# Citation drift method note — 2026-11-18

This is a **single-engine, single re-run**. It is not a multi-engine study and it is not a time series beyond one original snapshot (Nov 3–6, 2026) and this one re-run.

## When and how

- **Engine:** the same built-in research-agent web-search answer tool used for `runs-2026-11-03.csv`, `runs-2026-11-04-dtc.csv`, `runs-2026-11-05-local-services.csv`, and `runs-2026-11-06-saas.csv`. No other engine (ChatGPT, Perplexity, Gemini, Google AI Overviews, or a raw web index) was used.
- **Queries:** the same 48 query strings, copied from those four CSVs. One run per query.
- **When:** Wednesday, 2026-11-18, about **8:32 AM to 8:35 AM PT** (America/Los_Angeles; the box clock showed PST, UTC-8). Each query's timestamp is in `drift-2026-11-18.csv` (`rerun_timestamp_pt`) and in the raw batch files under `raw-2026-11-18/`.
- **What was recorded:** the tool returned five source links and a short synthesis for every query (48 × 5 = 240 citation rows, no missing links). We stored the full URLs in `raw-2026-11-18/source-urls.csv` and the raw batch JSON. The public re-run table uses the same four columns as the original files: `query`, `cited_domain`, `domain_type`, `first_brand_recommended`. Domains are the URL host with a leading `www.` stripped; other subdomains (for example `blog.hubspot.com`, `learn.g2.com`, `uk.pcmag.com`) are kept. Repeated domains in the same query are kept.

## first_brand_recommended

Taken from the tool's synthesis text, not from a source-page title. It is the first brand the answer recommended (the lead recommendation). If the synthesis was conditional ("choose X if …, Y if …"), the first named brand is recorded, and that is noted in the raw batch JSON (`first_brand_note`). This dataset still does not store the full answer text.

## domain_type

Our classification, not a label from the engine. Each query keeps the vocabulary of its original run:

- 2026-11-03 (`b2b-software`): `vendor`, `vendor (platform)`, `independent`, `agency`, `publisher`, `marketplace listing`
- 2026-11-04 (`dtc-consumer`): `publisher`, `independent`, `retailer`, `nonprofit`, `professional` (no `brand` label)
- 2026-11-05 (`local-services`): `aggregator`, `brand`, `directory`, `publisher`, `other`, `nonprofit`
- 2026-11-06 (`saas-product`): `brand`, `aggregator`, `publisher`, `other`

If a domain already appeared in that vertical's original file, the original label was reused. New domains were labelled with the same scheme (see `build-drift-2026-11-18.py`). A different labeller could recode edge cases, especially vendor blogs that also publish competitor roundups.

## Drift comparison (scripted)

Script: `build-drift-2026-11-18.py`. Inputs: the four original CSVs and the 2026-11-18 raw batches. Output: `drift-2026-11-18.csv`.

- **First-named brand changed:** `no` if the two names are equal after lowercasing and punctuation stripping, or if one name is a more specific product string that starts with the other (for example `Silk & Snow` and `Silk & Snow Hand-Knitted Weighted Blanket`, or `La Roche-Posay Anthelios` and `La Roche-Posay Anthelios Melt-In Milk SPF 100`). Otherwise `yes`. This is a brand/product-name match, not a judgment about whether the companies are related.
- **Citation overlap:** the number of **distinct domains** that appear in both the original five citations and the re-run five, reported as `shared_domains` and `overlap_out_of_5` (N/5). Repeats do not increase the overlap. Overlap does not use URL paths, only hosts.

