# Edition 1 restatement — 2026-09-03

**What this is.** A dated restatement of the three Edition 1 (July 2026) data files of the SA AI Visibility Index. The CSVs at this URL now carry a `restated` column reading `2026-09`; the files exactly as first published on 2026-07-05 are archived unchanged in `as-published-2026-07-05/`. Nothing was edited silently.

## What was wrong

The Index is scored by an LLM open-capture pass that reads each grounded answer and extracts the brands it names. That pass was given the raw answer *including the list of cited source URLs the engines append*. A brand whose name is also its domain — `property24.com`, `privateproperty.co.za`, `remax.co.za`, `rawson.co.za` — was therefore credited with a *mention* in answers where it appeared only inside a source URL and nowhere in the prose. A citation is not a mention; citations are counted separately. This overstated share-of-mention for exactly those brands.

A narrower related fault: in a handful of answers the extractor named a brand the answer text does not contain at all (a generic phrase or the prompt's context read as the brand). Those rows are not measurements either.

The equivalent fault in the deterministic cross-check was found and fixed on 2026-08-02; the fix never reached the pass that publishes. Found 2026-08-26 while locking the Edition 2 panel (`docs/aiv-index-edition2-2026-10/PANEL-LOCK-2026-08-26.md` §3).

## The fix

A deterministic post-filter at export time (`benchmarks/_pipeline/export_index.py`, shared matcher in `_text.py`): an extracted entity is published only if its name or an alias appears in the answer prose with cited URLs removed, or in the prompt the answer responds to (a comparison prompt that names the brand; the answer may use a short form outside the alias list). No model was re-run; the stored answers were re-exported. The same filter applies to every edition from here on, so the Edition 1 → Edition 2 delta is measured with one instrument.

Reproduce: `python3 benchmarks/_pipeline/export_index.py --sector <sector> --out <dir> --restated 2026-09`. Tests: `benchmarks/_pipeline/test_export_filter.py`.

## Measured effect on tracked brands

Answers per vertical: 432 (unchanged). Shortlist shares: unchanged in every vertical — every removed hit was on a non-shortlist prompt. Conviction moves where the removed hits were non-recommended mentions.

| Vertical | Brand | Named in answers | Share of mention |
|---|---|---|---|
| Property | RE/MAX of Southern Africa | 106 → 93 | 24.5% → 21.5% |
| Property | Rawson Property Group | 95 → 86 | 22.0% → 19.9% |
| Property | Property24 (platform) | 77 → 66 | 17.8% → 15.3% |
| Property | Private Property (platform) | 57 → 45 | 13.2% → 10.4% |
| Legal | Adams & Adams | 61 → 59 | 14.1% → 13.7% |
| Legal | LegalWise (platform) | 34 → 31 | 7.9% → 7.2% |
| Legal | Scorpion Legal Protection (platform) | 21 → 18 | 4.9% → 4.2% |
| Legal | Barnard Incorporated | 5 → 3 | 1.2% → 0.7% |
| Accounting | Xero (platform) | 70 → 69 | 16.2% → 16.0% |
| Accounting | PwC South Africa | 59 → 58 | 13.7% → 13.4% |
| Accounting | TaxTim (platform) | 22 → 20 | 5.1% → 4.6% |

Every other tracked brand is unchanged. Two rank changes follow: in property, Chas Everitt International (20.1%) moves above Rawson Property Group (19.9%) to fourth; in legal, STBB and Barnard Incorporated (both 3 answers) swap ninth and tenth.

Removed hits by mechanism (tracked brands): property 41 URL-only + 4 absent from the text = 45 of 1,053 (4.3%); legal 4 + 6 = 10 of 419 (2.4%); accounting 2 + 2 = 4 of 760 (0.5%). Two further legal hits (Scorpion, on the prompt that asks about Scorpion by name) fail the alias match but are kept by the prompt rule.

Untracked names: 18 of 1,412 property, 44 of 1,876 legal and 7 of 1,534 accounting untracked hits were removed on the same rule — almost all names lifted from source URLs (ooba, MyRoof, attorneys.co.za, Old Mutual). On the published "named without being tracked" lists this moves Burger Huyser Attorneys 46 → 43, Sage 32 → 30, Tax Consulting South Africa 26 → 25 and QuickBooks 21 → 20.

## Correction to the 2026-08-26 measurement

The panel-lock note estimated 47 property hits, including 6 for Lew Geffen Sotheby's International Realty. Those 6 were an artefact of the measurement, not of the data: the answers write "Sotheby’s" with a curly apostrophe and the alias list uses a straight one. The shared matcher now folds apostrophes; Lew Geffen Sotheby's is unchanged. Legal was estimated at 2 and is 10, because the 2026-08-26 pass counted only URL-only hits and did not count extractor names absent from the text.