1. Executive Summary

The Bottom Line

On May 19, 2026, Google announced a redesigned Search experience and made Gemini 3.5 Flash the default model inside AI Mode. It did not announce AI Mode as the global default, replace ordinary Search with an entity-only ranking system, or establish a universal business impact. Normal SEO foundations remain relevant.

SOURCE
bind every external claim to an accessible primary or reviewed authority
SCOPE
retain cohort, query class, locale, device, account state, and definition
DATE
record collection period, model, rollout state, and expiry
PROOF
keep answer, referral, lead, and revenue evidence separate

Key Findings

  1. ClickRadius joins measurement and governed delivery. The product measures five weighted AI Readiness evidence families, records separate monthly AI Presence observations from ChatGPT, Gemini, Perplexity, Claude, and Grok, prepares eligible corrections, and identifies guided off-site authority work. A supported effect is applied only with current authorization for the exact revision and a capable connection. Copilot is not currently supported.
  2. On-site and off-site evidence serve different jobs. A technically sound, useful website establishes first-party facts; accurate profiles, independent coverage, reviews, and citations can corroborate them. ClickRadius is designed to evaluate both evidence surfaces without assigning an unsupported universal split between them.
  3. Opportunity is measured, not presumed. Freeze a prompt panel and record engine, model, locale, date, mention, link, and recommendation states. A miss in that sample is not proof of universal invisibility or future market share.
  4. The platform is technically deep. ClickRadius combines website analysis, configured strategy selection, citation observation, entity workflows, and governed deployment paths. Customer-outcome training is not currently active. Patent Pending.

2. Google's Algorithm Revolution — What Changed

2.1 Google I/O 2026 (May 19, 2026)

Google VP of Search Elizabeth Reid called this "the biggest upgrade to our Search box in over 25 years." CEO Sundar Pichai separately described it as "our biggest upgrade to Search ever." Those descriptions establish Google's characterization, not a global rollout or outcome statistic.

AI Mode Model and Interface Changes

Google described a redesigned intelligent Search box, a flow from an AI Overview into AI Mode, and Gemini 3.5 Flash as the default model inside AI Mode. The announcement did not establish AI Mode as the global default or formally demote ordinary results, and Google said users continue to receive a range of results.

Search Box Redesigned

Google announced a redesigned Search box with AI-assisted interactions. That does not establish that autocomplete was universally replaced, that ranking became entity-only, or that a particular business will be cited.

AI Overview Coverage

Coverage varies by measurement source, query cohort and query class, collection date, locale, device, account state, and feature definition. This report does not assert a universal current percentage.

Information Agents

Google previewed Information Agents for AI Pro and Ultra subscribers with an announced summer rollout and described synthesized updates with web links. Announcement, eligibility, observed availability, task completion, displayed links, and referral behavior are separate evidence states.

Evidence boundary: The announcement establishes product direction, not a universal referral rate or a private citation-selection mechanism. Measure the interface and outcome relevant to each declared cohort.

2.2 The Impact — Required Measurement Record

Evidence class Required scope Observed artifact What it does not prove
Interface rollout Account, locale, device, query, date Rendered capture Global availability
AI Overview coverage Source, cohort, query class, date, locale, device, definition Numerator and denominator Universal current rate
Click behavior Feature exposure, position, impressions, period, device Clicks joined to impressions Why behavior changed
Answer observation Prompt, engine, model, locale, time Raw response and displayed links Full candidate set
Referral Referrer rule, bot filtering, landing URL, time Analytics event Citation causation
Lead or conversion Event definition, join key, attribution window Customer outcome event Incremental lift
Public-record evidence Publisher, entity, fact scope, date, conflicts Captured public record Engine recognition
Readiness Submitted URL, evidence family, scanner version, time Observed setup evidence Citation probability
Trend Same protocol and comparable collection windows Versioned observations Causal attribution

2.3 What Can Be Inspected

Normal SEO foundations remain relevant. A broader audit can also inspect structured identity and independent public records without claiming that Google uses one universal weighting:

On-page evidence

  • Crawl and index controls
  • Page purpose and supported claims
  • Valid structured data matching visible content
  • Performance and mobile usability
  • Attributable first-party evidence

Separate evidence and outcomes

  • Entity records — identity, publisher, fact scope, date, and conflicts
  • Independent evidence — source-controlled coverage, listings, and reviews
  • Answer observations — prompt, engine, model, response, links, and mentions
  • Referral evidence — classified sessions with bot filtering
  • Customer outcomes — defined events and attribution windows
  • Uncertainty — missing evidence remains unknown rather than zero

2.4 Citation, Referral, and Outcome Separation

A displayed citation is not itself a referral or a business outcome. A defensible record keeps these stages distinct:

The opportunity test: run a declared prompt panel for the business and comparison set. The result describes that sample only; it does not establish universal invisibility, future market share, or a compounding provider trust mechanism.

2.5 Evidence Principles

Expert commentary can motivate a hypothesis, but a person's reputation is not evidence for a private engine mechanism or a customer outcome. A reusable claim needs its exact source, wording, scope, date, and applicability; an attributed paraphrase without that record is not used here as authority.

Useful first-party content, accurate structured data, and independently controlled public records are distinct review surfaces. Whether a named engine crawls, retrieves, selects, cites, or refers traffic from any one of them requires direct observation under a declared protocol.

