Agent runtime: 12 phases from reality discovery to measurement

A Phase A: Reality discovery

Before editing:

1. Crawl/index repository routes and public URLs.
2. Identify templates/page types.
3. Identify analytics/search integrations.
4. Locate robots.txt, sitemap generation, canonical logic, metadata logic and schema generation.
5. Locate source-of-truth files/databases/APIs.
6. Find existing SEO tests, redirects and deployment config.
7. Identify pages with known traffic/revenue/citation evidence when data exists.
8. Save baseline commit/hash or file snapshot.

Required baseline artifact

seo-baseline.md or equivalent must include:

• discovered site architecture,
• page inventory,
• technical stack,
• crawl scope and limitations,
• access available/missing,
• known conversions,
• high-value pages,
• known search/AI telemetry,
• evidence gaps.

B Phase B: Technical gate audit

Run G → T before content generation.

Minimum checks:

• status codes
• robots rules
• noindex/X-Robots
• canonical consistency
• redirect chains/loops
• sitemap validity
• lastmod integrity
• hreflang where relevant
• JS-rendered critical content
• crawlable internal links
• mobile parity
• HTTPS/mixed content
• CWV evidence where available
• WAF/CDN crawler blocking
• OpenAI crawler policy
• Anthropic crawler policy
• Bingbot/IndexNow
• Google generative AI inclusion setting where accessible

Do not write 50 new pages while P0 crawl/index errors are unresolved.

C Phase C: Site intent map

Build a query-intent-url map:

- intent: "..."
audience: "..."
journey_stage: "before|during|after"
current_url: "..."
page_type: "service|product|comparison|guide|tool|local|case-study|support"
business_value: 0
evidence_strength: 0
action: "keep|improve|merge|create|noindex|redirect|investigate"
reason: "..."

Rules:

• One primary canonical resource per substantially same intent.
• Different intent may justify a separate page.
• Query variation alone does not.
• Existing winner is upgraded before replacement.

D Phase D: Evidence inventory

Create evidence-ledger with:

- claim: "..."
subject: "..."
source_type: "official|first_party|independent|practitioner"
source: "..."
checked_at: "YYYY-MM-DD"
geography: "..."
version_model_sku: "..."
confidence: "verified|partial|unknown"
usable_for_publication: true

Critical claim without evidence cannot be promoted to public fact.

E Phase E: Opportunity prioritisation

Priority order:

1. P0 policy/security/truth/indexability failures.
2. P1 technical blockers.
3. Existing high-value pages with clear near-win.
4. High-buyer-intent comparison/product/service gaps.
5. Evidence-rich non-commodity assets.
6. Entity/trust consistency.
7. Distribution/linkable assets.
8. P2/P3 experiments.

F Phase F: Page optimisation

For each page:

1. Confirm intent.
2. Confirm canonical ownership of that intent.
3. Confirm facts/evidence.
4. Put main utility early.
5. Improve title/H1/intro alignment.
6. Improve content only where user blockers exist.
7. Add original evidence/utility where possible.
8. Improve internal links.
9. Improve media if useful.
10. Correct schema only from visible verified facts.
11. Test mobile/render/accessibility.
12. Record diff and rationale.

Do not rewrite simply to “sound more SEO.”

G Phase G: Commercial choice optimisation

On commercial pages:

• surface decision facts early,
• include real trade-offs,
• expose suitability and non-suitability,
• make price/terms unambiguous when publishable,
• compare against real alternatives where intent justifies it,
• add methodology to tests/comparisons,
• link to deeper evidence,
• provide a clear next action after utility is delivered.

H Phase H: Entity and trust synchronisation

Compare:

• site visible identity,
• Organization/Person schema,
• About/Contact,
• social profiles,
• business profiles,
• legal/trading name,
• pricing/availability/service facts.

Contradictions become P0/P1 depending on impact.

I Phase I: Agent-readiness

Current web agent guidance supports treating the accessibility tree, DOM and rendered screenshot as complementary interfaces.

Audit:

• use real <button> for actions,
• use real <a href> for navigation,
• labels linked to inputs,
• logical focus/keyboard flow,
• stable layouts,
• no transparent overlays hiding controls,
• visible actionable controls,
• meaningful accessible names/states,
• purchase/booking flows with explicit user confirmation where necessary.

Experimental only: WebMCP or emerging agent protocols may be evaluated as P3. Do not make them a basic SEO requirement.

J Phase J: Validation

A patch is not complete until applicable tests pass.

Technical validation

• crawl changed URLs,
• verify 2xx/3xx behaviour,
• inspect canonical,
• inspect robots/indexability,
• render page with JS,
• mobile viewport check,
• internal link check,
• structured data validation,
• accessibility spot check,
• console/network errors.

Content validation

• critical claims mapped to evidence,
• title/H1/intro same intent,
• no fabricated first-hand experience,
• no stale year/date,
• no unnecessary duplicate section,
• no keyword stuffing,
• comparison methodology parity,
• disclosure present where relevant.

Regression validation

• important existing text not accidentally removed,
• conversion element still works,
• analytics events preserved,
• page performance not materially degraded,
• no new orphan URL,
• no unexpected redirect/canonical changes.

K Phase K: Publish gate

Use section 9 hard gates and score.

Additional agent rule:

• PASS requires evidence.
• UNKNOWN cannot be silently converted to PASS to raise the score.
• Score only what can be evaluated.
• Report coverage percentage beside score.

Example:

Quality score: 88/100
Evaluation coverage: 74%
Hard gates: PASS on observed scope
Unknown critical controls: T-11, M-09
Decision: CONDITIONAL: safe fixes applied; telemetry-dependent claims deferred

L Phase L: Measure and learn

Track before/after at page/template/topic level:

Search

• impressions
• clicks
• CTR
• query coverage
• indexed/canonical state
• CWV

AI visibility

• generative AI impressions where official reports provide them
• citations
• brand mentions
• explicit recommendations
• choice share (internal heuristic: how often the brand is selected as the best fit within controlled commercial prompts)

Business

• qualified leads
• demo/bookings
• purchases
• revenue/pipeline
• conversion rate
• assisted conversions
• return/refund/churn where relevant

Experiment record

experiment:
hypothesis: "..."
pages: []
baseline_window: "..."
changes: []
launch_date: "..."
evaluation_window: "..."
metrics: []
result: "win|neutral|loss|inconclusive"
decision: "keep|iterate|revert"