Agent Journeys
Being recommended isn’t the same as being able to
buy it
Your next customer may send an AI agent first. Write that customer’s goal in plain language. Agent Journeys sends a browser agent to your site to attempt it, then scores whether the task actually completed — with the trace, the screenshots, and the exact point where it broke down.
- Included in every plan
- Stays on the domain you set
- Stops before checkout
Customer goal
I need a rain jacket that works for hiking in cold weather, shipped before the 14th.
Starting URL
example-store.com/outerwear
Success condition
Correct item in cart
Runner profiles
Guardrails
Simulated shopping-assistant profiles styled after Amazon’s Alexa for Shopping (formerly Rufus) and Walmart’s Sparky. These are Omni Impact’s own agent configurations — not those assistants, and not affiliated with, endorsed by, or operated by Amazon or Walmart.
The other half of the funnel
You can be cited everywhere and still lose the transaction.
A brand can be recommended constantly and still lose agent-driven transactions, because the agent cannot determine compatibility, delivery timing, eligibility, booking availability, or return rules. Visibility measures the recommendation. Journeys measures the outcome.
01
Discover
02
Recommend
AI Visibility
03
Understand
04
Act
05
Complete
Agent Journeys
AI Visibility
Can AI systems discover, understand, cite, and recommend the business?
Prompts, mentions, citations, competitive share of voice, content and structured-data opportunities.
Agent Journeys
After an agent selects the business, can it accomplish the customer’s intended task accurately and successfully?
Journey runs, completion scores, traces, screenshots, blockers, incorrect interpretations, and recommended fixes.
How a run works
One run, end to end.
Every journey follows the same path, and every score traces back to the run, the steps, and the artifacts that support it.
Validate
Before anything runs, the platform checks entitlement limits, the journey configuration, and the domain allowlist. A journey stays on the domain you set.
Execute
An isolated browser session launches and runs the versioned runner profile. Each action and observation is recorded as an ordered step, with evidence uploaded as it goes.
Score
Deterministic success checks run first — observable URL, DOM, and application state. Model-assisted grading follows, for interpretation and qualitative diagnosis rather than as the only proof of completion.
Report
Findings are deduplicated against existing open issues, so a recurring problem stays one item with a first-seen and last-seen date. Reports update and notifications dispatch.
Scoring
Six dimensions, each tied to evidence.
A pass or fail on its own is not useful. Every dimension records the grader version, prompt version, model, inputs, and confidence — so a score from six months ago is still interpretable today.
Task completion
Whether the defined task reached the permitted success state.
Preferred evidence
Deterministic URL, DOM, and state checks plus the final trace.
Selection quality
Whether the agent chose the right product, service, configuration, or path.
Preferred evidence
Ground-truth comparison and evidence of the selected entity.
Factual accuracy
Whether policies, pricing, compatibility, timing, and requirements were represented correctly.
Preferred evidence
Your configured truth records against the page content actually cited.
Friction
Extra steps, loops, navigation failures, inaccessible content, or ambiguous forms.
Preferred evidence
Step count, retries, error events, screenshots, agent observations.
Action readiness
Whether the information and controls needed to finish were actually available.
Preferred evidence
Missing fields, inaccessible state, blocked or unsupported interactions.
Consistency
Whether results change across runner profiles or across repeated runs.
Preferred evidence
A comparable set of runs measured over time.
Deterministic where possible
If a success condition can be expressed as observable application state, it is checked that way — not inferred by a model.
Failures are classified
Execution, website, model limitation, authentication, bot protection, and infrastructure failures are kept distinct.
No speculative revenue claims
Lost-revenue estimates are not presented as fact. Where impact is modelled, the assumptions and method are labelled.
Templates
Start from a journey that already knows where to stop.
A non-technical user should be able to define a journey without writing a browser test. Each template translates a plain-language goal into a versioned execution plan — including the point past which it will not go. The product also offers a growing library of journey templates across eight categories.
01 / 5
Goal
Find the correct product for stated requirements, validate compatibility, and add it to the cart.
Safe stopping point
Stops before checkout.
What gets checked
- Requirement match
- Compatibility validated
- Correct variant selected
- Cart state confirmed
Safeguards
An agent on your live site, with limits it cannot exceed.
Every boundary below is enforced at execution, not left to the model’s judgement. The run stops at any control that would place an order, confirm a booking, send a message or create an account.
Stays on the domain you set
Journeys are restricted to the domain you set. The allowlist is checked before the run starts, and the browser is turned back if it tries to leave the domain.
Irreversible actions are gated
Runs stop by default before payment, reservation confirmation, account deletion, message sending, or production ticket creation.
No protection bypass
CAPTCHAs, bot protection, rate limits, and access controls are never circumvented. When one blocks a run, it is recorded as a finding.
Identified, never disguised
Every request carries our OmniImpactJourneyBot token, and simulated shopping profiles are always labelled as simulations. A run never pretends to be a human visitor.
Sensitive fields are never typed
The agent never enters payment, password, or personal-identity data anywhere, and stops before submitting any form other than a site search — so a run cannot create a lead, order, or account.
Evidence expires
Artifact URLs are short-lived and authorization-checked. Screenshot, trace, video, DOM, and log retention are configured separately.
Reporting
Every score links back to its evidence.
Findings carry a category, severity, recommendation, and the run that produced them. Your team can mark each one valid, a false positive, an accepted risk, resolved, or ignored.
Delivery date not exposed before cart
The agent could not determine whether the order would arrive by the stated deadline, and abandoned the task at the product page.
Return window stated as 14 days
Configured ground truth records a 30-day window. The agent read the shortened figure from a legacy policy page still in the index.
Size guide opens in a modal the agent cannot read
Compatibility information is present but only reachable through an interaction the runner could not complete, adding four retry steps.
Journey health and trend
Completion rate over time for each journey, so a slow regression is visible before it becomes a quarter-end surprise.
Step-by-step replay
The latest run reconstructed action by action, with screenshots, the intended action, what actually happened, and the observation at each step.
Ranked open findings
Issues ordered by severity and recurrence, deduplicated by fingerprint so a persistent problem stays one item with a first-seen date.
Cross-profile consistency
The same journey compared across runner profiles and repeated runs, which is often where a site turns out to work for one agent and not another.
Ground-truth discrepancies
Every place an agent stated a policy, price, or requirement that disagrees with the facts you configured.
Combined funnel view
Recommendation performance and task completion side by side in a single picture, across both halves of the funnel.
Included in your plan
Every plan runs journeys.
Run allowances scale with your tier, and the journey definitions, runner profiles and schedule options scale with them.
Starter
1 run a month
One journey, run on demand.
Pro
10 runs a month
Three journeys, weekly schedule.
Premium
40 runs a month
Ten journeys, daily schedule.
Questions