3. On-Site Foundations and Off-Site Corroboration

AI and search systems can encounter both first-party website content and independent information elsewhere on the web. The relative influence varies by engine, query, market, and time, so this report treats the two surfaces as complementary evidence rather than assigning them a universal percentage.

3.1 On-Site Foundation Scope

A conventional technical and content audit can inspect a publisher-controlled website: crawl directives, metadata, structured data, performance, accessibility, content, and internal links.

Those observations establish the measured site state only. They do not establish how an external engine will rank, retrieve, cite, recommend, or send traffic to the page.

3.2 The Expanded Model: Independent Sources Can Corroborate First-Party Facts

Independent public records create additional observable evidence surfaces beyond the publisher's own website. Whether a particular engine crawls, retrieves, uses, or cites any record remains provider-controlled and must be measured rather than inferred:

FOUNDATION

On-Site Optimization

Schema markup, meta tags, content quality, technical SEO, page speed, mobile optimization

Publisher-controlled inputs; external effect unproven

CORROBORATION

Off-Site Entity Authority

  • Directory presence — Data Axle, Foursquare, Bing Places, and industry-specific directories
  • Knowledge Graph entity — Google KG, Wikidata, and linked data sources
  • Social authority signals — LinkedIn, Reddit, and platform-specific content
  • Attribute consistency — NAP (Name, Address, Phone) comparison across the declared, successfully observed platforms
  • Third-party mentions — press, reviews, industry publications, forums
  • Cross-platform attribute comparison — agreements, conflicts, unavailable lookups, and source ownership kept distinct from identity or engine recognition

Independent profiles, editorial references, reviews, and directory records can corroborate identity and business facts found on a company's own site. Their effect cannot be reduced to one universal citation-rate comparison: ClickRadius must observe each engine, query set, and time window before reporting an outcome.

The Product Opportunity: Site diagnostics alone do not describe a business's complete evidence footprint. ClickRadius is designed to connect website observations with governed off-site identity and citation observations while keeping measured, unavailable, and externally controlled outcomes distinct.

3.3 Why This Can Matter for a Local Business

Consider a local PI attorney. A bounded on-site audit might:

But none of that tells the attorney whether:

ClickRadius connects these evidence surfaces. It records the website, supported outside-source observations, and monthly engine-answer samples without treating one surface as proof that another provider used it or that an external result will follow.

4. ClickRadius — Platform Overview

ClickRadius is more than a reporting interface — it is an AI-citation intelligence platform. It connects scoring, monitoring, entity evidence, fix generation, and governed deployment paths. Patent Pending describes status only; it is not presented here as proof of product capability or legal scope.
VERSIONED
API Route Inventory
CONTRACTED
Database Schema
57
Cataloged Scheduler Jobs in Current Source
5
Supported Engines Sampled Monthly
GATED
Entity and Distribution Effects
Pending
Patent Status

4.1 On-Site Capabilities

Capability What It Does
Scoring Engine (five weighted evidence families) Off-site Authority (28%), On-page GEO (26%), Schema Engine (18%), Entity / E-E-A-T (18%), Crawler Access (10%). These are starting estimates, not weights fitted to customer outcomes. SEO Health and measured AI Presence are reported as separate scores
Schema Analyzer Parses JSON-LD, Microdata, and RDFa; recognizes 30+ Schema.org types; scores depth, industry relevance, and completeness
Auto-Fix Engine Prepares eligible corrections; a supported effect is applied only with current authorization for the exact revision and a connection proven capable of that effect. Delivery evidence varies by connection
Content Engine Article-draft generation with internal content checks, evidence and review states, media preparation, and governed publication paths. An internal threshold does not prove factual accuracy, citation utility, image rights, or public delivery
Search Agent Scanner Six-dimension internal checklist covering semantic HTML, accessibility, structured data, crawlability, action markup, and content signals. Its score is not proof of Information Agent eligibility, discovery, or task completion

4.2 Off-Site Capabilities (the differentiator)

Capability What It Does
Entity Orchestrator Build-plan and adapter paths exist for selected directories, professional/social surfaces, and community work. Each external effect requires platform authorization, exact-item approval where applicable, capability, receipt, and readback; adapter presence is not proof that an entity was built
Citation Monitor Monthly answer and citation observations from Claude, ChatGPT, Gemini, Perplexity and Grok, with change and cross-engine summaries limited to the sampled responses; Microsoft Copilot is not currently supported
AI Overview Tracker The scheduled collector is held pending validation. Simulation records must remain labeled as simulations and cannot be presented as observed Google AI Overview presence, citations, or competitor share
Knowledge Panel Monitor Attempts Google Knowledge Graph and Wikidata lookups and records a configured completeness heuristic. Provider failure, no match, attribute agreement, identity, and engine recognition remain separate states
Entity Verification Compares available entity attributes across configured public sources and records discrepancies and coverage. Agreement is not credential verification, provider recognition, or proof of current authority
Outside Signals Scanner Configured checks reference 20+ publishers and platforms across five legacy categories. A result is usable only with its actual source coverage and collector state; the composite is an internal rubric, not measured citation influence
Conversational Query Analyzer Produces conversational variants and records a model-simulation comparison with an explicit measurement kind. A simulation is not an observed citation likelihood or Google AI Mode result

