Index changelog

Index Changelog: Prompt & Method Changes

Logged with the edition it takes effect. Editions are immutable; corrections publish as amendments.

2026-10-31 · index: 2026-10 edition; ChatGPT answers from the app and the OpenAI API. 500 runs, one per prompt across 20 categories, Prompt Set v1.0. 248 answers were captured from the ChatGPT app between 27 and 30 October 2026 (gpt-6-thinking). The app kept signing the capture session out, so 252 were answered through the OpenAI Responses API with web search on 30 October (232 gpt-6-luna at medium reasoning effort, 10 gpt-6-luna at high effort, 10 gpt-6.1-terra). Each run records which way it was collected. API answers keep only the links shown in the answer, as app captures do. On the same prompts the API answers named 3.8 database vendors per answer against 4.5 for September's app captures, so rates in API-heavy categories read lower partly because of the instrument. Engine status now lists every model that answered.
2026-10-31 · index: Ten vendors added after the 2026-10 review. Each was named in at least two October prompts as a vendor of its category and checked against its own site; PMS modules and services-only compliance firms were not added (same bar as September). Docuverus and Two Dots (tenant screening), Parkva and Laneo (parking), Reliance Utility Services and Studebaker Submetering (utility management), Pro Specialty (renters insurance), Caliper Air (HVAC filter delivery), SuiteSpot (maintenance) and Concierge Plus (resident experience).
2026-10-31 · index: Claude added as an engine from the 2026-10 edition. 500 Claude answers, one per prompt, through the Anthropic Messages API with the web search tool (approximate US location), Claude Haiku 5.5 at medium effort, captured 31 October 2026 and stored as engine `claude`, collection `anthropic_messages_api`. Runs keep the links shown in the answer. A vendor's rate in the 2026-10 edition is the mean of its ChatGPT and Claude rates, so every October rate and rank changed from the ChatGPT-only figures first published that day. The protocol now lists five engines; the API is not the claude.ai app, and the methodology says so.
2026-10-31 · index: Thirty-one vendors added after reviewing Claude's answers. Same bar as the earlier October review (named in at least two prompts as a vendor of its category, checked against its own site; PMS products, local or regional service firms and general-purpose tools not added). LeaseGuard (renters insurance, also deposit alternatives), Deposit Choice, URBI (community engagement and resident experience), Haven AI, Tenant Turner, Realty AI, Leasey.AI, Nurture Boss, PLACA.AI, ClearNow, Paymentus, Billee, Utility Ranger, UB West, American Submetering, Move Concierge, MeterNet, Hocutt, Epic Charging, GoPower, REVS, ChargeLab, HOA Express, Effortless HOA, Filter-Smart, National Tenant Network, Checkr, Igloohome, IBM Maximo, Planon and Asset Essentials.
2026-10-31 · index: 2026-09 edition marked published. The September edition had stayed labeled "preview" after its last data change on 2026-10-03 (500 ChatGPT runs, 20 categories). It is now published with that date and frozen; its notes no longer call it a preview, and no figure changed.
2026-10-05 · index: Funnel alias trimmed. The alias "LeaseHawk ACE" was removed from Funnel because no source ties that product to the April 2025 asset purchase; "LeaseHawk" stays, supported by Funnel's release. Matching is unchanged for the 2026-09 runs, since "LeaseHawk" already covers the same mentions.
2026-10-03 · index: Run 1 of the seventeen added categories captured; 87 vendors added after review. 425 ChatGPT web-app answers captured on 2 and 3 October 2026 (temporary chats set to Unpersonalized; model recorded from the conversation record, gpt-5-6-thinking on every answer), bringing the 2026-09 edition to 500 runs across 20 categories. Unknown names from all 20 categories were reviewed against each company's own site; Ryan approved the promotions, and data/vendors.yaml grew from 154 to 241 vendors, including new vendors in the three launch categories (NetVendor, AppWork, Wayline, VendorPM, DoorlineOps; Aiphone, ProdataKey, 2N, Akuvox, Verkada, Kastle, LiftMaster myQ Community; Boom, Esusu, RentPlus, Stake, RIS, Updater, Resident-Link), which adds rows to their September tables without changing their top rankings. Property management systems, local service providers and lock hardware brands are not ranked.
2026-10-02 · index: Ten more categories added to the Index; all twenty measured. package_management, resident_experience, pest_control_programs, renters_insurance_programs, tenant_screening, ai_leasing_assistants, rent_payments, security_deposit_alternatives, hvac_filter_delivery, smart_home registered in categories.yaml with their Prompt Set v1.0 drafts approved as written by Ryan. These ten hold 5 to 9 sourced vendors each, below the ten-vendor test in spec §4.2; Ryan waived the test for them, and vendors the engine names that are not in the database go to the review queue. The cost hold on the seven categories added 2026-09-06 is lifted. Run 1 for all seventeen is captured from the ChatGPT web app.
