Retail AI agent · Discovery and ranking
Product-Recommendation & Search AI Agent
Rank and recommend inside the ranking policy your merchandising owner approved, serve paid placement only into slots the disclosure template owns, and hold the weights, the boosts and the surfaces for a person.
- Typical delivery
- 4–6 weeks
- Deployment
- Your stack
- Merchandising
- Labelled slots
- After launch
- Agent Care
What this agent does
Ranks inside a policy someone signed
In
- 01
Read the query or session, the catalogue, live inventory and the ranking config from supported search.
- 02
Standardise queries, synonyms and product attributes, and carry each ranking input with the source it came from.
Reason
- 03
Score and order candidates using the approved policy and the feature set signed off alongside it.
- 04
Separate the slots — which are organic, which are monetised, which the disclosure template owns.
- 05
Attach to each surfaced result the inputs that lifted it, commercial signals included.
Decide
- 06
Mark any result whose position rests on a paid, supplier-funded or own-brand signal.
- 07
Send ranking-policy and surface changes to the category merchandising manager rather than applying them.
Out
- 08
Retain the query, the ranking inputs, the disclosure state and the set that was served.
- 09
Execute write actions only inside the approval boundaries agreed during implementation.
Product statement
The agent re-ranks inside an approved policy; the category merchandising manager owns the weights, the boost rules and which surfaces carry paid placement at all.
Example workflow
One session, query to served set
- 1Query or session receivedOn-site search box, category page, personalised carousel or app surface
- 2Signals assembledCatalogue records, live inventory, eligibility state and the signed ranking config
- 3Ranking proposedOrdered results, slot assignment, commercial-signal contribution and confidence
- 4Gates appliedConfig-version checks, feature-set checks, slot and disclosure checks and confidence threshold
No human action required
Stages 1 to 4 run unaided and no ranking policy moves at any of them — the merchandising lane opens at the confidence gate.
Serves inside the approved config.
Takes an ad-operations read first.
Merchandising approval
The ranking is held with its inputs, its marked slots and the confidence.
Approve · Adjust · Send to ad ops
Approved — released to serve
- 6Search and personalisation systems updatedOnly where write access and approval policy allow it
- 7Session evaluatedDisclosure render rate, adjustment rate, relevance outcomes and post-release rollbacks
Reorderings
Every merchandising adjustment is counted in the evaluation.
What should not run autonomously
Human approval stays in control
Outside the boundary — human approval required
8 items
- Approving the ranking policy and its boost weights.
- Deciding which surfaces carry paid placement.
- Writing the sponsored-disclosure wording or its placement.
- Setting the own-brand promotion rule.
Automation boundary
Agent acts unaided
- ✓Re-score and re-rank inside the approved policy and feature set.
- ✓Refresh embeddings.
- ✓Fill approved slot templates from live inventory.
- ✓Emit per-impression decision records and disclosure-render assertions.
Any write happens inside the boundaries agreed at implementation, never ahead of merchandising approval.
- Clearing the feature set against protected-characteristic proxies.
- Varying price, discount or availability by individual shopper.
- Configuring child-directed or mixed-audience surfaces.
- Publishing the ranking-parameter disclosure for EU shoppers.
Example output
One impression, annotated
Everything the agent surfaces carries the inputs that put it where it is.
Ranking output · single impression
Illustrative example- Session
- Logged-out search
- Surfaced result
- Mid-weight running shoe returned third on a head query
- Slot type
- Organic slot
- Commercial signal
- None recorded
- Confidence
- 91%
- Disclosure state
- Not required, logged
As received
Taken from the catalogue, the inventory feed and the signed config — nothing on this side is inferred.
Ranking inputs used
- Query-intent match
- Live inventory state
- Session behaviour
Why this position
Relevance and stock lifted it and no funded signal entered the score.
What the score decides
Below the configured threshold the ranking picks up an ad-operations read before anything is.
Action
ApproveAdjustSend to ad opsValue
Where AI adds value
The same four claims, placed at the point in the workflow where each one applies.
Where the value lands
Value 01 – 04
Every session
From the query and the catalogue
Order inside the approved config
Draw on live inventory, the approved feature set and the ranking policy the merchandising owner signed.
Rank on what is in stock
Head queries in stable categories come back ordered without anyone touching them.
Put the monetised slots in front
Paid, supplier-funded and own-brand placements are marked, so review lands where disclosure risk sits.
Build an evidence trail
The query, the ranking inputs and the surfaced set stay attached to the session.
Integrations
Typical integrations
Five system groups connect to the same agent. Which of them are in scope is decided in discovery.