4.3 Intelligence & Automation

Capability What It Does
Configured Strategy Selection Samples stored Beta-Binomial arm state to order eligible auto-fix strategies and records the decision evidence; customer-outcome training is not currently active
Outcome Measurement Records grounded pre/post citation observations and quarantines missing, sparse, or confounded evidence; it does not establish causation
Content Distribution Processes eligible, specifically approved revisions for authorized and capable destinations. Provider acceptance, public availability, engagement, entity authority, and citation outcomes remain separate evidence states
White-Label System Reseller-scoped branding exists in supported portal and report paths. Complete surface coverage, tenant isolation, and production behavior require separate release evidence
Reseller Infrastructure Source includes reseller registration, client/site management, onboarding, report, and billing workflows. Their implementation states and production proofs are tracked separately; this list is not an end-to-end completion claim

4.4 Competitive Advantages

  1. Entity work is represented separately from scoring. ClickRadius can prepare governed entity and distribution work without treating a higher score, an adapter call, or a provider receipt as proof of an external result.
  2. Monthly citation observations across five supported AI engines. ClickRadius samples ChatGPT, Gemini, Perplexity, Claude and Grok because each integration can return different answers and citation evidence. Microsoft Copilot is not currently supported, and one monthly sample is not a continuous view of every user response.
  3. Governed correction delivery. ClickRadius prepares eligible schema, metadata and technical corrections. A supported effect is applied only with current authorization for the exact revision and a capable connection; otherwise it remains for review or customer implementation. Provider acceptance and independent public-origin observation remain separate evidence.
  4. Traceable strategy selection. The current engine records selected and competing samples plus an available decision-time baseline. That is decision evidence, not a claim that the platform has learned what works.
  5. Explicit outcome boundaries. Missing baselines, sparse coverage, and co-occurring same-site decisions are quarantined rather than presented as causal proof.
  6. Patent status. ClickRadius identifies its technology as Patent Pending. Filing status, implementation state, and legal scope remain separate evidence questions.

5. ClickRadius — Technical Architecture & Specifications

This section inventories current ClickRadius source paths, configured mechanisms, and explicit hold boundaries for technical evaluation. Each item states whether it is an active measurement, a model simulation, a governed candidate path, or a held capability. Implementation, integration, production release, provider execution, and public effect are separate evidence stages; source presence is not itself a statement that a capability is live for every customer.

5.1 Scoring Engine — Five-Family Weighted Analysis

The scoring engine crawls a submitted site and runs analyzers across five weighted evidence families. The composite score (0-100) is an estimated setup-evidence index; it is not a probability that an engine will cite the site, a forecast, or a customer outcome:

Off-site Authority
28%
On-page GEO
26%
Schema Engine
18%
Entity / E-E-A-T
18%
Crawler Access
10%

Schema Engine analyzer (18% weight)

Parses three structured data formats: JSON-LD (application/ld+json script blocks), Microdata (itemscope/itemprop attributes), and RDFa (typeof/property attributes). Recognizes 30+ Schema.org types across three tiers:
  • Business types: LocalBusiness and 20+ subtypes — LegalService, Dentist, Restaurant, Store, MedicalBusiness, AutoDealer, RoofingContractor, Electrician, LocksmithService, MovingCompany, Hotel, ChildCare, Florist, TaxiService, and more
  • Organization types: Organization, Corporation, GovernmentOrganization, NGO, EducationalOrganization, MedicalOrganization, SportsOrganization, Airline
  • Content types: Article, BlogPosting, WebPage, WebSite, CreativeWork, Event, VideoObject, FAQPage, HowTo, BreadcrumbList, Product, Service
Scoring factors: LocalBusiness presence (+25), industry-specific subtype (+15), Organization fallback (+10), FAQPage (+8), HowTo (+5), BreadcrumbList (+5), schema depth (nested properties), content schema bonus (+5).

Legacy GEO Content Heuristic

This internal metric counts three page-content patterns. It does not evaluate the likelihood that an AI engine will retrieve, select, cite, or recommend the page. Its configured dimensions are:

GEO Score = min(100,
  (min(quotations, 5) / 5) × 33 +
  (min(statistics, 8) / 8) × 33 +
  (min(citations, 5) / 5) × 34
)
Quotation detection: Counts <blockquote>, <cite>, and <q> HTML elements.

Statistics detection (7 regex patterns): Percentages with context (e.g., "X% of/increase/decrease"), dollar amounts with context ("$X per/in/worth"), numeric counts ("X years/clients/cases"), approximations ("over/more than X"), compound counts ("X+ years"), ratings ("rated/score X.X"), and multipliers ("Xx faster/more").

Citation detection: Text patterns ("according to", "source:", "research shows", "study found") plus external links with citation-related anchor text.

Structural bonus signals: Table presence, heading depth (H2+H3 hierarchy), FAQ section detection, ordered lists, definition lists — tracked but not scored, used for content improvement recommendations.

5.2 Citation Monitor — Five Supported Engines; Monthly Sampling

Each month, the Citation Monitor sends queries through the five supported engine integrations, analyzes the returned responses for brand mentions and URL citations, computes sentiment, and records change over time. Microsoft Copilot is not currently supported and returns an explicit not-supported response.

Supported AI Engines

Engine Integration Method Current source default (environment-overridable)
Claude (Anthropic) Official SDK (@anthropic-ai/sdk) ANTHROPIC_SONNET (default claude-sonnet-5)
ChatGPT (OpenAI) Official SDK (openai) OPENAI_SMART for grounded checks or OPENAI_FAST without hosted search (defaults gpt-5.1 / gpt-5-mini)
Gemini (Google) Official SDK (@google/generative-ai) GEMINI_PRO (default gemini-2.5-pro)
Perplexity REST API (api.perplexity.ai) Sonar
Grok (xAI) OpenAI-compatible SDK (api.x.ai) XAI_GROK (default grok-3-mini)
Copilot (Microsoft) Not supported; no ClickRadius observation data is collected. No future integration or timing is promised.

