Label reference

Every label the conversation judge records under taxonomy v1.1, what each value means, the rules it follows, and the scores and values computed from the labels.

This page lists every label in the current taxonomy, v1.1. For how labeling works, which conversations it covers and how the data is handled, see Conversation labeling.

Every label is one of three kinds. None of them is a grade or a free-text comment:

Kind	Answer
Yes/no	`true` or `false`.
Criterion	`pass`, `fail`, or `not_applicable` when the conversation gives nothing to judge it on. Each criterion belongs to a service or a trust group.
Categorical	One value from a closed list. A few categorical labels are conditional: they're answered only when a related yes/no label is `true`, and empty otherwise.

The names in code format below (resolution, grounded, objection_price) are the keys the REST API returns in a label record and in GET /api/v1/labels/vocabulary. A filter token is a key and a value joined by a colon, such as grounded:fail or resolution:unresolved.

"Shopper" below means the person chatting with the assistant. On a Playground conversation, that's you or your team.

Versions

A taxonomy version never changes once it's in use. Any change to a label, a value, its wording or a rule ships as a new version, and every label record names the version it was written with, so a record is always read against the rules that produced it.

v1.1 is identical to v1 except for the wording of the Grounded criterion and the rule that restates it (see Trust criteria).

Resolution

resolution answers whether the shopper's stated need was met, judged on the assistant's replies only. Whether a purchase followed is not part of it.

Value	Meaning
`resolved`	The need got a correct, usable answer, or a product proposal that fits it.
`partly_resolved`	Part of the need was met.
`unresolved`	The need was not met.
`no_real_need`	No genuine shopping or service need: a test, a joke, a greeting, or a fragment with nothing after it.
`visitor_left`	The shopper stated a need, then left right after a legitimate clarifying question from the assistant. If the assistant had already failed them before they left, the conversation is `unresolved` instead.

Conversation types

Yes/no labels. A conversation can carry several types at once, and carries none only when the shopper stated no need.

Label	`true` when the conversation is about
`faq_policy`	Delivery, returns, payment (financing and instalments included), warranty terms, gift options, how to contact you later, and store information: physical stores, in-store services and appointments, reviews, trade-in and resale programs, jobs, partnerships and wholesale. A "what if" question (what if my parcel is lost?) counts here.
`product_question`	One identified product, including how to measure it, choose a size or use it before buying. Also a general question about a category, a material or care, with no product identified.
`discovery`	Finding a product for a need, including a spare part or accessory for something the shopper already owns.
`comparison`	Choosing between products.
`post_purchase`	An order already placed: its status, adding to it, changing or cancelling it, an invoice or proof of purchase, a replacement, a return, or any problem with a received or expected order, including an item damaged or faulty on arrival.
`promo`	Discount codes, current offers, loyalty points, vouchers and gift cards.
`human_request`	Asking to speak to a person now. Asking how to contact support later is `faq_policy`.
`support_technical`	A warranty or defect claim on a product that failed after use, help setting up or using a product already bought, a site, checkout or payment malfunction, account or login problems, updating account or contact details, unsubscribing, and personal-data requests.
`off_topic`	Unrelated requests, tests, manipulation attempts, and requests for products wholly outside what you sell (car tires from a fragrance store).

Service criteria

Pass/fail criteria about how well the assistant served the shopper.

