Home › Research › How AI decides who to recommend
SEMPITE Research · AI Visibility Index · September 2026 edition · By Jacob Gerrish
ChatGPT Skips 9 in 10 of Google’s Top-Ranked Local Businesses. Google’s Own AI Skips 2 in 3.
We took the top 3 Google Maps businesses in 50 local markets (5 industries × 10 major U.S. cities) and asked AI assistants who to hire in each. ChatGPT answered with named businesses in all 50 markets. Almost none of them were the businesses Google ranks highest.
Correction, 17 September 2026
The July (32%) and August (43%) ChatGPT figures we published earlier were wrong. The rule that decided whether an AI answer mentioned a business counted generic words from the Google listing (the city name, “real estate”, “personal injury”, “plumbing”) as a mention. Those words appear in almost every answer, so the rule reported businesses as visible when the AI never named them.
We rebuilt the matcher, checked it by hand against every answer, and re-scored everything we still had. August 2026, re-scored: ChatGPT never named 96% of Google’s top-3 businesses, not 43%. July’s raw answers no longer exist, so July cannot be re-scored and is withdrawn. The original August page is archived here with this note attached. Every figure below uses the corrected method, and the code is published so anyone can check it.
How stable is a single answer? Update, 20 September 2026
Every figure on this page comes from one answer per market. We re-ran 10 of the 50 markets five times each with the identical prompt to find out what a single answer is worth. Not one of the 30 businesses Google ranks in its top three 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 invisibility averaged 76.0% with a standard deviation of 22.2 points, against the 68.0% reported below from a single draw. The confidence intervals on this page describe sampling across markets and do not include how much one market moves between identical queries, so the real uncertainty is wider than they show. Read these rates as one sample of a moving quantity, and trust the ordering between engines more than any individual number. The sub-study, its code and all 150 stored answers ship with the dataset.
Three engines, three different answers
September 2026, same 50 markets, same questions:
| Engine | Top-3 businesses not named | #1 business named | Markets with none of the top 3 |
|---|---|---|---|
| ChatGPT (no web search) | 95% (143/150; 95% CI 90.7–97.7%) | 4% (2/50) | 43/50 |
| ChatGPT with web search | 89% (134/150; 95% CI 83.4–93.3%) | 20% (10/50) | 36/50 |
| Google AI Mode | 68% (102/150; 95% CI 60.2–74.9%) | 42% (21/50) | 17/50 |
Each engine answered all 50 markets. CI = 95% Wilson interval. An engine that returns no answer for a market is left out of that market, never counted as a miss.
By industry
Share of Google’s top-3 businesses each engine actually named (30 businesses per industry):
| Industry | ChatGPT | ChatGPT + web search | Google AI Mode |
|---|---|---|---|
| Plumbers | 7% (2/30) | 40% (12/30) | 53% (16/30) |
| Dentists | 7% (2/30) | 3% (1/30) | 33% (10/30) |
| Injury lawyers | 7% (2/30) | 0% (0/30) | 13% (4/30) |
| Real estate agents | 3% (1/30) | 3% (1/30) | 10% (3/30) |
| HVAC contractors | 0% (0/30) | 7% (2/30) | 50% (15/30) |
Plumbers are the one industry where web search pulls ChatGPT toward Google’s map rankings. Injury lawyers and real estate agents are nearly invisible to every engine, Google’s own included.
What this means
AI assistants are already answering “who should I hire?” with specific names, and they are mostly not the names Google ranks first. A business that has spent years earning a top-3 spot on Google Maps should not assume that position carries into ChatGPT. For nine in ten of them, it does not.
Turning on web search does not close the gap on its own. With search on, ChatGPT recommends real, currently operating local businesses, complete with star ratings and street addresses, but for dentists, injury lawyers, real estate agents and HVAC contractors they are rarely the ones Google puts in its top 3. The two systems are reading different signals.
Google’s AI Mode sits in between. It names Google’s #1 map business in 21 of 50 markets, far more often than ChatGPT, yet still leaves two in three top-3 businesses out. Treating “AI search” as one channel averages engines that disagree about who deserves the recommendation. If you have invested in ranking on Google, checking what each assistant says about you is the first step; the playbook is in our AI Search Visibility guide.
Month over month
| Run | ChatGPT: top 3 not named | ChatGPT + web search | Google AI Mode |
|---|---|---|---|
| August 2026 (re-scored) | 96% (144/150) | not measured | not comparable1 |
| September 2026 | 95% (143/150) | 89% (134/150) | 68% (102/150) |
1 August stored only the first 1,500 characters of each answer. Every AI Mode answer was longer than that, so August AI Mode cannot be scored fairly and is not published. August’s ChatGPT answers were all shorter and are complete. July 2026 is withdrawn: its raw answers were overwritten by a software fault on 1 August and cannot be re-scored. ChatGPT without search barely moved between August and September, which is what you would expect from a model answering from memory. Two data points are still not a trend.
