Glossary · AI classifier atlas
The labels an AI model uses to pick a business, defined, sorted and measured.
When gpt-oss is asked for “the best roofer in Denver” and told to write out how it is deciding, it names the things it weighs: hidden_intent_possible, trust_risk, no_verifiable_local_address. 659 of those names are defined here, in the order the model uses them, with what a business can do about each.
The list started as 809 names SeerSEO collected from its scans. CrawlCheck rewrote every definition, reviewed every row’s namespace, merged 125 near-duplicates, and set apart 25 rows that are not labels (12 business names that leaked in as keys, 13 container fields). Then it started measuring: which labels the model actually uses this week, and which facts about a business change its ranking.
What these labels are, and are not
They are field names gpt-oss, OpenAI’s open-weights reasoning model, writes when it is asked to describe its evaluation before answering. They are not documented OpenAI internals, nothing here shows ChatGPT uses them, and a model’s description of its reasoning is not the computation itself. They are still the most direct view there is of what a model says it weighs, and the census below counts them instead of assuming them. First collected by SeerSEO.
16 stages
In the order the model works
- 01Inputs & Scope33 labelsWhat the model takes as the brief before it looks at anyone: what is in and out of the request, and the situation it describes.
- 02Sourcing2 labelsWhere candidates come from: the directories and aggregators the model treats as places to find businesses.
- 03Query Understanding177 labelsWhat the person wants, read from their words: the job, the urgency, the budget, the trade-offs they are likely to make.
- 04Query Complexity4 labelsHow hard the request is to answer well, and what it leaves unknown.
- 05Reasoning Classifiers95 labelsThe risks and trade-offs the model reasons about while it weighs the request: safety, money, time, trust.
- 06Candidate Screening94 labelsThe tests a business must pass to stay in the running at all.
- 07Signal Extraction35 labelsFacts read off each candidate: hours, years in business, prices, service lists.
- 08Digital Presence Signals74 labelsWhat the web says about the business: its own site, its listings, its profiles and whether they agree.
- 09Competitive Context37 labelsHow each candidate stands against the others and the local market.
- 10Temporal Analysis8 labelsHow current the evidence is, and whether the timing of the request matters.
- 11Groundedness Evaluation14 labelsWhether what the model is about to say is backed by a source it read.
- 12Scoring46 labelsHow candidates are weighted, scored and ranked.
- 13Shortlisting18 labelsThe rules for building the list the person will see: how many, and who is left out.
- 14Answer Justification8 labelsThe reasons the answer gives for each recommendation.
- 15Answer Generation134 labelsHow the answer is written: its format, tone, sections and caveats.
- 16Handoff5 labelsWhat the person is told to do next, and the limits the answer states.
Measured weekly
Which labels the model used this week
97 questions asked this week (2026-W42); the model wrote a readable evaluation in 91 of them, using 2518 fields. 107 of the 659 atlas labels have appeared so far; 200 field names appeared that no list had.
| Label | Stage | Share of runs | Example the model wrote |
|---|---|---|---|
review_signal | Digital Presence Signals | 50.5% (49) | “0.05” |
geography | Query Understanding | 49.5% (48) | “Austin, TX and surrounding metro area” |
reviews_and_ratings | Candidate Screening | 47.4% (46) | “0.2” |
disclaimer | Handoff | 33% (32) | “Encourage the user to interview agents to find the best personal fit.” |
service | Inputs & Scope | 32% (31) | “moving company” |
service_scope_signal | Digital Presence Signals | 26.8% (26) | “10” |
availability_claims | Groundedness Evaluation | 24.7% (24) | “flexible scheduling” |
long_operating_history | Signal Extraction | 23.7% (23) | “0.1” |
structured | Answer Generation | 23.7% (23) | “Provide a brief intro, list the top 3 recommended locksmiths with key facts (rating, reviews, license, contact” |
license_and_insurance_verification | Candidate Screening | 21.6% (21) | “true” |
high_weight_sources | Reasoning Classifiers | 21.6% (21) | “Google Business reviews, Yelp ratings, Angi (formerly Angie's List), Better Business Bureau (BBB) accreditatio” |
pricing_transparency | Digital Presence Signals | 18.6% (18) | “true” |
official_website | Digital Presence Signals | 17.5% (17) | “...” |
licensing_verification | Candidate Screening | 15.5% (15) | “Verified Texas real estate license” |
trust_and_credentials | Scoring | 15.5% (15) | “Licensed, insured, and member of industry associations (e.g., National Garage Door Association).” |
ranking_focus | Query Understanding | 15.5% (15) | “Why these were chosen (high ratings, insurance, services).” |
phone | Digital Presence Signals | 14.4% (14) | “...” |
location_distance | Query Understanding | 12.4% (12) | “Boise, ID” |
notes | Candidate Screening | 10.3% (10) | “24/7 emergency service, quick response time.” |
address | Digital Presence Signals | 10.3% (10) | “Street, city, zip” |
Field names no list had
intro request_type shortlist_process average_rating name shortlisting_process bbb_rating request_understanding request_parsing candidate_sources signal_sources scoring_method list additional_tips rating_score weight step_1 step_2 step_3 step_4 company_name data_sources_considered tips step_5
Fact sensitivity
Which facts about a business move its ranking
Five fictional businesses, identical in shape. One fact about the third is removed, and the same model ranks all five again. Repeated across trades and cities, the rank it loses is what that fact is worth to the model. Nothing about a real business is used.
