How long does an AI recommendation last?
Gemini answered the same four questions, three times each, on 2,194 local business pages of the AIGROW Index in September, October and November 2026. We measured how long its recommendations lasted, net of the change that simply asking again causes.
Published by Guglielmo Paesano, founder of AIGROW · working paper, 5 November 2026
- Kinds of business
- 25
- Cities
- 88
- Countries
- 20
- Languages
- 11
- Answers per edition
- 26,328
Working paper, not peer reviewed
How long does an AI recommendation last? AIGROW Labs compared Gemini's answers on 2,194 local business pages from September to November 2026. With the model unchanged, recommendations held over 10 days as well as on a simple re-ask. A model update cut them to 0.71 of that level in one step, mostly among rarely named businesses.
Findings
Gemini, the one assistant asked the same way in all three editions. The tables below give every figure with its 95% interval.
Asked the same questions again, Gemini named only 56.1% to 57.1% of the businesses its two other runs had named.
That is a re-ask in the same edition with the same model: 56.1% in September, 56.5% in October, 57.1% in November. On the same pages in November, ChatGPT reached 48.8% and Claude 34.1%. One answer is a noisy measurement, and much of the change people see between two checks is this noise.
With the model unchanged, nothing was lost in 10 days. Across a model update, recommendations kept 0.71 of their re-ask level.
The figure divides the share of earlier businesses a later run named by the share a re-ask named, so 1 means nothing changed beyond re-asking. September to October, same model: 1.00, and the interval contains 1 in each of the 25 kinds of business. October to November, when the model behind the product changed from gemini-3.1-flash-lite to gemini-3.1-flash together with the access route: 0.71, a November run naming 40.2% of the businesses against 56.5% for a re-ask. September to November, 39 days: 0.71 as well, so the loss came in one step with the update and did not grow with time.
The loss fell on businesses named rarely: named in one of eight answers, they kept 0.51 of their re-ask level; named in six or more, 0.94.
Across the update, the first three names of an answer kept 0.98 and the business a page named most kept 0.92. The new model also gave shorter lists: 3.21 names per answer against 5.03 in October, with 78.1% of November answers naming exactly three businesses.
What it means for a business
Check more than once before concluding anything: one answer is a noisy measurement. A recheck tells you most after a model update. A business named rarely is likely to drop out at the next update, and one named in most answers is likely to stay named.
What lasted across the model update
The share kept is a later run’s rate divided by a re-ask’s rate: 1 means nothing changed beyond re-asking.
- September to October, same model
- October to November, model update
- October 2026
- November 2026
Named again by a test run
Two runs of the earlier edition select the businesses, and one more run tests them. Re-ask: the test run comes from the same edition. Later run: it comes from the later edition. Each figure is the share of selected businesses the test run named, with its 95% interval.
| Comparison | Days apart | Re-ask | Later run | Share kept |
|---|---|---|---|---|
| September to October, same model | 10 | 56.1% (55.5% to 56.8%) | 56.1% (55.4% to 56.7%) | 1.00 (0.99 to 1.00) |
| October to November, model update | 29 | 56.5% (55.8% to 57.2%) | 40.2% (39.7% to 40.7%) | 0.71 (0.71 to 0.72) |
| September to November | 39 | 56.1% (55.5% to 56.8%) | 40.0% (39.6% to 40.5%) | 0.71 (0.71 to 0.72) |
The head of the list
Share kept, with its 95% interval. First three names counts a business only among the first three names of an answer. Most named business is the single business the earlier answers named most, on pages without a tie.
| Names | September to October, same model | October to November, model update | September to November |
|---|---|---|---|
| All names | 1.00 (0.99 to 1.00) | 0.71 (0.71 to 0.72) | 0.71 (0.71 to 0.72) |
| First three names | 0.99 (0.99 to 1.00) | 0.98 (0.97 to 0.99) | 0.98 (0.97 to 0.99) |
| Most named business | 1.00 (0.99 to 1.02) | 0.92 (0.90 to 0.94) | 0.93 (0.91 to 0.95) |
Three assistants in November
The same 2,194 pages, November only, so this is the noise of re-asking and not persistence: the share of businesses named in two runs that a third run named.
| Assistant | Re-ask | Names per answer | Answers naming no business |
|---|---|---|---|
| Gemini | 57.1% (56.4% to 57.8%) | 3.21 | 0.0% |
| ChatGPT | 48.8% (48.1% to 49.7%) | 2.99 | 2.8% |
| Claude | 34.1% (33.3% to 35.0%) | 3.97 | 23.9% |
Claude ran at its default temperature rather than 0.3, so its lower figure mixes the assistant with the setting.