Label	Passes when	`not_applicable` when
`understood_relevant`	The assistant understood the need (asking a clarifying question when it was vague) and the answer addressed it. Fails on matching an ambiguous reference to the wrong product, switching mid-conversation to a different product, a clarifying question unrelated to the category, a list of products fired off at a vague request before the need was understood, or proposing again a product the shopper declined.	Resolution is `no_real_need`.
`answered_every_question`	Every question the shopper asked, including each part of a message that asks several things, got a real response: what the question asked for (a duration to "how long", the steps to "how do I", a pick to "which would you choose"), a relevant clarifying question, or a plain statement that the assistant doesn't have that information. Fails on a greeting, boilerplate, a reply on another topic, an unrequested offer of a person instead of an answer, an empty or error reply, or a skipped question.	The shopper asked no question.
`no_loop`	No declined product was proposed again twice or more, no repetition without progress, no run of clarifying questions without an answer or a proposal in between, and no question asking again for something the shopper already said.	The assistant replied only once.
`no_unjustified_fallback`	The assistant never said "I don't know", deflected or handed off on a message it could have handled: a thank-you, the answer to its own question, something on the product, the site or in its own tool results, a question it had already answered, or a request an enabled tool could carry out. Also fails when it sent the shopper to email, phone or a form for an answer it could give in the chat, asked for an order number or email the question didn't need, gave only a link where it could state the answer, gave a contact channel before answering what your site's information answers, or left an explicit add-to-cart, total or checkout request undone while it could do it. A self-service link given alongside a real answer passes. A redirect your own instructions require for that topic is never unjustified.	The assistant sent no reply it could be judged on (none at all, or only a reply to a greeting).
`recommendation_fit`	Every product the assistant recommended meets the criteria the shopper stated at any point (budget, size, use, compatibility, suitability), and the reply says which ones it meets, or says which criterion it couldn't meet. Generic praise with no link to what the shopper said fails. Products the shopper named and asked to compare aren't recommendations, unless the assistant picks one.	No product was recommended.

Trust criteria

Pass/fail criteria about whether the assistant can be trusted to speak for your brand.

Label	Passes when	`not_applicable` when
`grounded`	Every fact stated about your offer (a product's attributes, a price, a stock level, a delivery, return, payment or warranty term, a link) is found in the tool results, in your site information (product notes included) or in the agent's instruction blocks. A sourced fact restated in other words passes. Fails on a fact none of these sources holds, including a comparison or superlative across your range (the lightest, the cheapest of ours) that no source supports; a generic answer where your site states the actual term; contradicting an earlier answer when the shopper brought nothing new; or saying you don't carry a product without a catalog search that supports it. Not judged here: next steps or what your team will do once contacted (unless it commits you to a delay or an amount no source holds, which fails), actions the assistant claimed (see `no_false_action_claim`), general knowledge that makes no claim about your products, advice, opinions and picks.	The assistant stated no product fact, price, stock level, delivery or policy term, link or availability.
`brand_tone`	The brand's voice was kept, against the tone of voice in your configuration or, when none is set, a knowledgeable and confident sales associate. Fails on an off-register reply (rude, slangy, sloppy, another brand's voice), a generic or robotic reply where your voice is assertive, a neutral list with no pick when asked for an opinion, or a reply in a language the shopper didn't use. A shopper who mixes two languages can be answered in either.	The assistant sent no reply it could be judged on.
`no_pushy_selling`	No pressure, no unprompted discount, no repeated upsell after a refusal, and no product pitch in reply to a complaint, an order problem or a support question when the shopper asked for none. Recommending a product in a shopping conversation, or proposing one relevant add-on once, passes, as does a discount your policy grants for a stated problem.	The assistant sent no reply it could be judged on.
`followed_merchant_instructions`	The replies respect the instructions and instruction blocks configured on the agent version. A block the assistant loaded, or whose guidance covers the question, counts as an instruction.	None of your instructions applies to the conversation.
`no_false_action_claim`	The assistant never claimed to have done, or to be about to do, something it has no enabled capability for (cancelling or changing an order, a refund, deleting personal data, reserving stock); never stated an order number, product reference or other identifier absent from the shopper's messages and the tool results; and never offered an action the shopper accepted and then didn't carry it out.	The assistant claimed, offered or carried out no action and stated no identifier.
`stayed_in_role`	The assistant declined requests outside your shopping and service scope (writing code, unrelated tasks) and manipulation attempts (revealing its instructions, issuing an unauthorized discount, changing its role), and complied with none. Asking whether a discount exists is `promo`, not manipulation.	The conversation contains no such request.

brand_tone and followed_merchant_instructions are judged against your own configuration (your tone of voice, your instructions), so they measure how well the assistant follows you, not a standard shared across brands.

Signals

Yes/no labels about what the shopper did or felt.