Citation Score Formula

baseScore = min(100,
  mentionRate × 0.5 +
  (cited > 0 ? 30 : 0) +
  avg_confidence × 20
)

citationScore = min(100, baseScore × 0.7 + weightedMentionRate × 0.3)
Configured response-classification values: Brand mentioned + URL cited = 0.95 | Brand mentioned, no URL = 0.70 | Not mentioned = 0.30. These constants are labels used by the current implementation, not calibrated probabilities.

Configured engine coefficients: Each AI engine can have a scoring profile stored in engine_profiles. For example, current configuration can assign different coefficients to Perplexity citation observations and Gemini schema-related inputs. These are product settings, not weights fitted to customer outcomes and not proof of how an engine actually selects sources.

Sentiment analysis: 11 positive keywords (recommend, excellent, top, best, leading, trusted, great, quality, reliable, expert, outstanding) and 9 negative keywords (avoid, poor, bad, worst, scam, unreliable, expensive, overpriced, disappointing) with weighted comparison.

Citation Velocity Tracking

A weekly aggregation job can store snapshots in the citation_velocity table from retained citation checks; it does not collect a new engine response. Scheduled customer answer and citation collection remains monthly on the first day at 4:00 AM ET. Rate-of-change is computed via SQL LAG() across stored periods. The labels below are configured change bands, not business outcomes:
Accelerating (>5% increase) | Growing (>0%) | Stable (>-5%) | Declining (>-15%) | Dropping (<-15%)

Cross-Engine Consensus Scoring

consistencyScore = max(0, 100 - mentionStdDev × 2)

consensusScore = consistencyScore × 0.3 + crossPlatformRate × 0.4 + sentimentAgreement × 0.3
Summarizes how consistently the brand appeared across the five monthly sampled engine responses. High or low consensus describes dispersion in that measured set; it does not prove entity authority or identify why an engine produced a different answer. Microsoft Copilot is excluded because it is not currently supported.

Configured Per-Engine Guidance

An engine profile can include static review tips. They are configured guidance, not customer-outcome measurements or validated causal models of an engine. Evidence-safe examples are:
  • ChatGPT: inspect separately verified public identity records, including Wikipedia or Wikidata where they exist; absence remains unknown and does not predict citation
  • Gemini: validate structured data against the visible page; publisher markup does not establish retrieval, selection, or citation
  • Perplexity: "Use FAQ schema only when visible question-and-answer content supports it; markup does not guarantee citation"
Engine-labeled suggestions can focus a review, but a difference between sampled responses does not identify why an engine answered differently. Any causal or quantitative recommendation requires its own versioned evidence packet, applicability limits, and expiry.

5.3 Entity Intelligence Layer

Entity Orchestrator — Governed Build Plans and Adapters

The Entity Orchestrator can inspect site data, prepare a build plan, and route eligible work toward a modular adapter architecture. An adapter module establishes a code path only. It does not establish current platform permission, usable credentials, provider acceptance, a public record, or an AI-engine outcome.

Platform adapters:
PlatformFunctionIntegration
Data AxleDirectory candidate workAdapter path; authorization, terms, receipt, and readback required
FoursquareLocation-data candidate workAdapter path; authorization, terms, receipt, and readback required
Bing PlacesBusiness-profile candidate workAdapter path; authorization, terms, receipt, and readback required
LinkedInProfessional-profile or content candidate workExact account and item authority required; publisher receipt is not public proof
RedditCommunity-response candidate workPer-item approval and truthful affiliation disclosure required; no source-rank claim

Build Plan Intelligence

The orchestrator doesn't blindly submit to every platform. It generates a prioritized build plan based on the Outside Signals score gaps:
Signal CategoryConfigured TriggerCandidate Work
Brand VisibilityScore < 40Prepare supported directory work for review and authorization
Social AuthorityScore < 50Prepare an exact account-bound LinkedIn item
Earned MediaScore < 30Prepare an evidence-bound distribution item
Community PresenceScore < 40Prepare a disclosed response for per-item human approval
These thresholds are configured prioritization rules, not outcome-fitted cutoffs. An associated correlation_event can preserve descriptive linkage for later analysis; it does not establish that an action happened publicly or caused a citation change.

Knowledge-Record Cross-Reference

Current source contains two provider lookup paths. A successful response can support a dated record comparison; transport failure, no match, and verified absence must remain distinct, and a first search result is not identity proof:
  • Google Knowledge Graph Search API — entity lookup returning name, types, description, KG ID, relevance score, image, and URL
  • Wikidata API — entity search + claims extraction (P856: official website, P571: inception date, P17: country)
Configured legacy completeness heuristic (0-100): these are internal rubric points, not calibrated engine-recognition or authority probabilities.
SignalPoints
Google Knowledge Graph presence+25
KG has description+10
KG has detailed description+10
KG has image+5
KG has URL+5
Wikidata entity exists+15
On-site Organization/LocalBusiness schema+15
sameAs links in schema+10
Logo in schema+5
Founder/foundingDate in schema+5
Trend analysis can compare stored heuristic scores across periods and assign configured change labels. A score delta does not establish that an external knowledge record changed, that an engine recognized the entity, or that a business outcome improved.