On-site search
- Algolia
- Constructor
Personalisation
- Bloomreach
- Klevu
Retail media
- Criteo Retail Media
- CitrusAd
- Epsilon
Agent
Recommendation & search
- Reads the catalogue
- Orders the results
- Holds policy changes
Testing and consent
- Optimizely
- Dynamic Yield
- OneTrust
- Ketch
Observability & evaluation
- OpenTelemetry
- Langfuse
- Supported monitoring/evaluation sources
Integration availability depends on the client's existing systems and API access.
Agent controls
Six layers between the model and the shopper
Every layer wraps the next. What none of them catches is in the map below.
L6 · Outermost — last line of defence
Inward → L1 · closest to the model
- L1 – L2
- Test whether a ranking may stand
- L3
- Puts the release in a merchant's hands
- L4 – L5
- Hold the inputs the ranking rested on
- L6
- Falls back to the default ordering when signals slip
How Nestack evaluates it
Evaluate the whole ranking path — not only the top result.
Coverage runs the whole depth of the workflow, and every layer is cut by slice.
Surface — the set the shopper sees
- E1Final-output evaluationWas this the order the approved policy produces, with each label rendered?
- E2Step-level evaluationDid the agent use the signed config, the approved feature set and live inventory?
- E3Tool evaluationDid it read and write the correct index and the correct surface?
- E4Confidence calibrationDo low-confidence rankings actually attract more merchandising adjustments?
- E5Slice evaluationHow does performance change across specific query classes?
- E6Business outcomeHow many impressions carried a commercial placement with no rendered label?
Floor — what the shopper sees ranked
Failure modes
Where each failure originates in the agent
Seven ways the ranking goes wrong, by stage.
Agent lifecycle
Direction of processing →
01 · Retrieval
1 mode
- RK-03
Stale monetisation manifest
A supplier-funded item is scored as organic and served unlabelled.
Stage gathers
Catalogue, inventory, with the source each came from
02 · Reasoning
2 modes
- RK-04
Proxy feature promoted
A location-derived affluence signal enters the score as a proxy.
- RK-06
Review set trimmed by rating
Low-star reviews are down-weighted while the summary reads as complete.
Stage proposes
Ordered results, slot assignment and confidence
03 · Tool / write
2 modes
- RK-02
Config written past sign-off
A new ranking config reaches the feature store before merchandising sees it.
- RK-05
Cohort enrolled without consent
A mixed-audience surface adds sessions with no parental consent on file.
Stage writes
Only where write access and approval policy allow it
04 · Output
1 mode
- RK-01
Disclosure not visible
The sponsored label renders in low contrast and below the fold.
Stage returns
The set the shopper sees and the labels on it
05 · Change / Version
1 mode
- RK-07
Parameters move, text does not
An upgrade shifts the main ranking parameters while the published explanation still describes the last set.
Stage tracks
Model, prompt, ranking config and feature set
- Sev-1 · a held act performed by the agent
- Sev-2 · an unlabelled paid placement is served
- Sev-3 · signals degrade, ranking routes to review
Affected slices
One surface can carry most of the exposure
Start at the all-session baseline: across ordinary traffic the undisclosed-commercial-influence rate sits low. Read the same rate by cohort and a few surfaces carry several times.
Slice performance — reported separately, not only in aggregate
Illustrative example| Slice | Failure rate | Lift | Lift vs. threshold | Status |
|---|---|---|---|---|
| Mixed-audience sessions, sponsored slots | 4.1% | 3.7× | Review | |
| EU and UK regulated categories | 3.1% | 2.8× | Review | |
| Logged-out long-tail queries | 2.0% | 1.8× | Watch | |
| Logged-in head queries, organic only | 1.0% | 0.9× | Normal |
Bar: undisclosed-commercial-influence lift vs. the all-session baseline · scale 0–4.0× · tick marks the 2.0× review threshold
2 of 4 slices over threshold
Evidence-linked improvement
A cycle is measured by what it adds
The loop shuts when the miss is a case in the suite, not when someone has explained it.
Improvement cycle · five stages
Switchback — the path turns at Improve and returns at Learn
- 01
Detect
Undisclosed placements cluster on one surface.
- 02
Diagnose
The category merchandising manager reads the impressions and the configs behind them, and does not stop at the first plausible cause.
- 03
Improve
Whatever changes ships against a version, with the queries that prompted it attached.
- 04
Verify
Nothing ships until the affected cases pass a second time.
- 05
Learn
The suite grows by one case; so does the ranking policy.
Learn → Detect
The return edge. What is detected next is measured against a suite one case longer.
Typical build scope
Twelve workstreams across six weeks
The build scope read against the delivery timeline. Week structure follows the six-week plan — discovery, sources, ranking workflow, evaluation, integration, then production validation and handover.