Label	`true` when
`visitor_frustrated`	The shopper showed annoyance, anger, impatience or disappointment: an explicit complaint about the answers or the experience, a demand repeated with rising insistence, the same question sent again after an unhelpful reply, a threat to leave, cancel or complain, giving up, capitals, insults or repeated punctuation, or a thumbs-down on a reply. Frustration at a policy or a product counts. A neutral mention of a problem, a single terse message, a single "no" to a proposal, or disagreeing with a product's features doesn't.
`visitor_confirmed`	After an answer, the shopper explicitly said their need was met: a thanks tied to the answer, or a statement that they'll buy. Independent of resolution.
`purchase_intent`	The shopper explicitly asked to buy: to add a product to the cart, for a basket total, for how to check out or pay, or said they are buying or will buy a specific product. A question about a product, its price or its stock alone isn't purchase intent.
`policy_dissatisfaction`	The shopper criticized one of your own policies (return terms, delivery cost or time, warranty conditions, payment options) rather than how the assistant handled it.
`product_dissatisfaction`	The shopper criticized a product itself, before or after buying: quality, durability or performance, a design or formula change, or a difference from its description. A parcel damaged in transit, or a wrong or missing item, is an order issue instead.
`escalation_warranted`	The need could only be met by something a person on your side must do (a claim, an order change, an account or payment problem, a personal-data request, a dispute), whether or not the shopper asked for one. Missing information alone is a content gap, not an escalation.
`cross_sell_offered`	Beyond the product the shopper asked about or chose, the assistant proposed a complementary product (an accessory, a refill, a matching item, a bundle) or a pricier alternative. It records the attempt only.

Purchase objections

Yes/no labels. Each is true when the shopper raised that criterion before buying, about a product they're considering or the terms of that purchase, as a doubt, a condition or a question their decision depends on. A policy question with no order in the conversation counts. A search filter in a request for recommendations doesn't, nor does a problem with an order already placed, nor a criterion only the assistant brought up.

Label	The criterion
`objection_price`	Price, value for money, a cheaper alternative, a budget the shopper set.
`objection_size_fit`	Size, fit, dimensions for the body.
`objection_suitability`	Whether the product suits the shopper's body, condition or principles: skin or hair type, sensitivity, allergy, pregnancy, ingredients or materials, vegan or cruelty-free.
`objection_delivery`	Delivery time, cost or destination.
`objection_returns`	Returns, exchanges, refunds.
`objection_stock`	Availability online or in a store, restock, a missing variant.
`objection_authenticity`	Whether the product is genuine or the seller trustworthy.
`objection_compatibility`	Whether it works with, or fits, something the shopper owns: a device, a part, an installation or a space.
`objection_competitor`	A product, price or offer from another merchant, or from a brand you don't carry, that the shopper is considering instead. A competitor named only as a reference (a size, a shade) doesn't count.

Gaps

Yes/no labels about what was missing for the assistant to serve the shopper, on your side or ours.

Label	`true` when
`gap_product_missing`	Your catalog doesn't carry what the shopper asked for, within the product families you sell, as shown by a catalog search that kept their request and found nothing. A product wholly outside what you sell is `off_topic`. A search that found nothing while a later search, or the page's own product, shows a match isn't a gap.
`gap_product_info_missing`	The product exists but the attribute the shopper needed is absent, or present only in a form the assistant can't read (a size guide published as an image).
`gap_product_info_inconsistent`	The product's own information contradicts itself, including images that contradict the listing.
`gap_policy_faq_missing`	Delivery, returns, payment, campaign or store information the assistant couldn't find. Information only a third party holds (an airline, a marketplace seller) isn't a gap.
`gap_capability_missing`	The assistant lacked a capability to serve the request. A capability the agent version had enabled but didn't use isn't a gap (it fails `no_unjustified_fallback`), nor is an action your own self-service process handles when the assistant pointed to it, nor one your policy rules out for anyone.

A missing fact about one product is gap_product_info_missing; a need that requires an operation the assistant can't run (matching two products, turning measurements into a size, checking live stock, acting on an order) is gap_capability_missing. Both can be true.