Entity Attribute Comparison

Compares available attributes from on-site schema, Google Knowledge Graph, Wikidata, and declared social profiles. These are not four independent authorities: several records can be publisher-controlled, incomplete, stale, or about a different entity.

Social platform inventory: Facebook, LinkedIn, Twitter/X, Instagram, YouTube, Yelp, BBB, Wikipedia, Wikidata, and Crunchbase URLs can be extracted from sameAs. An HTTP reachability response does not verify ownership, identity, credentials, or current authority.

Comparison checks: Name and URL equality and profile reachability can identify records for review. A match is consistency evidence only; a mismatch requires identity-aware adjudication rather than automatic replacement.

Configured review priorities: labels such as Critical, High, and Medium are workflow configuration, not measured citation harm. Missing or unavailable records remain explicitly unknown where absence has not been proven.

5.4 Outside Signals Collector — Coverage-Bound Public Evidence

The current collector attempts configured public-source checks across five legacy categories. Every result needs source, observation time, collector state, and coverage. This route's internal composite is separate from the five-family AI Readiness score and is not an outcome-fitted model of citation influence.
Signal Category Weight Platforms Scanned Key Metrics
Brand Visibility 25% Google Search (via DataForSEO API), Bing Search, Google News, Google Knowledge Graph Mention count, KG panel presence, search result diversity
Reviews 20% Google Business Profile, G2, Capterra, Trustpilot, Yelp, BBB Platform count, average rating, total review volume
Community 20% Reddit (OAuth API), Quora, Stack Overflow, Hacker News (Algolia API) Mention count, subreddit diversity, upvote engagement, HN story count
Media 20% News publications, guest posts, podcasts, syndication networks, Wikipedia News mentions, external link count, Wikipedia presence
Social 15% LinkedIn, Twitter/X, YouTube, Facebook Platform presence, sameAs schema count, channel discovery, diversity score
Scoring examples:
Brand Visibility: 15+ mentions = 40pts, KG panel = +30pts, any signals = +10pts, 2+ signal types = +20pts (max 100)
Reviews: Each platform = 10pts (max 40), avg rating ≥4.5 = 30pts, 100+ reviews = 30pts (max 100)
Community: Reddit observations, subreddit diversity, and upvote engagement can be recorded by the configured rubric. Any source-specific score weight is product configuration, not a cross-engine source ranking.
Media: Wikipedia presence can receive configured rubric points when it is actually observed. That score is a publisher-record signal, not evidence that every engine checked the record or that it caused a citation.

Composite: overall = brand × 0.25 + review × 0.20 + community × 0.20 + media × 0.20 + social × 0.15

5.5 AI Overview Observation Hold & Search Agent Optimization

AI Overview Observation — Collector Held

The current route can retain a keyword list, but observed Google AI Overview collection is held pending live-collector validation. Manual and batch observation endpoints return collector_unavailable before provider spend, result persistence, or timestamp mutation. Historical rows created by the retired language-model simulation path remain available only as provenance: their outcome fields and predicted-source lists are suppressed from customer observation metrics.

Scheduled state: The Wednesday 3:00 AM ET slot emits a durable held receipt with zero attempted keywords, simulations, and Google observations. It does not perform a weekly check.

Contained legacy formula: The formula below documents the earlier implementation only. While the collector is held, the current summary returns aio_visibility_score: null and no observed rate:
aioVisibilityScore = min(100,
  citationRate × 0.6 +
  (aioTriggerRate > 50 ? 20 : aioTriggerRate × 0.4) +
  (clientCited > 0 ? 20 : 0)
)
Citation-share boundary: The current endpoint returns a null client share, zero eligible citations, and an empty domain list. It reports counts of stored historical simulations separately; model-predicted sources are not competitor observations and are not eligible for an AIO share claim.

Search Agent Publisher-Surface Scanner

Inventories 6 publisher-controlled dimensions with 70+ individual checks. This internal checklist does not measure Information Agent availability, eligibility, retrieval, selection, citation, task completion, or referral:

Dimension Weight What It Checks
Semantic HTML 15% 14 semantic tags (nav, main, article, section, aside, header, footer, figure, figcaption, details, summary, time, mark, address), semantic-to-generic ratio, non-semantic interactive element penalties
Accessibility 15% Alt text coverage, ARIA landmarks, lang attribute, heading order validation (H1 first, no skipped levels), skip links, tabindex audit, ARIA-labeled elements
Structured Data 20% Observed structured-data type strings, including Organization, LocalBusiness, WebSite, Article, FAQPage, HowTo, Product, Service, BreadcrumbList, SiteNavigationElement, and SearchAction. Presence is not vocabulary, factual-parity, page-scope, eligibility, or agent-use proof
Crawlability 15% robots.txt policy checks for named documented agents, sitemap presence, RSS feed detection, IndexNow header, canonical tags, and noindex checks. Policy is not proof of a successful fetch, indexing, or agent use
Action Readiness 15% CTA elements, forms with proper labels, tel: links, mailto: links, and contact-page links. These are observable visitor affordances, not proof that an external agent recognized or used an action
Content Originality 20% Author attribution, publication dates, about section, testimonials, statistics density, source citations, experience markers ("in our experience", "we've found"), minimum word count

5.6 Conversational Query Analyzer

This analyzer generates natural-language variants and asks a configured language model to simulate the original keyword and up to four variants. Every returned figure is labeled model_simulation; no search engine is contacted and observed_comparisons_completed remains zero.

