AI assistants rarely give the same list twice, and a model update can reshuffle it. AIGROW Labs followed Gemini on 2,194 local business pages for three months: with the same model nothing was lost beyond re-ask noise, and across a model update recommendations kept 0.71 of their re-ask level, mostly lost by rarely named businesses.
Need the answers for your own brand? The AI visibility audit puts 20 to 40 of your buyers’ questions to four AI models and records who they name instead of you.
- One AI answer is a small sample. Asked the same questions again, Gemini named 56.1% to 57.1% of the businesses its two other runs had named, on 2,194 local business pages of the AIGROW Index.
- ChatGPT was less steady: a re-ask on the same pages in November named 48.8% of them, and Claude 34.1%.
- With the same model, Gemini lost nothing over 10 days beyond that re-ask noise. Across a model update it kept 0.71 of its re-ask level, and the loss came in one step.
- Rarely named businesses took the loss. Named in one of 8 earlier answers, they kept 0.51 of their re-ask level; named in 6 or more, they kept 0.94.
- You cannot stop a model update. A check repeated every month on the same questions shows when one has moved you, and the free checker gives a first reading.
What probably happened
A customer tells you ChatGPT suggested your business. A month later you ask it the same thing, and your name is not in the answer. Most owners start looking for something they did wrong.
Usually one of two things has happened. The assistant gives a different answer each time it is asked, so a single answer, yours or your customer's, is a small sample. Or the model behind the product changed, and the new one picks a different set of businesses.
AIGROW Labs measured both. 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. A page is one kind of business in one city.
Between October and November the model behind Gemini changed. The study measured how long its recommendations lasted, net of the change that simply asking again causes.
One answer is a sample
Ask an assistant the same question twice and you get two different lists. The study measured how different. On each page, two runs of Gemini's four questions picked out the businesses it named, and a third run in the same month was compared with them.
That third run named only 56.1% of those businesses in September, 56.5% in October and 57.1% in November.
ChatGPT and Claude answered on the same 2,194 pages in November, 26,328 answers each, and they were less steady:
| Assistant | Named again on a re-ask | Names per answer | Answers naming no business |
|---|---|---|---|
| Gemini | 57.1% | 3.21 | 0.0% |
| ChatGPT | 48.8% | 2.99 | 2.8% |
| Claude | 34.1% | 3.97 | 23.9% |
Claude ran at its default temperature, unlike the other two, so its lower figure mixes the assistant with the setting.
So if ChatGPT named you once and not the next time, that alone tells you very little. A business at the edge of the list can be in one answer and out of the next on the same day.
With the same model, nothing was lost
From September to October the model behind Gemini stayed the same. The two readings were a median of 10 days apart, and a later run named 56.1% of the earlier businesses, the same share a re-ask in September named.
The study calls the later rate divided by the re-ask rate the share kept: 1 means nothing changed beyond asking again. From September to October it was 1.00, and its interval contained 1 in each of the 25 kinds of business.
Over those 10 days, then, a later answer differed from an earlier one only as much as two answers on the same day did. If nothing changed behind the product, a missing answer is most likely that noise.
A model update reshuffles the list in one step
Between October and November the Gemini model that answered changed from gemini-3.1-flash-lite to gemini-3.1-flash, and the way the study reached it changed at the same time. Across that change the share kept fell to 0.71.
| Comparison | Days apart | Re-ask | Later run | Share kept |
|---|---|---|---|---|
| September to October, same model | 10 | 56.1% | 56.1% | 1.00 |
| October to November, model update | 29 | 56.5% | 40.2% | 0.71 |
| September to November | 39 | 56.1% | 40.0% | 0.71 |
September to November spans 39 days and gives the same 0.71, so the loss came in one step with the update and did not grow with time. Its interval lay wholly below 1 in each of the 25 kinds of business, from 0.68 for physiotherapists to 0.75 for plumbers.
Counted over all 12 answers on a page, 50.1% of the 42,534 businesses Gemini named on the October pages were named again on the same page in November. From September to October, with the same model, 68.3% of 42,814 were.