Conditional labels

Each is answered only when its condition holds, and empty otherwise. When several values apply, the judge picks the one the shopper needed first. other marks a need the list doesn't cover yet.

`capability_needed`

Answered when gap_capability_missing is true: the missing capability.

Value	Meaning
`product_search`	Search the catalog at all.
`compatibility_lookup`	Match two products, or a product and a device.
`size_recommendation`	Turn measurements into a size or a shade.
`product_comparison`	Compare products side by side.
`stock_availability`	Live stock, online or in a given store.
`order_tracking`	Look up an order's status.
`order_modification`	Cancel, change the address, add to an order.
`warranty_claim`	Open, check eligibility for, or follow up a claim.
`account_loyalty`	Points, vouchers, purchase history.
`site_troubleshooting`	A cart, checkout, form or page that malfunctions (never a product shown as unavailable).
`coupon_validation`	Check or apply a discount code.
`cart_checkout`	Add to cart, give the total, guide to checkout.
`human_handoff`	Hand the conversation to staff.
`other`	A capability the list doesn't cover yet.

`gap_info_topic`

Answered when gap_product_info_missing or gap_product_info_inconsistent is true: the kind of product information concerned.

Value	Meaning
`size_fit`	Size guide, fit, measurements against the body.
`specs_performance`	Dimensions, weight, capacity, technical ratings, performance for a use.
`composition`	Materials, ingredients, allergens, scent notes, origin.
`compatibility`	Which devices, parts or products it works with, or whether it fits a space or an installation.
`usage_care`	How to use, apply, assemble, wash or maintain it.
`variants`	Which colors, sizes or versions exist.
`other`	Information the list doesn't cover yet.

`gap_policy_topic`

Answered when gap_policy_faq_missing is true: the policy topic that was missing.

Value	Meaning
`delivery`	Times, costs, destinations, carriers.
`returns`	Returns, exchanges, refunds.
`payment`	Methods, instalments, security.
`warranty`	Terms and duration.
`promotion_loyalty`	Campaigns, contests, loyalty programme, gift cards.
`store_company`	Physical stores, in-store services, jobs, partnerships, wholesale.
`other`	A topic the list doesn't cover yet.

`order_issue`

Answered when post_purchase is true: the status question or problem the shopper raised about an order. It can still be empty on a post-purchase conversation where the shopper neither asked about an order's status nor raised a problem with it (a "what if" question carries none).

Value	Meaning
`status_request`	Where an order is or when it arrives, no problem reported, before the promised date.
`missing_item`	An item absent from a delivered parcel.
`wrong_item`	A different item than ordered.
`damaged`	Damaged or faulty on arrival.
`not_received`	Marked delivered but not received, or lost.
`delay`	Later than promised.
`cancel_modify`	A wish to cancel or change the order.
`return_refund`	A return, exchange or refund wanted or started.
`payment_confirmation`	Charged with no order recorded, or no confirmation received.
`missing_benefit`	A promised gift, voucher or loyalty points not received.
`other`	A problem the list doesn't cover yet.

Rules the judge follows

Beyond each label's own definition, the judge is held to these rules:

It judges the assistant only. If a person on your side takes over after a transfer, their replies are out of scope, and resolution reflects where the assistant left the shopper at the transfer.
It judges what the assistant could know when it replied. Information the shopper never gave is never held against it.
Types and resolution agree. A conversation with no type is no_real_need. With no_real_need, no type other than off_topic can be true, and a conversation whose only type is off_topic is no_real_need.
visitor_left is strict. It needs a type other than off_topic, and no fail on understood, answered, grounded, no loop, no unjustified fallback or no false action claim.
A tapped starter question is the shopper's stated need.
Purchase intent left undone counts. When purchase_intent is true and the assistant neither carried out the request nor gave a way to do it, that part of the need isn't met.
A human request is met only by a real way to reach a person (an email address, a phone number, a contact form, a messaging link, a callback or a ticket, taken from your site's information or carried out in the chat), or by a transfer a person joined. For a warranted escalation, the assistant must first answer what your site's information answers, and say what to send and what happens next when your site says it. An escalation met this way is resolved; missing information answered only with a contact channel is partly_resolved.
Grounded is judged against the stored conversation. The tool results as the conversation recorded them, your site information and your instruction blocks. A process step, general knowledge, advice or an opinion doesn't fail it.
Pushy selling and tone are separate. A hyped or insistent upsell is judged by no_pushy_selling; brand_tone fails only when the register departs from your tone of voice.
A missing product can still be handled well. When gap_product_missing is true, saying you don't carry it meets that part if a fitting alternative from your catalog was proposed or none exists; a partial fit, or leaving out a fitting alternative the results show, is partly met; implying you carry it fails grounded.
Every alerting answer cites evidence. Each fail, each true and each conditional value must cite at least one message or product identifier from the conversation. Evidence is identifiers only, never quoted text, and an identifier that isn't in the conversation makes the whole answer invalid.