| Fact removed from one business | Pairs ranked | Dropped in the ranking | Mean places lost | Labels that appear without it |
|---|---|---|---|---|
| Higher rating Rated 4.8 of 5 → Rated 4.1 of 5 | 10 | 100% (72.2–100%) | 3.8 | candidates |
| Serves the asker's city Serves {city} and nearby → Serves a town 40 miles from {city} | 10 | 80% (49–94.3%) | 1.9 | shortlist shortlist_process shortlisting |
| Many reviews 212 Google reviews → 9 Google reviews | 9 | 66.7% (35.4–87.9%) | 1.56 | candidates_read final_ranking |
| Own website Website with service and pricing pages → No website (a social media page only) | 10 | 70% (39.7–89.2%) | 1 | candidates shortlist candidate_pool |
| Insurance stated Insured and bonded → No insurance information | 9 | 44.4% (18.9–73.3%) | 0.78 | candidate_pool shortlisting candidates shortlist_logic |
| Licence number shown State licence #48213, active → No licence information | 10 | 40% (16.8–68.7%) | 0.5 | candidates shortlisting_logic shortlist_logic candidate_pool |
| Hours stated Hours: Mon-Sat 7am-6pm → Hours: not listed | 11 | 36.4% (15.2–64.6%) | 0.45 | candidates |
| Emergency service 24/7 emergency service → Weekday appointments only | 10 | 20% (5.7–51%) | 0.2 | summit_personal_injury_lawyer candidates emergency_answer appointments candidates_read |
| Years in business In business since 2009 → Founded in 2025 | 9 | 11.1% (2–43.5%) | 0.11 | candidate_pool |
| Phone listed Phone (555) 010-4477 → No phone listed (contact form only) | 10 | 10% (1.8–40.4%) | 0.1 | reviews_and_ratings candidates phone copperline_fence_company ranking_focus |
| Prices published Prices published, free written estimates → Call for pricing | 10 | 10% (1.8–40.4%) | 0.1 | shortlist reviews_and_ratings granite_peak_fence_company long_operating_history pricing_transparency |
| Fast reply Replies to inquiries within 1 hour → Replies to inquiries within 3-5 days | 10 | 0% (0–27.8%) | -0.3 | candidate_pool shortlist candidates |
23 namespaces
What each label is about
A second axis beside the stage: a label in Scoring can be about reputation, credentials or price. Every row was reviewed; where the review moved a label out of its provisional namespace, the label’s entry says so.
query.intent 105offer.fit 98service.fit 85evidence.assessment 70answer.presentation 51reputation.evidence 47claims.verification 47evaluation.scoring 35query.context 34comparison.context 28query.scope 25credentials.status 25presence.consistency 21query.geography 17candidate.eligibility 16evidence.freshness 14selection.policy 14action.handoff 10claims.extraction 9entity.identity 9query.uncertainty 8evidence.provenance 8answer.justification 8
From label to fix
The CrawlCheck check behind each label
A label is only worth knowing if something on your site or your listings feeds it. These are the checks that read those facts.
Set apart
Rows that are not labels
Kept so the count can be checked, never counted as labels: additional_notes rule rule_1 rule_2 rule_3 black_roofing_and_waterproofing country_side_campgrounds_llc douglass_colony_group edgemont_rv_park elite_real_estate_signage empire_roofing_and_sheet_metal_inc flynn_bec_lp_denver mountain_view_rv_retreat notes rationale realtysign_pros signage_direct sunset_valley_rv_resort tecta_america_colorado_llc classifier_output model_takeaway additional_observations implication observation note.
Use the data
The whole atlas is signed at /api/v1/ai-classifiers (one label: ?key=trust_risk), the week’s census at /api/v1/ai-classifiers/census, the fact tests at /api/v1/ai-classifiers/sensitivity. For a business on a licence, /api/v1/business/ai-screening shows how the model screened it for its own trade and city. Verify any of them with the public verifier.