Process:
  1. Takes a traditional keyword (e.g., "personal injury lawyer chicago")
  2. AI generates 4 conversational reformulations with intent classification (informational, transactional, comparative, local) and complexity rating (simple, moderate, complex)
  3. The model returns a simulated citation estimate, position, source candidates, and suggested content signals for the original and each usable variant
  4. The system computes a comparison only when both the original and at least one variant produced usable model output; unavailable output remains unavailable rather than zero
Hypothesis generation:
  • Opportunity hypothesis: The model estimated more simulated coverage for conversational variants
  • Risk hypothesis: The model estimated the original but not the usable conversational variants; this is not a traffic forecast
  • Modeled difference: A percentage-point difference between model-returned estimates, not measured citation lift
  • Review input: Model-suggested content signals, frequency-ranked across usable variants
Stored state: Comparable simulations can be classified as conversational_better, keyword_better, or similar. Those labels describe model output only; they are neither Google AI Mode observations nor forecasts of citations or traffic.

5.7 Content Engine — Governed Article-Draft Pipeline

The current source creates exact-revision article candidates through bounded generation, safety, authority, research, source-evidence, corpus, and publication gates. A generated draft is not automatically a factual, citable, published, or externally observed article.

1
Fact Authority — Binds authorized client facts and an accountable author snapshot to the request. Crawled or extracted text is not promoted to a verified fact merely because it appeared on a page
2
Research Receipt — Retrieves source candidates and requires a versioned receipt with at least three independently owned domains before governed admission; retrieval is evidence about the fetched source, not proof that every source claim is true
3
Bounded Draft Generation — The configured Anthropic Sonnet model receives the reader task, authorized facts, reviewer context, and retrieval-observed sources. The prompt explicitly rejects invented identities, business facts, quotations, numbers, credentials, and URLs
4
Safety and Retry Loop — Up to five bounded generation attempts can repair response shape, HTML-policy, depth, and exact source-evidence failures. If admission still fails, the route returns a typed refusal with diagnostics rather than persisting a successful candidate
5
Exact Source-Evidence Admission — Inventories visible factual units, fetches declared sources through the governed verifier, checks source revisions and support, and requires zero uncovered or unsupported factual units. This verifies the recorded source-support contract, not universal truth or future citation
6
Business-Fact Consistency Screen — A narrow model check compares business- and author-specific statements with the bound facts. Missing, failed, or unparseable checks remain unavailable and fail admission; a passing screen is not a blanket factuality certification
7
Diagnostic Screens — The legacy GEO heuristic records quotation, statistic, and citation-pattern counts, with 75 displayed as a screening target. The separate E-E-A-T model estimate can score four dimensions. Neither score is an admission proof or an external-engine prediction, and unavailable E-E-A-T remains null
8
Corpus Admission — Binds the exact corpus snapshot and rejects duplicate or near-duplicate intent/content under versioned thresholds. "Not duplicative" does not itself establish positive information gain, usefulness, rankings, traffic, leads, or revenue
9
Schema and Claim Verification — Derives BlogPosting markup from the visible draft, bound author, publisher context, keyword, description, and admitted source URLs. A background terminal check recomputes the exact source-evidence receipt before the publication gate
10
Media Hold and Governed Publication — Current generation records media as held_contract_required; it does not generate or deliver a featured image until applicability, rights, accessibility, connector, approval, receipt, readback, and withdrawal contracts clear. An approved exact article revision can publish only with current authority and a capable connection; provider receipt and public-origin observation remain separate
Editorial workflow: Generated candidates begin in review. Approval, scheduling, exact-revision authority, fact/source clearance, connector capability, provider receipt, and independent origin readback are distinct states; an hourly worker processes eligible scheduled obligations but its invocation is not publication proof.

Prompt and HTML defenses: Selected prompt-bound strings are normalized and length-limited, untrusted-data boundaries are repeated in system/user prompts, and generated HTML passes an allowlist inspection. These are defense-in-depth controls, not a guarantee that all possible input or model output is safe.

RSS feed: Public endpoint per site serving RSS 2.0 with Atom self-link, up to 50 published articles.

5.8 Strategy Engine — Bayesian Selection, Outcome Training Inactive

The Strategy Engine currently uses Thompson Sampling over stored Beta-Binomial arm state to order eligible auto-fix strategies. It records each selection and an available grounded citation baseline. The implemented selection mechanism must not be confused with a causally validated customer-learning loop; customer-outcome training is not currently active.

How it works:
  1. Active optimization strategies are stored by category and represented as arms with configured Beta(α, β) state
  2. When deciding what to optimize, the engine samples from each arm's Beta distribution using Marsaglia-Tsang Gamma sampling + Box-Muller normal distribution
  3. The batch auto-fix path selects the top three samples, writes strategy-decision records, and captures one grounded baseline when available
  4. Those simultaneous same-site decisions are treated as confounded by the scheduled outcome observer and are recorded without training
  5. A shared causal-eligibility and exactly-once update contract, including the admin path, must close before customer-outcome training can be activated

Informative Priors by Category

Different strategy categories start with configured values. Their implied means are initialization choices, not measured customer success rates:
CategoryPrior (α, β)Implied Success RateRationale
Schema(3, 2)60%Configured initialization; not an observed outcome rate
Meta(3, 2)60%Configured initialization; not an observed outcome rate
GEO(2, 2)50%Configured initialization; not an observed outcome rate
Content(2, 2)50%Configured initialization; not an observed outcome rate
Technical(2.5, 1.5)63%Configured initialization; not an observed outcome rate
Entity(1, 1)50%Configured initialization; not an observed outcome rate

Stored-State Blending

blendFactor = min(1.0, total_pulls / observations_needed)

effectiveAlpha = blendFactor × alpha + (1 - blendFactor) × parent_alpha
effectiveBeta = blendFactor × beta + (1 - blendFactor) × parent_beta
The selection code blends arm and parent state according to total_pulls / observations_needed. Reward propagation exists in source, but the normal batch path is currently quarantined as confounded and the admin path does not yet share the full causal and exactly-once gate. This is why the blending mechanism is documented without claiming active customer learning.