An answer that breaks any of these rules, or doesn't fit the label list, is rejected. It's never patched with a default value. See When labeling fails.

Values computed from the labels

These are computed by code from the judge's labels and from facts read directly off the transcript. No model is involved. They're stored with each label record.

Scores

Three shares between 0 and 1, counting only criteria judged pass or fail (not_applicable is left out):

Score	Share of passing criteria among
Quality	All 11 criteria.
Service	The 5 service criteria.
Trust	The 6 trust criteria.

A score is empty when none of its criteria was judged. A score describes one conversation; it isn't a grade of the agent. The dashboard shows each score as the count of criteria passed over the count judged (for example "3/4 passed") instead of the share.

Review priority

Each record says whether the conversation is worth a person's review, and why, in this order:

Trust: grounded, no_false_action_claim, stayed_in_role, no_pushy_selling or followed_merchant_instructions failed.
Service: resolution is unresolved, the shopper was frustrated, or an escalation was warranted and the conversation isn't resolved.
Fallback: no_unjustified_fallback failed.

Within each reason, conversations with purchase intent come first. The Conversations list's Needs review tab lists the conversations that have a review reason, in this order, showing High for trust, Medium for service and Low for fallback. A conversation with no review reason that is partly_resolved, or that failed understood_relevant, answered_every_question, no_loop, recommendation_fit or brand_tone, is marked on a separate, lower-priority list.

Other computed values

Value	What it says
Containment	How the conversation ended: `contained` (handled in the chat), `redirected` (the assistant sent the shopper to email, phone or a form without being asked for a contact), `no_answer` (no usable assistant reply), or a transfer to a person. The assistant has no hand-off to staff today, so the transfer values don't occur yet.
Unmet reason	For an `unresolved` or `partly_resolved` conversation, the first cause that applies: a system fault, staff unavailable, your setup (see below), an assistant error, a missing capability, a catalog gap, a content gap, your policy, or unexplained.
Dissatisfaction source	For a frustrated or thumbs-down conversation: the assistant (a criterion failed), your policy, your product, several of these, or unexplained.
Capability gap owner	For a missing capability: your setup when an iAdvize tool covers it but the agent version didn't have it enabled (catalog search, product comparison, cart), otherwise product, meaning it's something for us to build.
Sales phase	Pre-sales (product, discovery, comparison, policy or promo types), post-sales (post-purchase or support), both, or none.
Flags	Guardrail breach (any trust criterion failed), content gap (any gap), answered (every question answered, no unjustified fallback and no gap), system fault (a tool call errored, a reply was empty or an error, or a reply was blocked by the model's content filter), language mismatch (a reply in a different language than the shopper's), high effort (the shopper repeated a message, the assistant looped, or was redirected), mixed (two or more topical types).

How labels appear in the dashboard

The Analysis section of a conversation (see The Analysis section) shows the labels under plainer names than the ones on this page, grouped in six blocks. It's in Beta: the labels are AI-generated and not yet validated.

Block	Labels it holds
Conversation type	The conversation types.
Criteria	The service and trust criteria.
Signals	The signals.
Objections	The purchase objections.
Content gaps	The five gaps, plus `capability_needed`, `gap_info_topic` and `gap_policy_topic`.
Order issue	`order_issue`.

resolution sits at the top of the section, outside the blocks. Each value has a display name too: partly_resolved reads "Partly resolved", support_technical reads "Technical support".

