13 Things That Changed in ChatGPT in 2 Months: 40,616 Answers
Maya analyzed 40,616 ChatGPT answers across 8 sectors: sources narrowed on Sept 29–30, and answers asking questions jumped to 80–93% on Oct 9. 13 changes, with data.
Prefer Maya AI in Google
Highlight our stories in Search, AI Mode & AI Overviews.
Maya Research · Last checked: November 2, 2026
The way ChatGPT answers changed in several steps between August and October. From July 22 to October 26, 2026, we regularly asked ChatGPT 1,916 prompts across 8 sectors and analyzed 40,616 unique answers: fitness and weight loss, hair salons, beauty, restaurants, banking, card and miles programs, health, and security. We ran the source analysis on 603 prompts and 13,726 answers in 4 local sectors.
Each prompt was compared with its own history, and break dates were estimated from the data. Here are the 13 changes we saw.
In short:
- August 8–9: Background searches increased.
- September 29–30: Displayed sources narrowed overnight.
- October 9: ChatGPT began asking questions, answers got longer, and business cards appeared.
Sources
1. It started running more searches in the background
To answer a question, ChatGPT runs a series of searches in the background (fan-out). The number of these searches increased on August 8–9. In national sectors, search queries per answer rose from about 0.75 to 2.9 in banking and from 0.7 to 1.8 in card programs.
2. Then it cut the sources it shows overnight
On September 29–30, sources (citations) shown per answer in local searches fell in all four sectors within the same two days:
- Fitness: 10.0 → 2.9 (−71%)
- Hair salons: 9.5 → 2.1 (−78%)
- Restaurants: 10.6 → 3.7 (−66%)
- Beauty: 10.7 → 2.4 (−77%)
Among prompts whose source count changed, it fell in every case and rose in none.
3. The sources that disappeared were not cited in the text
Most sources were "background" sources that ChatGPT looked at but did not link in the answer text. These fell from 4.7–5.8 per answer to 0.7–0.9 (−82% to −87%). In-text links stayed at 1–2 per answer. The citations users see within answers barely changed; the narrowing happened in the background.
4. Sources concentrated on a few sites
For an equal number of answers, the number of distinct source sites fell 58–78%. In fitness, it dropped from 1,218 sites to 406. In beauty, the 10 most frequent sites went from 52% to 74% of all sources: three quarters of sources now come from ten sites.
5. It still shows the site of the business it recommends
When ChatGPT recommended a business, it cited that business's site in 100% of answers in all four sectors, both before and after the narrowing. Because total sources fell, brand sites' share of sources rose:
- Fitness: 19% → 34%
- Hair salons: 2% → 8%
- Restaurants: 7% → 18%
- Beauty: 6% → 24%
6. It no longer shows the site of a business it doesn't recommend
A business's site used to be able to appear as a background source even when the business wasn't recommended in the answer. After September 30, this visibility almost disappeared:
- Fitness: 14.4% → 0.8%
- Beauty: 11.2% → 0.0%
- Restaurants: 2.4% → 0.0%
- Hair salons: 10.6% → 9.2% (the only exception)
7. Directory, booking, and review sites pulled back
We tested, one by one, the 153 sites cited in more than 3% of answers in at least one period. 51 dropped out of the source list entirely: 23 directory and booking platforms, 15 business sites, and 10 media outlets and blogs. Some of the rest fell sharply:
- A large booking platform in beauty: 53% → 10%
- A services marketplace in beauty: 45% → 13%
- Yandex in beauty: 41% → 14%
- Tripadvisor in restaurants: 35% → 14%
- Yandex in restaurants: 21% → 4%
- Şikayetvar (a Turkish complaints site) in hair salons: 15% → 3%
- A health and directory site in fitness: 18% → 3%
The rate at which directory platforms were linked in the text also fell: from 60% to 32% in beauty and from 49% to 32% in hair salons.
Questions
8. On October 9, it began asking users questions
Nine days after sources narrowed, the format of answers changed. Share of answers asking a question (September 24 – October 8 → October 9–16):
| Sector | Before | After |
|---|---|---|
| Fitness and weight loss | 3% | 80% |
| Hair salons | 2% | 89% |
| Beauty | 4% | 86% |
| Restaurants | 1% | 80% |
| Banking | 10% | 89% |
| Card and miles programs | 7% | 90% |
| Health | 7% | 85% |
| Security | 7% | 93% |
The change wasn't gradual: in six of the eight sectors, the break fell exactly on October 9. In beauty it was October 8, and in health October 11.