The new model also gave shorter lists: 3.21 names per answer against 5.03 in October. In November 78.1% of the answers named exactly three businesses, where 73.7% of the October answers had named five or more. A shorter list leaves fewer places for the businesses at its edge.
Who drops out, and who stays
The loss did not fall evenly. The study grouped businesses by how many of the 8 earlier answers on their page named them, two runs of four questions, and measured the share each group kept across the update.
| Named in the earlier answers | Share kept across the update |
|---|---|
| 1 of 8 | 0.51 |
| 2 to 3 of 8 | 0.66 |
| 4 to 5 of 8 | 0.80 |
| 6 to 8 of 8 | 0.94 |
| All businesses | 0.71 |
The head of the list held. The first three names of an answer kept 0.98 of their re-ask level, and the single business a page named most kept 0.92.
If an assistant named you in most answers, the update rarely moved you. If it named you now and then, often far down a long list, you were in the group most likely to disappear. A recommendation you heard about once, from one customer, may well have come from that group.
Want to know where your business stands? The free checker shows whether ChatGPT, Claude, Gemini and Grok name it.
What you can control
You cannot choose when a company changes the model behind its assistant, or how the new model draws up its list. You can choose how you read the answers and when you look.
Count over many answers
Ask several questions your customers would ask, in more than one assistant, and count how often you are named. One answer cannot tell a drop from chance.
Check every month on the same questions
In the study, the drop came in one step between two monthly readings, and before it nothing had moved. A monthly check on fixed questions shows that step when it comes. Keep the wording the same each time, or you will be measuring your own changes.
Look again before you act
If you are missing, check again on another day. A drop that holds on a second check and across most of your questions is a real change. One question missing once is most likely noise.
See who replaced you
The businesses named in your place tell you whether the new model favours different kinds of business or simply reshuffled the same rivals.
Keep your facts the same everywhere
The study did not test what makes a model name a business more often, so this part is general practice and not one of its findings. A model answers from what it learned from the web, and assistants with search also read pages at the moment they answer. The facts an owner controls sit on their own website, their Google Business Profile, the directories for their trade and the reviews customers leave.
Keeping them complete and consistent gives each new model one clear picture of the business.
How to check your own business
The free checker gives you a first reading. It shows whether ChatGPT, Claude, Gemini and Grok name your business and who gets named instead. Treat it as a snapshot: the study shows that one reading can miss in either direction, so repeat it before you change anything.
To follow your business over time by hand, keep a simple sheet:
- Write down the questions a customer would ask, with your town or area in them.
- Ask each one in the assistants your customers use, the same way each time, and note whether you are named and who is named instead.
- Repeat on the same questions once a month, and again when you hear that an assistant has a new model.
- Read the share of answers that name you, month by month, and leave single answers aside.
Where Watch fits
Watch is the AIGROW plan that runs those checks for you. It puts your 10 questions to ChatGPT, Claude, Gemini and Grok every month and saves each answer. You see whether you were named and in which place, who was named instead, and your share of voice over time, with a CSV export.
When your score falls by more than 2 points, or a new problem appears, it sends you an email.
A score built on ten questions and four assistants moves less on chance than a single answer does, and a monthly history shows whether a fall came in one step, as it did at the model update in the study, and whether it held.
Monitoring records the answers the assistants give. It cannot stop a model update or promise that an assistant will name you. It tells you in which month a change happened and which questions it touched, so you can act on a real change.
What the study cannot tell you
- The answers came through each model's API with a fixed prompt and web search off. A consumer app with web search, memory and the user's location can show something else.
- Between October and November the model, the way the study reached it and the passing days changed together, and the study cannot tell them apart.
- The month-to-month figures are Gemini's, the one assistant asked the same way in all three editions. ChatGPT and Claude were measured in November only.
- Business names were grouped by an automatic matcher, so a wrong merge or split changes the counts.
- It did not test what makes a business named more often.
- AIGROW sells monitoring, Watch among it. The study is a working paper, not peer reviewed, and its method and every figure with its interval are on the study page.
Check your own business in AI answers
Four assistants, your own buying prompts, five working days. If it turns out you are already named, the audit says so and costs you nothing further.
One fixed price, paid once
Frequently Asked Questions