5.9 Outcome Observation — Current Evidence Boundary

The current scheduler compares grounded citation observations before and after a strategy decision. That comparison can support a descriptive history record; by itself it cannot establish that an optimization caused the change.

Methodology:
  1. Capture a trailing grounded citation rate, or null when no grounded baseline exists, at decision time
  2. Re-observe eligible pending decisions after 14 days and before they age beyond the current 28-day query window
  3. Leave decisions pending when grounded after-data is unavailable
  4. Record but withhold from training when the baseline is missing, fewer than 10 grounded samples exist, or multiple same-site decisions are confounded
  5. Keep customer-outcome training inactive until treatment identity, measurement windows, causal eligibility, and exactly-once updates are enforced across every writer
Current eligibility checks:
grounded baseline present ∧ grounded after-data present ∧ samples ≥ 10 ∧ no co-occurring same-site decision

Launch boundary:
descriptive observation ≠ causal attribution ≠ training eligibility
Why this matters: A score change and a citation change can move together for many reasons. ClickRadius preserves the observations and known confounding state without labeling that correlation as causal proof.

5.10 Auto-Fix Engine — Prepare, Govern, Observe, Revert Where Supported

The Auto-Fix Engine prepares eligible corrections. A supported site effect is applied only with current authorization for the exact revision and a capable connection. The three connection methods below have different proof boundaries; manual customer implementation is a fourth delivery path, not an automated connector:
Fix Type What It Generates
Meta DescriptionAI-generated candidate description bound to the governed metadata-effect workflow; it does not predict citation
Title TagAI-generated title tags with entity and topic focus
Alt TextAI-generated image alt text for accessibility and structured signals
Internal LinksAI-suggested contextual internal links for topic clustering
Deployment Method How It Works
WordPress REST APICapability is probed per connection and effect. An approved exact revision can be delivered through supported fields; a provider receipt is not, by itself, proof of the public response.
JavaScript SnippetClient-side delivery only. It does not change raw origin HTML, so public-origin verification remains unknown rather than being promoted to verified.
Cloudflare WorkerEdge response rewriting where the connection demonstrates the required capability; provider acceptance and an independent public-origin reread are recorded separately.
ManualThe proposed correction remains for customer implementation when governed connector delivery is not authorized or capable.
Strategy-guided batch fixes: The batch endpoint samples configured arm state, selects up to three mapped fix types, and writes strategy-decision records. Because those decisions occur together on one site, the current scheduled observer treats them as confounded and withholds them from training.

Post-delivery observation: After a connector attempt, the system waits and independently re-reads the public page. A provider receipt and the returned public-origin HTML are separate evidence. Server-rendered changes may be observed in that response; client-side JavaScript-snippet effects are absent from raw origin HTML and remain not verifiable at the public origin rather than being promoted to verified. Records verified_at and verified_ok; neither field proves that an external engine observed the effect.

Revert capability: Effects deployed through a connector that declares a bound revert path store their prior value, and revert restores the original through the same deployment channel. Where a connector declares no reversal capability for an effect, the change is recorded and recovery is operator-owned rather than one-click.

Schema lock protection: Sites with schema_locked = true are blocked from all auto-fix operations. This prevents automated changes from overwriting carefully crafted custom schema.

5.11 Automation — Selected Cataloged Operations

The current source catalog defines 57 named scheduler jobs on cadences from minutes to monthly; the table below is a selected functional view, not the complete denominator. Registration, a lock attempt, and source implementation do not prove a job ran successfully in production. Completion requires a run-specific heartbeat or outcome receipt.

Job Schedule Function
Citation Monitor1st of month, 4:00 AM ETAttempt due query-engine observations across five supported integrations and record measured, partial, unavailable, or failed population state
Citation FeedbackDaily 5:30 AMAnalyze retained response associations without promoting them to causal effects
Conversation FinderDaily 9:00 AMCollect eligible conversation candidates; discovery is not approval or publication
Content DistributionDaily 10:00 AMProcess due, authorized, exact-revision distribution work; provider and public effects remain separate
Publish ScheduledHourlyProcess due approved publication obligations with retry and provenance states
Health CheckEvery 6 hoursRecord the configured reachability checks and their failures
Engagement TrackingEvery 12 hoursAttempt configured engagement observations with explicit coverage and provider state
Freshness ScanEvery 3 daysRefresh due freshness evidence without treating a failed fetch as a measured zero
Citation VelocityMonday 5:00 AMAggregate retained citation checks; this job does not collect a new engine response
Entity BuildTuesday 7:00 AMProcess eligible authorized build-plan work; an adapter call is not public-effect proof
Weekly Auto-FixWednesday 5:00 AMProcess eligible exact-revision corrections through the applicable authority and capability gates
AI Overview ObservationHeldCollector validation pending; no simulated result is represented as observed tracking
Outcome MeasurementFriday 8:00 AMRecord eligible descriptive before/after observations and quarantine missing, sparse, or confounded evidence
Outside SignalsSunday 6:00 AMAttempt supported off-site collectors and preserve coverage and error state
Entity Check1st of monthCompare available entity attributes without converting unavailable evidence into absence
Knowledge Panel1st of monthAttempt configured Knowledge Graph and Wikidata lookups while retaining provider, match, and coverage state
Entity Verification1st of monthCompare available cross-platform attributes without treating unavailable evidence as absence
Report Delivery1st of monthProcess period-bound report generation and delivery; generated, sent, delivered, and accessible are separate states
Agent Readiness15th of monthSearch Agent optimization scan (6 dimensions)
Conversational Query1st & 15th of monthGenerate explicitly labeled model simulations; no observed engine comparison is claimed
Retry QueueEvery 2 hoursProcess eligible failures from workflows that use the scheduler retry queue; it is not a universal retry path
Advisory locks: Cataloged jobs have distinct lock identities, and registered runners attempt a PostgreSQL advisory lock before the job body. A skipped overlap and a completed run are separate states; locking reduces concurrent execution risk but is not a result receipt.