2026-09-20 · index: The 2026-09 ChatGPT API run is retired; all three categories collected from the app. The single ChatGPT run collected on 2026-09-03 from the Responses API (prompt mvn-best-01, gpt-5.5-2026-04-23) is moved out of the edition to index-engine/archive/retired-runs/, where it is kept unedited and auditable but is not loaded by the engine. The other 24 prompts in that category were never run against the API, so leaving it in place would have put one API run inside a prompt set of 25 app captures with nothing in the export to show the set was collected two ways. Every figure published from that run — 124 consulted URLs against 6 inline citations, and the source breakdown behind them — stays true and stays published, labeled as the API capture of 2026-09-03; it cannot be refreshed, because the app does not disclose the consulted list. In its place, run 1 of all three categories is now 75 web-app captures collected on 19 and 20 September — access_control 263 citations, maintenance_vendor_networks 184, resident_benefits 215. Collection cost was zero. The importer now refuses to write into a run slot already holding a differently-collected run. The resident_benefits set was re-captured after account memory was turned off, and the three-prompt canary that had leaked the operator's own occupation and an unprompted location came back clean.
2026-09-19 · index: First edition measures three categories; ChatGPT search context set to low. The seven categories registered on 2026-09-06 stay registered, with their prompts public, but are held out of the run schedule (measure false in categories.yaml) until the first edition's results and costs are reviewed. resident_benefits, access_control and maintenance_vendor_networks run as planned. The ChatGPT adapter's search_context_size moves from medium to low, which reduces how much page content the model reads per search. The one preview run collected on 2026-09-03 used medium and is kept as recorded.
2026-09-19 · index: ChatGPT is now captured from the web app, not the API. ChatGPT answers are collected from the consumer app at chatgpt.com, one temporary chat per prompt, driven by browser capture on an account kept for the Index with memory, custom instructions and chat-history search off. The OpenAI API path is retired and the engine refuses to run through it, so the search_context_size change logged above no longer applies to anything collected from 2026-09-19. Three consequences that can move a number. A capture carries no model id, so a ChatGPT row can move because OpenAI changed the default model with nothing in the run file to show it. A capture holds only the links in the answer, so the consulted-sources list the API exposed is gone for ChatGPT and the Source Map counts fewer URLs per run for that engine. ChatGPT's three runs per prompt are captured by hand on three days at least two days apart rather than by the scheduler. Runs collected before this date came from the API and stay as recorded, labeled by their own collection field. A capture that shows the operator's own business or a location no prompt supplied is discarded and re-run; the first resident_benefits capture of this edition was discarded on those grounds.
2026-09-06 · index: Prompt Set v1.0 drafts for the seventeen unmeasured categories. 425 draft prompts (25 per category, six types) written to data/prompts/<category_id>.yaml for the 17 Property Management Software categories the Index does not yet measure. Drafts are not in categories.yaml, are not run, and are not published; each awaits Ryan's approval (spec §4.3) and the category-addition test (spec §4.2, ≥10 viable vendors). The validator checks drafts to the same rules as live sets.
2026-09-06 · index: Seven categories added to the Index. utility_management, parking_management, ev_charging, community_engagement, hoa_management, facilities_management_cre, affordable_housing_compliance registered in categories.yaml. Each has at least ten sourced vendors in the database (spec §4.2) and an approved 25-prompt set (Prompt Set v1.0; approval delegated by Ryan on 2026-09-06). The Index now measures ten categories. The other ten categories stay guides with draft prompt sets until they reach ten vendors.
2026-09-03 · index: 2026-09 preview edition. One ChatGPT run (gpt-5.5, web search) for prompt mvn-best-01. Perplexity, Gemini and Google AI Overviews not yet run. Engine list prices recorded in config/engines.yaml for cost logging.
2026-09-02 · index: Prompt Set v1.0 approved. 75 prompts across three categories (25 each, six types) approved by Ryan York. Frozen for the 2026-09 edition.
2026-09-02 · index: Categories at launch. resident_benefits, access_control, maintenance_vendor_networks. Vendor database seeded with 29 sourced records.
2026-09-02 · index: Google AI Overviews capture provider set to Apify. SE Ranking's data API has no on-demand SERP endpoint; AI Overview text and sources are captured through the Apify Google Search Results Scraper (US, desktop, non-personalized). No first-party UI automation.

Part of the AI Shortlist Index. See the methodology, the prompt library, and the site-wide changelog.