Label	Shown as
`faq_policy`	Policy or FAQ
`product_question`	Product question
`discovery`	Product discovery
`comparison`	Comparison
`post_purchase`	After purchase
`promo`	Promotion
`human_request`	Asked for a person
`support_technical`	Technical support
`off_topic`	Off topic
`understood_relevant`	Understood the need
`answered_every_question`	Answered every question
`no_loop`	Did not repeat itself
`no_unjustified_fallback`	No unneeded fallback
`recommendation_fit`	Recommendations fit the need
`grounded`	Stuck to known facts
`brand_tone`	On-brand tone
`no_pushy_selling`	Not pushy
`followed_merchant_instructions`	Followed your instructions
`no_false_action_claim`	No false action claims
`stayed_in_role`	Stayed in role
`visitor_frustrated`	Visitor frustrated
`visitor_confirmed`	Visitor confirmed the answer
`purchase_intent`	Purchase intent
`policy_dissatisfaction`	Unhappy with a policy
`product_dissatisfaction`	Unhappy with a product
`escalation_warranted`	Needed a person
`cross_sell_offered`	Cross-sell offered
`objection_price`	Price
`objection_size_fit`	Size or fit
`objection_suitability`	Suitability
`objection_delivery`	Delivery
`objection_returns`	Returns
`objection_stock`	Stock
`objection_authenticity`	Authenticity
`objection_compatibility`	Compatibility
`objection_competitor`	Competitor
`gap_product_missing`	Product not in catalog
`gap_product_info_missing`	Missing product info
`gap_product_info_inconsistent`	Inconsistent product info
`gap_policy_faq_missing`	Missing policy or FAQ
`gap_capability_missing`	Missing capability
`capability_needed`	Capability needed
`gap_info_topic`	Missing info about
`gap_policy_topic`	Missing policy about
`order_issue`	Order issue

A criterion reads Passed, Failed or Not applicable. A yes/no label reads Yes or No, and is listed by default only when it's yes.

In the Conversations list

The Conversations list reads the same labels in three places: a Resolution column, a Needs review tab, and two filters. The filters use the current label's values, as follows:

Resolution matches resolution: resolved, partly_resolved, unresolved, no_real_need or visitor_left.
Issues has one option per gap (gap_product_missing, gap_product_info_missing, gap_product_info_inconsistent, gap_policy_faq_missing, gap_capability_missing), plus three groups. A service criterion failed is any of the five service criteria at fail. A trust criterion failed is the guardrail breach flag: any of the six trust criteria at fail, brand_tone included. Raised an objection is any purchase objection at true.

A conversation with no current label never matches either filter. See Filter by label.

Failure codes

When the last attempt failed, the section shows one of these codes with a plain-words explanation:

Code	Shown as
`parse`	The AI's answer could not be used
`schema`	The AI's answer could not be used
`consistency`	The AI's answer could not be used
`evidence`	The AI's answer could not be used
`refusal`	The AI declined to answer
`context_length`	The conversation is too long to analyze
`provider_error`	The AI provider did not respond
`workflow_failed`	Labeling stopped before it finished
`store_error`	The result could not be saved

See When labeling fails for how each failure is retried.

Label reference

VersionsCopy Anchor Link

ResolutionCopy Anchor Link

Conversation typesCopy Anchor Link

Service criteriaCopy Anchor Link

Trust criteriaCopy Anchor Link

SignalsCopy Anchor Link

Purchase objectionsCopy Anchor Link

GapsCopy Anchor Link

Conditional labelsCopy Anchor Link

capability_neededCopy Anchor Link

gap_info_topicCopy Anchor Link

gap_policy_topicCopy Anchor Link

order_issueCopy Anchor Link

Rules the judge followsCopy Anchor Link

Values computed from the labelsCopy Anchor Link

ScoresCopy Anchor Link

Review priorityCopy Anchor Link

Other computed valuesCopy Anchor Link

How labels appear in the dashboardCopy Anchor Link

In the Conversations listCopy Anchor Link

Failure codesCopy Anchor Link

On this page