AI Names About as Many Brands Every Week. Are They the Same Brands?

Key findings

1. −1.8% vs 51.3%: on ChatGPT, between adjacent weeks, the total number of brands in answers to the same questions fell by only 1.8%, but 51.3% of the brands that appeared the previous week did not appear again the next week.
2. 3.2 vs 40.3: for Gemini answers in adjacent weeks, the overlap score for cited web pages was only 3.2 (out of 100), while the brand list scored 40.3. The web pages change almost completely every week; the brands change much more slowly.
3. 24.0 to 36.9: across the same 204 questions recorded in all 17 weeks, ChatGPT's weekly brand-list overlap score moved between 24.0 and 36.9.
4. 24.2 vs 38.9: ChatGPT's brand-list overlap score was 24.2 for decision questions and 38.9 for consideration questions. On ChatGPT, the recommendation lists for decision questions change fastest.

Many teams track AI visibility by taking a screenshot every week, or by counting how many brands an answer lists. Last week it listed 5, this week it still lists 5, so everything looks fine. But a list of the same length doesn't mean it's the same brands, let alone that you're still on it.

The Dageno Team took 17 full weeks of data, from June 1 to September 27, 2026, from the Dageno AI answer database, covering ChatGPT, Google AI Mode, Google AI Overviews, Gemini and Copilot. For each question, country and platform, we took one representative answer per week and compared it with the next week's answer: which brands it named, and which web pages, websites and source types it cited. The ChatGPT portion covers 502 question combinations and 47,540 answers.

01 | Did the number of brands AI recommends each week change?

Hardly at all.

Adding up all adjacent weeks that could be paired, the total number of brands in answers (counted separately per question):

• ChatGPT: 27,604 in the previous week and 27,100 in the next, down 1.8%.
• Google AI Mode: 22,890 and 23,074, up 0.8%.
• Google AI Overviews: 15,092 and 15,034, down 0.4%.
• Gemini: 20,601 and 20,415, down 0.9%.
• Copilot: 12,731 and 12,715, down 0.1%.

The change was under 2% on all 5 platforms. If you only look at "how many brands each answer lists on average," AI recommendations look very stable over these 17 weeks.

02 | With the count unchanged, how many brands changed?

About half.

Comparing adjacent weeks: how many of the previous week's brands did not appear again for the same question the next week, and how many of the next week's brands were new:

Table: "gone" uses the previous week's brand count as the denominator and "new" uses the next week's, so the two columns can't be added.

For example: last week, the answer to a question listed four brands, A, B, C and D. This week it still lists four, but they are now A, C, E and F. The count didn't change, but B and D are gone and E and F are new. All 5 platforms look roughly like this: about half the brands are replaced each week, and about as many leave as join, so the total barely moves.

This kind of replacement isn't necessarily bad. When users ask under different conditions or product information gets updated, it makes sense for AI to give a different set of answers. But for a brand, it means: being on the list last week doesn't mean you're on it this week, and a stable list length doesn't mean your own position is stable.

03 | Web pages, websites, source types, brands: which layer changes fastest?

Web pages change fastest; brands and source types change slowly.

Here we use an "overlap score" to measure how similar two weeks' lists are: the number of items that appear in both weeks, divided by the number of distinct items across both weeks, times 100. For example, if last week was A, B, C and this week is B, C, D, 2 items are shared out of 4 distinct ones, so the overlap score is 50. The lower the score, the more changed. The overlap score is not a brand appearance rate; a score of 50 does not mean a brand has 50% exposure.

Average overlap score between adjacent weeks (web pages, websites and source types exclude identified brand websites):

• Web pages change fastest. Web page overlap scores ranged from 3.2 to 25.3 across the 5 platforms. Gemini scored only 3.2, meaning it cited almost entirely different pages every week.
• Source types are the most stable (except on Gemini). This layer is very coarse: two different media pages count as the same type, so overlap comes easily. It doesn't show that AI keeps using the same kind of content.
• Brand lists sit in the middle, between 32.9 and 40.3 on all 5 platforms. Gemini's web pages change almost completely, yet its brand list is the most stable of the 5 platforms.

This shows that pages coming and going and brands coming and going are two different things. If the review citing you is replaced by another one, your brand may still be there; if your brand is still there, the model, features or conditions AI describes may still be wrong. Looking only at cited links, or only at brand lists, misses the other half.

04 | With the same fixed questions, how much does each week change?

Quite a lot, and the conclusion depends on which set of questions you look at.

We also picked out the question combinations recorded in all 17 consecutive weeks (204 for ChatGPT), which gives 16 adjacent-week comparisons. This set of questions is completely fixed, closer to the practice of "monitoring the same set of questions over time."

• ChatGPT: the weekly brand-list overlap score ranged from 24.0 to 36.9.
• Google AI Overviews: from 29.0 to 50.3, the widest swing. It was 50.3 in the week of August 17 and 29.7 in the week of September 21.
• Google AI Mode: 22.9 to 42.2. Gemini: 32.8 to 45.4. Copilot: 28.5 to 44.0.

These are the ranges seen over these 17 weeks. They can't be used directly as a normal range or alert threshold for every brand. A low overlap score only means this week's list differs a lot from last week's; it doesn't show how much exposure any brand lost. A high score isn't necessarily good either: if you're missing in both weeks, the two lists of competitors can still look very alike.

