ARO Index Research Report · Vol 2

State of AI Recommendations: Charleston, Vol 2 (July 2026)

By Therese Grittner ARO Index Published July 2026 Cold-query observed selection, 4 AI models Data current as of July 24, 2026

Held their spot

Dropped out

New entries

1 of 3 Charleston businesses selected by AI models in early July lost their spot within three weeks. Of the 89 businesses selected in the Vol 1 census, 56 held their position, 33 dropped out, and 28 new businesses entered. The overall selection rate held nearly flat: 84 of 688 (12.2%) versus 89 of 688 (12.9%) in Vol 1.

Across the expanded current scope of 1,444 audited businesses and 82 categories, 241 of 1,444 (16.7%) were ever named in a cold buyer query. 1,203 of 1,444 (83.3%) were never named by any model.

In our first Nashville census, 39 of 485 audited businesses (8.0%) were ever named. 446 of 485 (92.0%) were never named.

The central finding of this volume is not who is selected. It is that selection does not hold still.

Suggested citation

ARO Index, State of AI Recommendations: Charleston, Vol 2 (July 2026), published July 2026.

Cite this finding

Of the 89 Charleston businesses selected by AI models in the July 8, 2026 census, 56 held their position three weeks later, 33 dropped out, and 28 new businesses entered, on an identical 688-business cohort and 43 categories. ARO Index, State of AI Recommendations: Charleston, Vol 2 (July 2026), aroindex.com/research/charleston-2026-vol2. Data current July 24, 2026.

Media contact

Therese Grittner, ARO Index, therese@tagmakessc.com

The Churn Finding: Selection Is Not Sticky

Vol 1 answered a static question: which businesses do AI models actually select when a buyer asks a cold question like "best plumber in Charleston"? Vol 2 asked the question again, three weeks later, under identical conditions: the same 688 businesses, the same 43 categories, the same query bank, the same four models in the same order (ChatGPT, Claude, Gemini, Perplexity).

The selected set turned over substantially. 33 of the original 89 selected businesses (roughly 4 of 10) did not repeat. 28 businesses that were invisible in Vol 1 appeared in Vol 2. The list changed while the rate barely moved.

This matters more than any single ranking. A business that confirms it is selected today holds no guarantee it is selected next month. A business that is invisible today is not permanently locked out. The board is live.

We want to be precise about what this finding is and is not. It is evidence that cold-query AI selection is volatile run to run. It is not evidence that any individual business did something wrong, or that the Charleston market shifted. Part of the observed churn is inherent model variance, which we address directly in the methodology section below.

Trend: Same Cohort, Same Categories

To compare against Vol 1 honestly, we froze the original conditions: the identical 688-business cohort and the identical 43 categories, re-run against fresh cold queries.

Vol 1 (data current July 8): 89 of 688 selected (12.9%)
Vol 2 (data current July 24): 84 of 688 selected (12.2%)

A change of 5 businesses on a cohort of 688 is within run-to-run variance for this method. We do not claim a decline, and readers should not either. The stable rate combined with the unstable membership is the story: the size of the selected class holds steady while its composition turns over.

Current Coverage: The Expanded Picture

The audited cohort has grown since Vol 1, from 688 to 1,444 businesses, and the category bank has grown from 43 to 82. Against this full current scope, the July 2026 census ran 984 cold queries (82 categories, 4 models, 3 query variants each).

Ever named: 241 of 1,444 (16.7%)
Never named: 1,203 of 1,444 (83.3%)

These coverage figures are reported separately from the trend figures above and are not comparable to Vol 1's 12.9%. The cohort and category expansions are disclosed methodology changes, not measurement drift.

First Look: Nashville

In July 2026, ARO Index ran its first cold market census outside Charleston. Across 264 cold queries in 22 Nashville categories, 39 of 485 audited Nashville businesses (8.0%) were ever named. 446 of 485 (92.0%) were never named by ChatGPT, Claude, Gemini, or Perplexity in any cold buyer query.

Nashville's invisibility rate exceeds Charleston's by a wide margin at this stage of measurement. A full Nashville report is in preparation and will apply the same methodology and disclosure standards as this series.

Methodology and Disclosed Changes

Census method. Each census run is a single cold API call: one category query variant against one model, in a fresh context with no prior conversation. The July 2026 Charleston census comprised 984 runs (82 categories x 4 models x 3 variants). Nashville comprised 264 runs (22 categories x 4 models x 3 variants). Model responses were parsed for named businesses, and every candidate name was verified for real-world existence before counting. Verified mentions: 2,857 (Charleston), 1,082 (Nashville).
Census reads are single-shot. ARO audits use a 2-of-3 majority vote per model before a business counts as selected. Census runs do not. Each census variant is one cold read. This means some portion of run-to-run churn reflects model output variance rather than durable change in model behavior. We state this plainly because it strengthens rather than weakens the instability finding: even the models themselves do not reliably repeat their own selections under identical cold conditions.
Selection, not citation. ARO measures whether a model names a business as a recommendation in a cold buyer-intent query. Being cited, linked, or mentioned in other contexts is a different phenomenon and is not counted here.
Disclosed changes from Vol 1. The audited cohort grew from 688 to 1,444 businesses. The category bank grew from 43 to 82. All trend comparisons in this report use the frozen Vol 1 cohort and categories. All coverage figures use the current full scope. The two are never mixed in a single claim.
Locked data. Vol 1 figures remain as published: 89 of 688 (12.9%), 1,616 verified mentions, data current July 8, 2026. Vol 2 batches are census_jul2026_charleston and census_jul2026_nashville, data current July 24, 2026. Published batch data is never retroactively modified.
Limitations. Single-shot census reads cannot separate model variance from durable selection change; repeated monthly measurement will narrow this. Cohort composition reflects businesses audited by ARO Index and is not a random sample of all Charleston or Nashville businesses. Verification confirms a named business exists; it does not assess recommendation quality.

Data Availability

A summary dataset (category-level selection counts by market and model, 416 rows across Charleston and Nashville) accompanies this report as a downloadable CSV. Researchers and journalists may use it with attribution.

Download

↓ vol2-data-availability.csv · 416 rows · market, service_category, model, distinct_businesses_selected_by_model_in_category, total_distinct_businesses_ever_named_in_category, total_runs_for_category

References

Grittner, T. (2026). The State of AI Recommendations: Charleston 2026. ARO Index. https://doi.org/10.5281/zenodo.21582986. Live at aroindex.com/research/charleston-2026.

Cite this report

Grittner, T. (2026). State of AI Recommendations: Charleston, Vol 2 (July 2026). ARO Index. https://aroindex.com [DOI: pending]

Contact

Questions about methodology or data access: reports@tagmakessc.com

ARO(tm), ARO Score(tm), and ARO Index(tm) are trademarks of Therese Grittner. Data current as of July 24, 2026. ARO Index publishes live, public research on which local businesses AI models actually select. The audit does not create your score. It reveals it.