By kind of business
Share kept, highest after the update first. From September to October the interval contains 1 in 25 of 25 kinds of business; from October to November it lies wholly below 1 in 25.
| Kind of business | September to October, same model | October to November, model update |
|---|---|---|
| Plumbers | 1.00 | 0.75 |
| Veterinarians | 0.98 | 0.75 |
| Electricians | 1.01 | 0.73 |
| Web design agencies | 0.99 | 0.73 |
| Cleaning services | 0.99 | 0.73 |
| Moving companies | 1.02 | 0.72 |
| Wedding photographers | 0.99 | 0.72 |
| Digital marketing agencies | 1.02 | 0.72 |
| Lawyers | 1.02 | 0.72 |
| Hotels | 1.01 | 0.72 |
| Hair salons | 1.01 | 0.71 |
| Roofers | 1.00 | 0.71 |
| Driving schools | 0.98 | 0.71 |
| Gyms | 0.99 | 0.71 |
| Insurance brokers | 0.99 | 0.71 |
| Dentists | 1.01 | 0.71 |
| Architects | 0.99 | 0.70 |
| Real estate agencies | 0.98 | 0.70 |
| Financial advisors | 0.98 | 0.70 |
| Interior designers | 1.01 | 0.70 |
| Dermatologists | 1.00 | 0.70 |
| Restaurants | 0.99 | 0.69 |
| Car repair shops | 1.01 | 0.69 |
| Accountants | 0.99 | 0.68 |
| Physiotherapists | 1.01 | 0.68 |
How the study was made
- The Index
- The AIGROW AI Visibility Index asks assistants four fixed questions about one kind of business in one city, three times each, in the local language, and publishes each month the businesses they name. A page is one kind of business in one city.
- The pages
- The 2,194 pages Gemini answered in full in all three editions: 25 kinds of business in 88 cities of 20 countries, in 11 languages, 26,328 Gemini answers per edition. Gemini answered from 22 September 2026 to 26 September 2026, on 5 October 2026 and on 3 November 2026.
- The questions
- The same system prompt in every edition, temperature 0.3, low reasoning effort and web search off, so the assistant answered from what the model already held. Each answer is a list of at most eight businesses.
- Access routes
- Only answers obtained through programmatic API access count: a model router, the vendor API or an API gateway. Answers through a vendor command-line tool on a consumer subscription, or through an API reseller, are left out, because the request that reached the model and its settings cannot be checked. That leaves Gemini as the one assistant observed in all three editions.
- The model update
- Between October and November the model reported by the route changed from gemini-3.1-flash-lite to gemini-3.1-flash, the route changed from a model router to an API gateway, and more days passed. These changed together, and the study cannot tell them apart.
- Share kept
- A business is selected when any of the eight answers of two runs of the earlier edition names it, and kept when any of the four answers of a test run names it. The test run comes from the same edition or from a later one. The share kept is the later rate divided by the same-edition rate, pooled over pages. Intervals are 95% intervals from 1,000 page bootstrap replicates, 400 per kind of business.
- What it does not show
- Offline answers of one product with a fixed prompt, not what a consumer app with web search, memory and location shows. Business names are grouped by an automatic matcher, so a wrong merge or split changes the counts.
- Competing interests
- The author founded AIGROW, which publishes the Index and sells services that aim to improve how businesses appear in AI answers. The study had no external funding.
Rankings from the AI Visibility Index
See which businesses and products ChatGPT, Claude, Gemini and Grok named for the categories and places below.
How to cite
Guglielmo Paesano, How Long Does an AI Recommendation Last? Held-Out Persistence of Local Business Recommendations across Three Months and a Model Update, AIGROW Labs working paper, 5 November 2026, aigrow.gg/research/ai-recommendation-persistence-2026.
A working paper, not yet peer reviewed. The figures on this page are the paper’s.
Computed from three editions of the AIGROW AI Visibility Index, archived on Zenodo: September 2026, DOI 10.5281/zenodo.23082267 · October 2026, DOI 10.5281/zenodo.23188442 · November 2026, DOI 10.5281/zenodo.23934866
Is your business named, and does it stay named?
The free scan asks ChatGPT, Claude, Gemini and Grok your own customers' questions and shows who they name instead of you.
AIGROW Labs · 5 November 2026