AI Sales Website and Lead-Journey Audit: A Manual 12-Check Worksheet
The most visible website problem is not always the most important sales problem. A form may work while follow-up has no owner; a chat may answer questions while losing the buyer's requested next step; a CRM may receive a record without the evidence needed to route it.
This manual audit follows one real lead from first visit to final disposition. It produces a prioritized repair list, not a conversion or ROI forecast.
Choose one journey and collect evidence
Define the audience, offer, entry page, conversion action, channels, destination team, and a recent review window. Use a test record that the team can identify and remove. Do not submit sensitive or deceptive data.
Journey audited:
Audience and intended next step:
Entry URL and device:
Test identity/reference:
Review window:
People observing website / routing / sales / CRM:
Google's form guidance covers accessible and usable data entry, while Google Analytics describes funnels as defined steps whose completion and abandonment can be observed. Those mechanisms help gather evidence; they do not diagnose the cause or promise an uplift. Review form guidance and funnel exploration.
Run the 12 checks
Mark each check pass, fail, or unknown. A screenshot, event, record, timestamp, transcript reference or owner confirmation must support the mark.
| Stage | Check | Evidence to capture |
|---|---|---|
| Intent | The page states who the offer is for and the next step | URL, message and CTA |
| Intent | Claims and expectations match the actual handoff | Claim and operating policy |
| Capture | Required fields are necessary, clear and usable | Form/chat replay and errors |
| Capture | Permission and channel expectations are visible | Exact copy and stored state |
| Qualification | Questions change an approved decision | Question-to-rule map |
| Qualification | Uncertainty routes to a person | Trigger and acceptance event |
| Routing | One current owner or monitored queue receives the lead | Rule version and owner record |
| Routing | Duplicate, missing and rejected routes have recovery | Replay and exception record |
| Follow-up | The buyer's requested next step and timing survive handoff | Handoff payload and timestamp |
| CRM | Source, goal, evidence, disposition and correction are structured | Field map and test record |
| Measurement | Start, denominator, outcome window and quality are defined | Metric specification |
| Governance | A named person can pause, correct and roll back the flow | Runbook and access record |
OpenAI's agent guide emphasizes bounded tasks, tools, guardrails, human intervention and evaluation. Apply those controls even when the “AI” part is only one step in a larger sales journey. Read the guide.
Prioritize findings transparently
For every fail or material unknown, assign three whole-number ratings:
In plain language, ask three questions: how badly does the issue block the journey, how much of the chosen path could face it, and how strong is the evidence? Those answers become severity, exposure and confidence:
- Severity:
1inconvenience,2blocks or corrupts a decision,3creates material customer, data, compliance or ownership risk. - Exposure:
1uncommon branch,2recurring segment,3default/high-volume path. - Confidence:
1hypothesis,2repeated observation,3reproduced and traced to evidence.
Priority = Severity × Exposure × Confidence
Valid results are integers from 1 to 27. Before scoring, the team records its own action thresholds, required evidence, approver and stop conditions. The method supplies no universal bands. Safety, privacy, legal or unowned-handoff stops override the score.
Completed fictional audit
Northstar chooses fictional thresholds before scoring: 18–27 repair now, 8–17 next cycle, and 1–7 evidence queue. These are Northstar's choices, not defaults.
Journey audited: Mobile bilingual service inquiry from paid search to accepted sales handoff
Audience and next step: Vietnam operations leaders requesting a discovery call
Entry URL and device: /services, mobile 390px
Test reference: E-204
Review window: 2026-08-09 to 2026-08-15 ICT
Observers: Web owner / routing owner / sales owner / CRM owner
Intent-1=pass (source and CTA); Intent-2=pass (policy replay)
Capture-1=pass (mobile/error replay); Capture-2=pass (permission record)
Qualification-1=pass (question-rule map); Qualification-2=pass (handoff accepted)
Routing-1=pass (queue record); Routing-2=pass (duplicate replay)
Follow-up=fail (Vietnamese preference absent from handoff/CRM)
CRM=fail (language field missing in E-204)
Measurement=pass (metric specification); Governance=pass (pause/rollback owner)
| Finding | Severity × exposure × confidence | Score | Decision |
|---|---|---|---|
| Vietnamese follow-up is absent from CRM | 2×3×3 | 18 | Repair now under Northstar's pre-set gate |
| Confirmation lacks change/help route | 1×3×3 | 9 | Next cycle under Northstar's pre-set gate |
Evidence: the first finding uses replay E-204 plus its CRM record; the second uses three device replays. The data owner owns the first repair and the web owner the second.
Recalculation: 2 × 3 × 3 = 18; 1 × 3 × 3 = 9. Northstar fixes the missing language field first. It does not claim the change will increase conversion; success means the field survives capture, routing, acceptance and CRM write in the retest.
Invalid states and sensitivity
Do not score a finding when there is no observable event, no defined journey, or no owner able to verify it. Keep it unknown and assign evidence collection. If the first finding's exposure is only 2, its score becomes 12, but it still precedes the score-9 issue. A privacy or safety stop remains first even with a lower arithmetic score.
Turn the audit into a repair queue
Finding:
Evidence link/reference:
Priority components and score:
Immediate containment:
Root cause to verify:
Owner and due date:
Acceptance test:
Rollback trigger:
Retest result:
Close a finding only after the same journey passes and downstream records agree. A UI change without a routing/CRM retest is incomplete.
Completed repair record for the first finding:
Finding: Vietnamese follow-up preference absent from CRM
Evidence: E-204 replay and CRM record
Priority: 2 × 3 × 3 = 18; Northstar repair-now gate
Containment: Bilingual queue checks preference manually
Root cause: CRM mapping omits preferred_language
Owner/due: Data owner / 2026-08-18
Acceptance: preference survives capture, routing, acceptance and CRM write
Rollback: mapping creates duplicate/invalid records
Retest: pending; finding remains open
Frequently asked questions
These questions clarify what the score and one test journey can—and cannot—show.
Does the highest score always come first? No. Stop conditions override scores, and the team must set thresholds before scoring.
Can one test lead represent all traffic? No. It verifies the path, not prevalence. Use production-safe samples and segments to assess exposure.
Does fixing a failed check prove more revenue? No. It proves only that the declared acceptance test passed.
Next steps, method, and limits
Choose the next worksheet. Use the website conversion audit when the defect is confined to the on-site path. Use the qualify-and-route template when the gap begins after capture.
Method and limits. The page stores no inputs and uses no benchmark defaults. Ratings are ordinal judgments that require evidence and independent review. The audit cannot establish causation, forecast uplift, certify compliance, or prove an Easy AI capability.