Retry and recovery: Some material workflows use durable retry or reconciliation records, including scheduler_retry_queue where that contract applies. Other jobs record failures and alerts without sharing that queue. No blanket automatic-retry claim applies to every scheduled job.

Run evidence: Activity, heartbeat, population accounting, alert, and effect receipts vary by job. Release validation compares the expected job denominator with those exact records; source registration or a dashboard timestamp alone does not prove successful work.

5.12 White-Label & Reseller Infrastructure

Current source includes reseller-scoped branding and agency workflows. The white-label system remains a partial product surface: implementation presence does not prove every portal, report, email, agreement, support path, or external artifact uses the reseller identity, and this page makes no zero-ClickRadius-branding claim.

White-Label Branding

Multi-tenant branding resolution: Site → Client → Reseller profile. The system traverses the ownership chain to find the correct branding. Customizable elements:
  • Company name — resolved by supported portal and report consumers; complete coverage requires rendered-path proof
  • Logo URL — available to supported portal/report consumers, subject to asset and accessibility validation
  • Primary color — available to supported themed components; contrast and complete-surface coverage remain separate checks
  • Portal title — configurable where the consumer implements the branding contract

Reseller Management System

Self-registration: Resellers create their own accounts with company name and credentials. Account created in pending_payment status until billing is configured.

Dashboard: 5 parallel aggregation queries showing client count, site count with health distribution (healthy/warning/critical), scan history, recent activity, and top-performing sites by score.

Onboarding source path: When an authorized reseller adds a client site, the implementation can request several downstream steps. Each step needs its own durable state and completion evidence:
  1. Triggers a full analysis scan in the background
  2. Generates content prompts for the site's business type
  3. Picks the top 3 keywords from auto-generated citation queries
  4. Generates 3 draft articles (1,200+ words, authoritative tone)
No completion time is promised by the presence of those calls; a queued request, generated draft, delivered report, and externally visible artifact are separate stages.

Branded reports: Monthly report generation and delivery paths can consume reseller branding. Exact PDF/email identity, current evidence coverage, delivery receipt, and customer access require run-specific proof.

Billing: Subscription, payment, cancellation, and access-control workflows exist in source. Live processor settlement, agreement binding, term calculation, entitlement change, and end-of-term access require reconciled commercial and production evidence.

Conclusion

Google's May 2026 announcement warrants direct observation, not panic or an invented universal funnel. Teams can preserve ordinary SEO foundations while adding explicit answer, referral, and outcome measurements.

ClickRadius was built for this moment. The platform connects configured strategy selection, citation observation, entity workflows, a multi-step content pipeline, and governed auto-fix preparation. An exact supported site effect is applied only with current authorization for the exact revision and a connection proven capable of that effect. Each capability still has to pass its own evidence and release gate; this report does not turn implementation presence into an outcome claim.

The on-site + off-site evidence model is the differentiator. Website diagnostics answer one class of questions; independent identity, presence, and citation observations answer another. ClickRadius is designed to connect those classes without representing an external platform action or an unobserved AI response as a measured outcome.

Opportunity is specific to a business, market, prompt panel, engine, model, locale, and date. Establish that baseline before making a timing or investment claim.

The Platform

Five AI engines are sampled monthly for answer and citation observations: ChatGPT, Claude, Gemini, Perplexity and Grok. Microsoft Copilot is not currently supported. Configured strategy selection, decision-evidence recording, governed deployment paths and white-label workflows retain their separate evidence and authorization boundaries. Patent Pending.

ClickRadius is an AI-search evidence and governed-delivery platform. Each capability and public claim remains at its evidenced release stage.

Primary Authorities and Claim Status

  1. Google Search I/O 2026: the first-party announcement supports the redesigned Search experience and Gemini 3.5 Flash as the default model inside AI Mode. It does not establish AI Mode as the global default.
  2. Google Search Central: current guidance says normal SEO foundations remain relevant to AI features and that no special optimization is required merely to appear in them.
  3. Third-party funnel and ROI figures: the earlier page did not retain accessible source links and independently reviewable claim packets for the displayed numbers. They have been removed rather than presented as current universal facts.
  4. Future quantitative claims: require source, cohort, method, query class, collection date, locale, device, model or feature version, applicability, uncertainty, and expiry before publication.