Which questions you pick also affects the result. ChatGPT's brand overlap score was 35.5 across all pairable questions and 30.2 for this fixed set; on the other 4 platforms, the two methods differed by less than 1 point. So a single overall score reflects both changes in the answers and "which questions were counted." After adding a new batch of questions, report them separately. Don't merge them into the original overall score and then claim the original scenarios improved.

05 | For which questions do brand lists change fastest?

It differs by platform. On ChatGPT, decision questions change fastest.

Brand-list overlap scores by the buying stage of the question:

• ChatGPT: awareness 35.0, consideration 38.9, decision 24.2. Decision questions change fastest.
• Google AI Mode: awareness 28.8, consideration 33.4, decision 34.5. The reverse: awareness questions change fastest.
• Google AI Overviews: 37.0, 39.3, 35.4. Gemini: 37.5, 43.0, 37.3. Copilot: 30.7, 37.9, 34.4.

By search intent, transactional questions had the lowest overlap score on 4 platforms: ChatGPT 26.7, Google AI Overviews 35.5, Gemini 32.5, Copilot 29.1. On Google AI Mode, informational questions were lowest (31.9).

So "the closer to purchase, the less stable" only holds on some platforms. On ChatGPT, being recommended once for a purchase question can't be treated as holding the spot long term. What's really worth watching are the specific questions that have commercial value, match your product and keep changing.

06 | Is it the same across industries?

No, but the industry differences are mixed with regional and sample differences.

ChatGPT brand-list overlap scores: charging and power banks 44.4, omnichannel customer messaging 40.2, AI gateways 39.1, web scraping APIs 31.1. The numbers of comparisons were 799, 1,197, 981 and 1,534.

The conditions for these 4 industry question sets aren't identical: web scraping APIs were observed from June, the other 3 from July; charging and power banks, AI gateways and web scraping APIs are U.S. questions, while omnichannel customer messaging covers Indonesia, Singapore and Thailand. So these differences can't all be attributed to industry. The takeaway: monitoring standards for one client can't be applied directly to another.

07 | If a brand is gone the next day, has it disappeared?

No. Missing for a day doesn't mean dropping out.

We also tracked, for each brand, the first day it was absent after it first appeared in the answer to a given question. Only 13.3% (Copilot) to 25.0% (Google AI Mode) were still there the next day; for ChatGPT, it was 14.0%. In other words, most brands were absent at least once by the next day.

But this is only the "first absence." If a brand appears on Monday, is absent on Tuesday and returns on Wednesday, its first absence falls on the next day, but you can't say it only existed for one day. So don't rewrite pages every day just because a brand "was gone the next day."

Other research has seen the same instability. A study published by SparkToro and Gumshoe in January 2026 (AIs are highly inconsistent when recommending brands or products) had about 600 volunteers ask the same set of questions 60 to 100 times each on ChatGPT, Claude and Google's AI. It found that the chance of any two runs returning exactly the same brand list was under 1 in 100. They recommend looking at "how often a brand appears across many runs" rather than its rank in a single answer. They asked repeatedly within a short time, while we compare week over week, so the two sets of numbers can't be compared directly.

08 | What should marketing teams do next?

1. Don't just look at list length and totals. Every week, look at the actual brand list: are you still on it, who came in, who dropped out.
2. Look at frequency, not single answers. How many weeks a brand appears and how many it's absent for the same question is more convincing than any screenshot.
3. Investigate pages and brands separately. If the cited page changed but your brand is still there, first see which content replaced it. If your brand dropped out, see who replaced you and what reasons the answer gave. If your brand is still there but the facts are wrong, correct your product information first.
4. Fix a set of core questions and keep a separate set of opportunity questions. Use core questions for week-over-week comparison and opportunity questions to find new scenarios. Report new questions separately; don't mix them into the original overall score.
5. Prioritize questions with commercial value that keep changing. On ChatGPT, decision and transactional questions change fastest; being recommended once doesn't mean long-term stability.
6. Keep a baseline when judging whether content works. Before and after changing content, record results with the same questions, platforms and time window, and note any product, pricing or PR changes in the same period. Don't mistake one swing for an effect.

Methodology

The data in this article comes from the Dageno AI answer database and covers 17 full weeks from June 1 to September 27, 2026, across ChatGPT, Google AI Mode, Google AI Overviews, Gemini and Copilot. For each question, country and platform, we took one representative answer per week and paired it only with an adjacent full week that has a record. The number of question combinations differs by platform: 502 for ChatGPT (47,540 answers; U.S., Spain, Indonesia, Singapore, Thailand), 352 for Google AI Mode, 330 for Google AI Overviews, 285 for Gemini and 178 for Copilot. Fixed question combinations recorded in all 17 consecutive weeks were calculated separately.

The overlap score is the average Jaccard similarity across comparisons, times 100. The "gone" and "new" brand shares are calculated after summing brand counts across all comparisons; the same brand is counted repeatedly across questions and weeks, so this is a different metric from the average overlap score. Web pages, websites and source types exclude identified brand websites; brands are automatically detected names and are not the same as recommendations. "First absence" only records the first time a brand didn't appear and does not mean it dropped out permanently. Question makeup differs by platform, industry and region, and we do not attribute these differences to model updates or content changes.

References

1. SparkToro, "NEW Research: AIs are highly inconsistent when recommending brands or products; marketers should take care when tracking AI visibility", January 28, 2026.