One-off audits
A single landing page, an individual page, or an email — reviewed in context, not in isolation.
- Landing pages and paid-traffic destinations
- Individual product, pricing and signup pages
- One-off marketing and lifecycle emails
Big Ears · Conversion Agent
Conversion intelligence that understands your organisation.
Not a generic screenshot audit. The Conversion Agent reads your live journeys, analytics, company context, previous experiments and competitor landscape — then identifies, prepares and measures improvements, with a human approving anything that ships.
Organisation-awareEvidence-linkedHuman-approved
Run · landing page audit
context loadedLoaded organisation context
Positioning · analytics · 3 previous tests
Walked the live page
Fold, form, proof placement, contrast
InferenceProof arrives after the decision
Trust signals sit below a full-height hero, so a visitor who bounces at the fold never reaches them. Suggested a placement test — held for human approval.
What you can point it at
One agent, many jobs — each read in full context.
From a single page to a standing loop, the same organisation-aware intelligence is applied. Nothing is judged in isolation.
Journey & funnel audits
Multi-step flows walked the way a real visitor moves through them, with drop-off read across the whole path.
- Acquisition-to-activation funnels
- Checkout and onboarding sequences
- Cross-page narrative and momentum
Lifecycle & sequence audits
Whole email sequences and lifecycle programmes reviewed as a set, for consistency and cumulative effect.
- Welcome and activation sequences
- Re-engagement and winback programmes
- Tone, cadence and offer coherence
Competitor sweeps
Direct, indirect, emerging and adjacent players scanned to separate table stakes from genuine white space.
- Direct and indirect competitors
- Emerging and adjacent challengers
- Category conventions vs. differentiation
Continuous optimisation
A standing loop that learns from human feedback and measured outcomes, so each run starts smarter than the last.
- Feedback memory across runs
- Outcome learning from shipped changes
- Re-prioritised backlog every cycle
Governed access
Connects through MCP, APIs and authorised web UIs — reading the systems you already run, with permission.
- MCP, API and authorised web UI access
- Scoped, revocable credentials
- Read-first, act-only-on-approval
How it reads your work
Every finding is evidence-linked and honestly hedged.
The agent distinguishes what it knows from what it suspects. Pick a surface to see the kind of output a run produces — illustrative, not a real customer report.
Landing page — paid search destination
Reviewed against your positioning, analytics for this path, and the last three tests run on it. Demonstration output only.
ObservationThe primary promise is restated three times above the fold
The headline, sub-headline and first badge row all carry the same claim in slightly different words, so the fold spends its attention repeating rather than building an argument.
Copy High held for human approval
InferenceThe proof section likely arrives after the decision is made
Trust signals sit below a full-height hero, so a visitor who bounces at the fold never reaches them. Placement, not credibility, is the constraint.
CRO Medium held for human approval
HypothesisA single-field form may lift starts on mobile
The current form asks for five fields before any value is shown. A staged first step is worth testing — framed as a hypothesis to validate, not a certainty.
UX Low held for human approval
FactContrast on the secondary CTA falls below 4.5:1
Measured from the rendered page: the ghost button text is #9aa4b2 on white. This is an accessibility fail, independent of any conversion argument.
Accessibility High held for human approval
Fact
Directly verifiable from the rendered page or data.
Observation
Something present and describable, not yet explained.
Inference
A reasoned read from the evidence, stated as such.
Hypothesis
A testable idea, paired with the metric that settles it.
Unknown
Named gaps, with the signal that would close them.
Competitive intelligence
Reads the market to find your opening, not to copy the pack.
Direct, indirect, emerging and adjacent players are swept together, then sorted through four lenses — so the recommendation is differentiation you can hold, not imitation that erases it.
DirectIndirectEmergingAdjacent
Table stakes
Conventions every serious player meets. Matching them buys entry, not distinction — so they are logged, not celebrated.
Parity traps
Places where copying a competitor would cost you your own advantage. The agent flags these before you follow the pack.
Transferable principles
Patterns that work for a reason you can borrow — the underlying principle, adapted to your context rather than cloned.
Strategic white space
Ground no one in the set is holding, mapped against your strengths so the opening is one you can actually own.
From finding to shipped change
Identify, prepare, approve, then learn.
The agent does the preparation and the measurement. A human stays in the loop for anything consequential — and every outcome makes the next run smarter.
01
Identify
Every finding is linked to its evidence and tagged by class — fact, observation, inference, hypothesis or unknown — so nothing reads as unearned certainty.
02
Prepare
Recommendations arrive implementation-ready: rewritten copy, Figma-ready layout direction, and the specific change, not a vague nudge.
03
Approve
Nothing consequential ships without a human. The agent proposes; a person reviews, edits and approves before anything reaches production.
04
Measure & learn
Shipped changes are measured, and both the outcome and your feedback fold into memory — so the next run is sharper than this one.
Nine specialist lenses
Each surface is reviewed the way a full team would review it.
PropositionCopyDesignUXCROTrustAccessibilityPerformanceCompetitive intelligence
Governed by design
Traceable runs, versioned logic, and a human in the loop.
The agent connects to what you already run, reads first, and acts only on approval. Everything it does is versioned and auditable through the Prompt Registry.
Prompt Registry
- Versioned promptsEvery prompt in the registry is versioned, so you can see exactly what logic produced a given run.
- QA & canariesChanges pass QA and a canary before wider use — new logic proves itself on a small slice first.
- Approvals & rollbackPromotions require approval, and any version can be rolled back to a known-good state.
- Run traceabilityEach run records which prompt versions, sources and approvals it used, so any finding is auditable end to end.
run landing-page-audit · surface=/pricing
prompt cro.finding@v14 · qa=pass · canary=pass
context analytics, brand, previous_tests(3)
gate human_approval=required
Access surfaces
MCP serversREST & GraphQL APIsAuthorised web UIsAnalytics & BISession replayCRM & pipelineResearch & brandFigma output
Context it reads
AnalyticsBISession replayCRMResearchBrandWritingPrevious tests
See it on your own pages
Point it at one page and watch it think.
We’ll run the Conversion Agent against a page or journey you choose, walk you through the findings, and show exactly where the human stays in control.