sempite/ai-visibility-index

The SEMPITE AI Visibility Index, 2026 Q3

Do AI assistants recommend the businesses Google ranks first?

Mostly not.

Canonical release: https://doi.org/10.5281/zenodo.22852300 This repository mirrors that deposit. Cite the DOI.

What is measured

For 50 US local markets (5 service categories x 10 cities) we take the three businesses Google Maps ranks highest, then ask AI assistants for a recommendation in that market and check whether each of those three is named in the answer.

One observation per (month, market, Google top-3 business, engine): 150 per engine-run.

Findings, share of Google's top 3 NOT named (95% Wilson intervals)

Month	Engine	Invisible
2026-08	ChatGPT (gpt-4o-mini, no web search)	96.0% (91.5 - 98.2)
2026-09	ChatGPT (gpt-4o-mini, no web search)	95.3% (90.7 - 97.7)
2026-09	ChatGPT (web search on)	89.3% (83.4 - 93.3)
2026-09	Google AI Mode	68.0% (60.2 - 74.9)

Every engine answered with specific named businesses in all 50 markets. The businesses they named were mostly not the ones Google ranks first.

Turning web search on closes some of the gap and nothing like all of it.

How stable is a single answer?

New in v1.1.0. Ten of the fifty markets were re-run five times each with the identical prompt on 20 September 2026, to find out what one draw is worth.

Engine	Of 30 top-3 businesses	Ever named but not every time
ChatGPT (no web search)	0 always, 7 sometimes, 23 never	100%
ChatGPT (web search on)	2 always, 4 sometimes, 24 never	66.7%
Google AI Mode	0 always, 17 sometimes, 13 never	100%

Not one business was named by Google AI Mode in all five repeats. Every business it ever named, it also failed to name at least once. One market's measured invisibility ranged from 33.3% to 100.0% across repeats of the same question on the same afternoon. Across repeats Google AI Mode averaged 76.0% invisible with a standard deviation of 22.2 points, against the 68.0% this edition reports from a single draw.

This changes how to read every rate above. The Wilson intervals describe sampling across markets. They do not include how much one market moves between identical queries, so the real uncertainty is wider than the intervals shown. Treat the rates as one sample of a moving quantity and the ordering between engines as more durable than any individual number.

The sub-study ships as data/stability_2026-09.json with all 150 stored answers, and its collector as code/avi_stability.py. It is 10 markets on one afternoon: it establishes that the quantity moves, not how much.

Contents

data/results.csv     one row per month, market, Google top-3 business, engine
data/summary.json    every published rate with its Wilson interval
data/answers.json    the Google top 3 and every AI answer in full, with prompts
code/avi_matcher.py  the matching rule, dependency-free, re-runnable
code/avi_score.py    produces summary.json and results.csv
METHODS.md           method and an eleven-point limitations section

avi_matcher.py can be re-run over answers.json to regenerate the mentioned column exactly. That is the point of shipping the answers in full.

Read these three limitations before citing

1. These are API models, not the consumer products. The ChatGPT figures come from gpt-4o-mini through an API, not from chatgpt.com, which uses different models, retrieval and system prompts. Read this as a measurement of a model's behaviour, not of a specific consumer app.

2. The intervals are optimistic. Wilson intervals are computed over 150 business-observations per engine-run as though independent. The three businesses in a market are scored against the same answer, so observations are clustered within market and a cluster-robust interval would be wider.

3. Earlier figures were wrong. The July (32%) and August (43%) ChatGPT figures published before 17 September 2026 used a matching rule that counted generic words from a Google listing title, such as the city or the trade, as a mention of the business. Those words appear in nearly every answer. The corrected August figure is 96.0% invisible, not 43%. Limitation 9 in METHODS.md documents it.

The matcher errs toward counting a mention wherever a call is ambiguous, so published invisibility rates are conservative.

Versions

Each quarterly edition is a new version of this record. The concept DOI always resolves to the newest edition.

Citation

@dataset{gerrish_avi_2026,
  author    = {Gerrish, Jacob Ezra},
  title     = {The SEMPITE AI Visibility Index: do AI assistants recommend
               the businesses Google ranks first? (2026 Q3)},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.22852300}
}

Write-up: https://sempite.com/research/ai-visibility-index/

Data CC BY 4.0, code MIT. Produced by SEMPITE.