9. It also asks on questions that don't seek a recommendation
On informational prompts such as "How do smart home systems work?" or "Where should I start saving for retirement?", the share of answers containing a question also rose from 2–9% to 83–91%.
10. The "Want me to compare them?" closing nearly vanished
Before, 67–88% of answers ended with an offer such as "I can compare them along with their prices if you like." After October 9, this fell to 7–28%. Questions asking the user for information took its place. Examples from real answers (translated from Turkish):
- Restaurants: "Which side of Istanbul suits you better?"
- Fitness and weight loss: "What's your budget per session?"
- Banking: "Which bank are you a customer of?"
- Card and miles programs: "Where do you travel most?"
- Health: "In which city?"
- Security: "How many cameras are you considering?"
The question doesn't replace the recommendations; it comes after them. ChatGPT recommends first, then asks.
11. Answers gained business cards and Google Maps links
Since October 9, local searches show a card and a Google Maps link for each recommended business. Answers containing a Google Maps link rose from under 1% to the following (September 30 – October 20):
- Hair salons: 16%
- Beauty: 13%
- Fitness: 11%
- Restaurants: 8%
Yandex links declined over the same period.
Answers and brands
12. Answers got 46–107% longer
Average words per answer rose in all eight sectors after October 9. The largest increase was in health: from 157 words to 325 (+107%). The smallest was in beauty: from 194 to 283 (+46%). Answers got longer for 93–100% of prompts.
13. But brands didn't increase, and fell in local searches
The extra space didn't go to brands: on October 9–20, no sector saw more brands recommended per answer. Brand counts fell in local sectors and security (compared with September 24 – October 8):
| Sector | October 9–16 | October 17–20 |
|---|---|---|
| Fitness and weight loss | −27% | −17% |
| Hair salons | −34% | −24% |
| Beauty | −23% | −14% |
| Restaurants | −16% | −14% |
| Security | −20% | −18% |
Brand counts didn't change in banking, card programs, or health. The difference between local and national sectors is significant (p < 0.001).
One caveat: if businesses whose names couldn't be read from the new-format cards are also counted, the local decline shrinks to −19% for October 9–16 and −2% for October 17–20. The drop got smaller after October 17, and it isn't yet clear whether it will last.
What does this mean for brands?
- ChatGPT now shows you through your own site. It shows far fewer sources, and the source it shows for a business it recommends is almost always that business's own site. Make your services, location, prices, and opening hours clear on your site; these are also the details ChatGPT asks users about.
- Don't rely on platform profiles alone. Directory, booking, and review platforms are still cited, but in far fewer answers.
- Keep your Google Business Profile up to date. Since October 9, business cards in answers link to Google Maps.
- Separate two break points in your measurements: September 29–30 and October 9. Otherwise a change in ChatGPT will look like a change in your brand's performance. GA4 shows a similar shift: around October 9, part of ChatGPT traffic began being recorded as Direct.
How we measured
- We used only answers from ChatGPT's user interface.
- Each prompt was compared with its own history. The 8-sector analysis used GEE clustered at the prompt level; the local analyses used paired Wilcoxon tests.
- Break dates weren't assumed: we searched the daily series for the day that best fit a single shift, and measured the uncertainty with bootstrapping.
- Holm and Benjamini-Hochberg corrections were applied for multiple testing.
- The source analysis is limited to local sectors, while the fan-out increase comes from national sector data.
For our overall measurement approach: How Maya measures
In which ChatGPT answers does your brand appear, and with which sources? You can see it with Maya: withmaya.ai
Source: Maya research — 8 sectors, 1,916 prompts, 40,616 unique ChatGPT answers; source analysis on 4 local sectors, 603 prompts, 13,726 answers (July 22 – October 26, 2026)
GEO researcher at Maya. Focused on measuring AI visibility — the metrics, benchmarks, and methods behind tracking how brands appear in AI answers.