Workstream
Week 1Week 2Week 3Week 4Week 5Week 6
- 01Ranking workflow discovery and boundary definition.Week 1
- 02Search and catalogue assessment.Week 1, Week 2
- 03Ranking-policy, slot and disclosure-rule mapping.Week 1, Week 2
- 04Catalogue and session ingestion.Week 2
- 05Ranking logic and input binding.Week 2, Week 3
- 06Confidence scoring and routing.Week 3
- 07Merchandising review workflow.Week 3, Week 4
- 08Commerce-platform integration.Week 5
- 09Relevance regression cases.Week 4, Week 5
- 10Guardrails and surfacing controls.Week 4
- 11Session-trail instrumentation.Week 5, Week 6
- 12Deployment, documentation and Agent Care handover.Week 6
12 workstreams · 6 weeks · bar shows the weeks a workstream is active — several run in parallel
Final scope and sequence confirmed in discovery
Engagement tiers
What each tier includes
Rows are the capabilities named in each tier's scope. Higher tiers include everything below them.
| Capability✓ in scope · — not at this tier | PilotOne surface, one market | ProductionProduction search traffic | AdvancedMultiple markets / brands |
|---|---|---|---|
| Introduced at Pilot | |||
| Ranking to your approved policy | ✓ | ✓ | ✓ |
| Merchandising approval | ✓ | ✓ | ✓ |
| Relevance baseline | ✓ | ✓ | ✓ |
| Introduced at Production | |||
| Reporting by query class | — | ✓ | ✓ |
| Approval workflow in your systems | — | ✓ | ✓ |
| Approved config write-back | — | ✓ | ✓ |
| Search-analytics integration | — | ✓ | ✓ |
| Introduced at Advanced | |||
| Multi-region ranking rules | — | — | ✓ |
| Multi-stage merchandising approvals | — | — | ✓ |
| High query volume | — | — | ✓ |
| Enterprise ranking controls | — | — | ✓ |
| Build price | From $5,000 | From $8,000 | Custom quote |
- Final build price
- Confirmed after discovery based on integrations, workflow complexity, transaction volume, approval controls and deployment requirements.
- Separate from build
- Build pricing is separate from recurring Agent Care, which covers managed monitoring, evaluations, incidents and verified improvements after launch.
What we need from you
What you bring, and what we build with it
Each input maps to a piece of build scope and a week in the delivery timeline.
You bring
We build with it
01Your ranking policy and its boost rules
→Ranking-policy, slot and disclosure mapping
Week 102Representative queries and sessions
→Ranking baseline, input binding and slot rules
Week 203Your surface map, organic and monetised
→Surface and disclosure mapping, and the automation boundary
Week 104Access to relevant APIs, feeds or exports
→Search, personalisation and retail-media assessment, then integration setup
Week 205Queries that surfaced the wrong thing
→Relevance cases and the evaluation suite
Week 406Which results may never be reordered automatically
→Confidence bands, surfacing rules and review controls
Week 307Named merchandisers to review the ranking
→Merchandising approval workflow, then pilot and production validation
Weeks 5–6
Nothing else is required
Deployment, documentation and Agent Care handover are ours.
Delivery timeline
Four phases across six weeks
The bands follow real work rather than a plan, so evaluation and pilot genuinely share the fifth week.
Phase
W1
W2
W3
W4
W5
W6
- DiscoveryW1
- BuildW2 – W3
- EvaluateW4 – W5
- Pilot & LaunchW5 – W6
Week focus
W1Ranking workflow discovery, surface mapping and the automation boundary
W2Catalogue and session data connected
W3Ranking logic, confidence scoring and surfacing controls
W4Relevance testing and disclosure guardrails
W5Retail-media integration, pilot surfaces and targeted corrections
W6A live ranking release under merchandising, then Agent Care handover
- Reading the band
- Nothing is padded to fill a week. Evaluation and pilot both land in week 5.
- At the end of W6
- Once the release validates, Agent Care owns the running agent.
- Duration
- Six-week plan shown · typical delivery 4–6 weeks depending on scope confirmed in discovery.
Next step · Retail AI agent
Build a recommendation agent around the policy your merchandiser already signs.
Show us your surfaces, your ranking policy and where paid placement sits today. Then the question the build turns on: which slots may an agent reorder unattended, and which ones does the merchandising owner keep?
- Typical delivery
- 4–6 weeks
- Deployment
- Your search and media stack
- Merchandising owner
- Labelled slots
- After launch
- Agent Care available
Nestack Agents · Recommendation and search
AGT-RT-04 · Agent Care available after launch