## Counted results (from drift-2026-11-18.csv)

- Queries compared: **48**
- First-named brand **unchanged**: **25/48**
- First-named brand **changed**: **23/48**
- Mean domain overlap: **2.10/5** (sum of shared domains = 101; 6 queries had 0 shared domains; 0 had 5/5)
- Overlap distribution: 0/5 × 6, 1/5 × 6, 2/5 × 20, 3/5 × 9, 4/5 × 7, 5/5 × 0

By original vertical:
- **B2B software (original 2026-11-03)** (12 queries): first-named brand changed in 7/12; mean overlap 1.58/5 (distribution: 0/5 × 4, 1/5 × 2, 2/5 × 2, 3/5 × 3, 4/5 × 1).
- **Consumer DTC (original 2026-11-04)** (12 queries): first-named brand changed in 3/12; mean overlap 2.75/5 (distribution: 2/5 × 6, 3/5 × 3, 4/5 × 3).
- **Local services (original 2026-11-05)** (12 queries): first-named brand changed in 7/12; mean overlap 2.17/5 (distribution: 1/5 × 3, 2/5 × 6, 3/5 × 1, 4/5 × 2).
- **SaaS products (original 2026-11-06)** (12 queries): first-named brand changed in 6/12; mean overlap 1.92/5 (distribution: 0/5 × 2, 1/5 × 1, 2/5 × 6, 3/5 × 2, 4/5 × 1).

Queries whose first-named brand changed:
- best email marketing platform for Shopify stores 2026: Omnisend → Klaviyo (overlap 2/5)
- best customer support help desk software for small ecommerce 2026: Gorgias → Shopify Inbox (overlap 3/5)
- best SOC 2 compliance automation tool for startups 2026: Sprinto → Secureframe (overlap 1/5)
- best returns management software for Shopify 2026: Loop → Redo (overlap 0/5)
- best product analytics tool for B2B SaaS 2026: Pendo → Amplitude (overlap 2/5)
- best loyalty and rewards app for Shopify 2026: BON Loyalty → Smile (overlap 0/5)
- best AI meeting note taker for teams 2026: Fireflies → Fellow (overlap 0/5)
- best mattress for back pain 2026: Helix Midnight Luxe → Nectar Premier Hybrid (overlap 2/5)
- best protein powder for women 2026: Transparent Labs → Promix Whey Protein Isolate (overlap 2/5)
- best office chair for back pain 2026: Steelcase Leap V2 → Herman Miller Sayl (overlap 2/5)
- best HVAC company Las Vegas 2026: Desert West → The Cooling Company (overlap 2/5)
- best plumber Austin 2026: Blue Dragon Plumbing → Beyond Wow Plumbing & Drains (overlap 1/5)
- best wedding photographer Miami 2026: Bells & Whistles Photography → Ramses Garcia (overlap 2/5)
- best personal injury lawyer Phoenix 2026: Hirsch & Lyon → Goldberg & Osborne (overlap 4/5)
- best roofing company Dallas 2026: Arrington Roofing → New View Roofing (overlap 1/5)
- best veterinary clinic Seattle 2026: North Seattle Veterinary Clinic → Elliott Bay Animal Hospital (overlap 2/5)
- best IT support for dental offices 2026: Darkhorse Tech → Pact-One (overlap 3/5)
- best CRM for SMBs 2026: Bigin by Zoho CRM → Pipedrive (overlap 3/5)
- best project management software for agencies 2026: Productive → Teamwork.com (overlap 0/5)
- best email marketing software for ecommerce 2026: Klaviyo → Omnisend (overlap 4/5)
- best password manager for teams 2026: 1Password Business → Bitwarden (overlap 2/5)
- best inventory software for Shopify brands 2026: Fabrikatör → Sumtracker (overlap 2/5)
- best SEO tools for agencies 2026: Semrush → Ahrefs (overlap 0/5)

## Files and hashes

Computed after the CSVs were written (SHA-256 of the file bytes):

- `runs-2026-11-18-rerun.csv` — sha256 `a4527d7ec1855294b924942c744d9ed1920f346a383b0fb407c30a1b6c4f1fcd` (9084 bytes)
- `drift-2026-11-18.csv` — sha256 `d41eee9600631c2c1ddd6782f682e3ef261adc28c131690e1e27b60b1dc9c829` (15479 bytes)
- Raw batches and full source URLs: `raw-2026-11-18/`

The four original run files were not edited. This re-run is one engine and one pass on 2026-11-18. It is not enough to claim a general ranking, a multi-engine comparison, or a stable trend.
