-visibility-measurement