Methodology
- Markets: plumber, dentist, personal injury lawyer, real estate agent and HVAC contractor, in New York, Los Angeles, Chicago, Houston, Phoenix, Philadelphia, San Antonio, San Diego, Dallas and Miami (50 markets, 150 businesses per engine).
- Baseline: the top 3 Google Maps results for “[industry] in [city]” (DataForSEO, United States).
- ChatGPT: gpt-4o-mini, asked “I need a [industry] in [city]. Who are the best [industry]s there? Please recommend specific businesses by name.” Once with web search off (answering from the model’s own knowledge) and once with it on. The web-search engine was added in September 2026.
- Google AI Mode: the query “best [industry] in [city]”.
- Matching: a business counts as named when the answer contains the distinctive part of its name (not the city, the trade or words like “Inc”), or its own website. Keyword stuffing in Google listing titles is ignored and street addresses are removed from answers first, so a firm on “Bellaire Blvd” does not match “Bellaire Air Conditioning”. Close calls are counted as mentions, so the invisibility rates here can only be understated, never inflated.
- Validation: every September match was checked by hand against the answer text (70 of 71 confirmed, 1 ambiguous and counted), along with 24 near-misses that a looser rule would have accepted (none was a real mention). All 150 August ChatGPT checks were reviewed against ChatGPT’s recommendation lists.
- Collection: September data was collected on 17 September 2026 (US Eastern time). One market (San Diego dentists) timed out and was re-collected the same evening. From October, runs happen on the 1st of each month.
- Limits: one run per month, and language models vary from run to run. Answers come from the ChatGPT and Google AI Mode endpoints via DataForSEO and can differ from what a signed-in user sees in the apps. Google’s map results come from a U.S.-level search naming the city, not a searcher standing in it. Claude is not measured: we have not funded API access, and we report it as absent rather than as zero.
- Cadence: monthly collection, quarterly editions. The next edition pools October, November and December 2026 and will be published in January 2027.
Data, code and citation
Everything behind this page is free to reuse under CC BY 4.0:
Cite as:
SEMPITE (2026). AI Visibility Index, September 2026 edition. https://sempite.com/research/ai-visibility-index/
Past editions and corrections
- 17 September 2026: September 2026 edition published with the corrected matcher; July and August figures corrected (see the note above).
- August 2026 edition (archived as published, superseded: its 43% figure was overstated).
- July 2026 edition: withdrawn. Its 32% figure was overstated and the raw data cannot be re-scored.
Cite this study
The full dataset, every line of collection and analysis code, and a methods document with its limitations are deposited on Zenodo under a permanent DOI. Anyone can download it, check the figures on this page against the raw records, and rerun the analysis.
Gerrish, J. E. (2026). The SEMPITE AI Visibility Index: do AI assistants recommend the businesses Google ranks first? (2026 Q3) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.22852300
Data CC BY 4.0, code MIT. Deposit maintained by SEMPITE.
Frequently asked questions
How many of Google’s top-ranked local businesses does ChatGPT recommend?
Very few. In the September 2026 run, across 50 local markets, ChatGPT never named 95% of Google’s top-3 map businesses (143 of 150) and named Google’s #1 business in 2 of 50 markets. With web search turned on it still missed 89% (134 of 150).
Does Google’s own AI recommend the businesses Google ranks?
Only partly. Google AI Mode left 68% of Google’s top-3 map businesses unnamed (102 of 150) and named the #1 map business in 21 of 50 markets. It agrees with Google Maps far more than ChatGPT does, but it is not a restatement of the map pack.
Why are these numbers so different from the August edition?
Because the August figures were wrong. The rule we used to decide whether an answer mentioned a business counted generic words from the listing (the city, “real estate”, “plumbing”) as a mention, which overstated visibility. Re-scored with the corrected rule, August’s ChatGPT figure is 96%, not 43%. The correction note on this page has the details.
Does turning on web search help a business show up in ChatGPT?
For plumbers, yes: 40% of top-3 plumbers were named with web search on, against 7% without. For the other four industries it barely moved. With search on, ChatGPT recommends real local businesses with ratings and addresses, but mostly not the ones Google ranks in its top 3.
How often is the index updated?
Data is collected every month and published as a quarterly edition, so each edition rests on three monthly runs instead of one. The next edition covers October to December 2026 and is due in January 2027.
Can I use this data?
Yes. The data, every prompt and answer, and the matching code are free to download and reuse under CC BY 4.0. Cite as: SEMPITE AI Visibility Index, September 2026, sempite.com/research/ai-visibility-index/.
Is your business in the answers?
Run the free 60-second check: the same engines, pointed at you.