CrawlCheck

AI classifier atlas · stage 11 of 16

Groundedness Evaluation

Whether what the model is about to say is backed by a source it read.

13 labels, plus 1 spellings merged into them. Field names gpt-oss writes when asked to describe its evaluation; not OpenAI internals.

13labels
6with a CrawlCheck check

Claim Classes claim_classes

The model sorts the claims it found about a business into types, such as numbers, credentials, availability or opinions, so it can check each type against the right kind of evidence.

claims.verification · also called claim_types, claim_categories

Groundedness Score groundedness_score

A number for how well the claims about a business are backed by evidence the model found. A roofer claiming 30 years with a matching licence record scores higher than one with no backing.

What you can do: Back each claim on your site with something checkable: licence numbers, dates, named certifications, linked sources.

claims.verification · also called credibility_score, support_score · same label written as claim_groundedness

Checked by The claims your site makes: years, prices, refund window · Term: grounding

Grounded Claims grounded_claims

Claims about a business that the model could confirm in evidence, such as a licence on the state register or hours matching across sources. These can be stated with confidence in the answer.

What you can do: State facts in checkable form and keep them consistent across your site and listings.

claims.verification · also called supported_claims, evidence_based_claims

Checked by The claims your site makes: years, prices, refund window · Term: grounding

Ungrounded Claims ungrounded_claims

Claims the model could not confirm, such as best in Denver or a warranty no source mentions. It may leave them out of the answer or add a caution.

What you can do: Replace superlatives with facts that can be checked, and support guarantees with written terms on the site.

claims.verification · also called unsupported_claims, evidence_lacking_claims

Checked by The claims your site makes: years, prices, refund window

Availability Claims availability_claims

Claims about whether a business or product is available, such as same-day appointments, 24/7 service or an item in stock. The model checks them against hours and other sources before repeating them.

What you can do: Only state availability you can meet, and keep hours and booking information current everywhere.

claims.verification · also called accessibility_claims, product_availability

Checked by Your hours, day by day, everywhere they are printed

Used in 26 of 100 census runs this week: “flexible scheduling”

Credential Claims credential_claims

Claims about licences, certifications, training or qualifications, such as a master electrician or a board-certified dentist. The model tries to confirm them before using them in its choice.

What you can do: Give licence and certification numbers and the issuer for each credential, with links to public lookups.

credentials.status · also called qualification_claims, expertise_claims

Numeric Claims numeric_claims

Claims that contain numbers, such as 25 years in business, 4.9 stars or 1,000 roofs installed. The model checks these against other sources and distrusts numbers that disagree.

What you can do: Keep numbers accurate and consistent everywhere, and update them, such as years in business, each year.

claims.verification · also called quantitative_claims, data_claims

Checked by The claims your site makes: years, prices, refund window · Term: self-claim-ledger

Subjective Quality Claims subjective_quality_claims

Claims that are matters of opinion or personal experience, such as friendly staff or the best pizza in town. The model treats them as weak evidence unless many independent reviews say the same.

claims.verification · also called opinion_claims, experience_claims

Complaint Pattern Claims complaint_pattern_claims

Claims about repeated problems in complaints, such as several reviews describing no-shows or surprise fees. The model treats a pattern as much more serious than one bad review.

What you can do: Fix the underlying problem, reply publicly to each complaint and note what changed.

reputation.evidence · also called recurring_complaints, issue_patterns

Concern Claims concern_claims

Claims raising concerns about a business, such as safety worries, pricing doubts or disputes. The model checks how well-supported they are before adding cautions to its answer.

What you can do: Address common concerns on the site, such as insurance, guarantees and complaint handling, and reply to concerns in reviews.

claims.verification · also called consumer_concerns, worry_claims

Service Scope Claims service_scope_claims

Claims about what work a business does and where, such as residential and commercial roofing across three counties. The model checks them against service pages and listings to decide fit.

What you can do: List each service and service area on its own page, and keep listing categories in line with them.

service.fit · also called service_range_claims, service_extent_claims

Checked by The places your site names against your Business Profile

Amenity Claims amenity_claims

Claims about extras and facilities, such as free parking, a pool at an RV park or a waiting room for pet owners. The model checks them before using them for queries that ask for those features.

What you can do: List amenities in text on the site and in listing attributes, and keep them current.

claims.verification · also called feature_claims, benefit_claims

Verification Preference verification_preference

The kind of proof the model prefers for a claim, such as an official register over the business's own word or many reviews over one testimonial. It decides how far a claim can be trusted.

What you can do: Point to the strongest proof available for each claim, such as official registers and independent reviews.

claims.verification · also called verification_methods, trust_verification

Also listed in this stage, defined under another: contradictions_found → contradiction

← Temporal Analysis  ·  Scoring →

All 16 stages · signed data