Quarterly report
State of AI Recommendations, Q3 2026
Who the AI models recommended, July to September 2026, read from the frozen monthly editions of the Index. Every number below comes from those records, and every one names the month it describes. It’s free to read, with nothing to sign up for. Rankings are never for sale.
The quarter in 8 findings
- 1
85 of 183 categories measured in every edition kept the same No. 1 from July to September 2026 (46%). ChatGPT and Dyson V15 Detect each held No. 1 throughout in 2 categories.July to September 2026
- 2
41% of categories crowned a different No. 1 in September 2026 than in August (78 of 191). Across the whole quarter, July to September, it was 46%.August to September 2026
- 3
The biggest climb of the quarter: Lippincott rose 16 places in branding agencies, from No. 20 in July to No. 4 in September 2026.July to September 2026
- 4
Every AI model with answers put the overall No. 1 first in 14% of categories in September 2026, against 9% in August. In 66% of categories at least half of them did.August to September 2026
- 5
82% of country boards named a different No. 1 from the global board in September 2026, highest in Canada and the United Kingdom, at 96% each. 125 of the 184 country boards measured all quarter disagreed with global in every edition.September 2026
- 6
ChatGPT pointed 42% of its citations at the ranked brands' own sites in September 2026, and Grok 3%. Across every AI model with citations it was 11%.September 2026
- 7
Across 193 categories, a median 48% of cited sources were new in the September 2026 edition, measured against August 2026.August to September 2026
- 8
4 AI models (ChatGPT, Claude, Gemini and Grok) changed model version during the quarter, and Google AI Mode joined in September.August to September 2026
Held No. 1 all quarter
85 of the 183 categories measured in every edition kept the same No. 1 from July to September 2026.
Show all 85
Shopifyelectronic signature software
Amazon Pharmacyonline therapyBiggest moves, July to September 2026
Rank change from July to September for brands ranked in both editions, across the 183 categories measured in both. July ran on earlier model versions, so read these as the quarter’s full swing, not a like-for-like step.
Climbed
Amazon Fire HD 10tabletsNo. 13 to No. 2▲ 11Slid
Do the AI models agree on the No. 1?
In September 2026, every AI model with answers put the overall No. 1 first in 14% of categories, and at least half of them did in 66%. On average 59% of the AI models backed the No. 1.
- All pick the same brand13% (23 categories)
- Three in four or more agree17% (32 categories)
- Split two ways8% (14 categories)
- Three or more picks62% (114 categories)
- All pick the same brand9% (18 categories)
- Three in four or more agree18% (35 categories)
- Split two ways9% (17 categories)
- Three or more picks63% (121 categories)
- All pick the same brand14% (33 categories)
- Three in four or more agree23% (54 categories)
- Split two ways10% (24 categories)
- Three or more picks54% (129 categories)
Where the country boards disagree with global
Share of country boards whose No. 1 differs from the global No. 1 in the same category, September 2026. 125 of the 184 country boards measured all quarter disagreed with global in every edition.
- Australia94%
- Canada96%
- United Kingdom96%
- United States42%
| Edition | AU | CA | UK | US |
|---|---|---|---|---|
| July 2026 | 96% | 87% | 91% | 50% |
| August 2026 | 93% | 89% | 93% | 39% |
| September 2026 | 94% | 96% | 96% | 42% |
Where the AI models get their sources
Each AI model’s citations in the September 2026 edition, by source type, with the ranked brands’ own sites counted on their own. Across every AI model with citations, 11% pointed at those sites.
- ChatGPT42%
- Claude12%
- Gemini9%
- Perplexity9%
- Google AI Mode7%
- Copilot5%
- Grok3%
Share of citations to the ranked brands’ own sites, September 2026.
- Ranked brands' own sites42% (1,068 citations)
- Directory8% (200 citations)
- Reference4% (90 citations)
- News3% (88 citations)
- Other44% (1,115 citations)
- Ranked brands' own sites12% (364 citations)
- News8% (249 citations)
- Directory2% (75 citations)
- Blog2% (51 citations)
- Other76% (2,319 citations)
- Ranked brands' own sites9% (326 citations)
- Social16% (591 citations)
- Forum6% (217 citations)
- News6% (210 citations)
- Other64% (2,372 citations)
- Ranked brands' own sites9% (407 citations)
- News12% (571 citations)
- Forum6% (272 citations)
- Directory5% (248 citations)
- Other68% (3,140 citations)
- Ranked brands' own sites3% (144 citations)
- News12% (517 citations)
- Directory3% (154 citations)
- Marketplace1% (60 citations)
- Other80% (3,546 citations)
- Ranked brands' own sites5% (112 citations)
- Directory7% (157 citations)
- News6% (140 citations)
- Social2% (47 citations)
- Other80% (1,791 citations)
- Ranked brands' own sites7% (160 citations)
- Search11% (232 citations)
- News8% (182 citations)
- Social8% (168 citations)
- Other66% (1,433 citations)
Sources turn over fast. Across 193 categories, a median 48% of cited sources were new in the September 2026 edition, measured against August 2026.
How brands get named
Not this quarter. Being named isn’t the same as being recommended, and too few of this quarter’s answers carry the labels that tell the two apart, so the split isn’t shown. The count is under How to read this report, below.
Which model answered
The model behind each AI model’s answers, read from every archived answer. Changes happen only between editions and are dated in the methodology changelog.
| AI model | July | August | September |
|---|---|---|---|
| ChatGPT | gpt-4.1-mini | gpt-5.5-2026-04-23 | gpt-5.5-2026-04-23 |
| Claude | claude-sonnet-4-6 (73%), claude-sonnet-5 (27%) | claude-sonnet-5 | claude-sonnet-5 |
| Gemini | gemini-2.5-flash | gemini-3.5-flash | gemini-3.5-flash |
| Perplexity | sonar | sonar | sonar |
| DeepSeek | deepseek-v4-flash | deepseek-v4-flash | deepseek-v4-flash |
| Grok | grok-3 | grok-4.5 (96%), grok-4.3 (4%) | grok-4.3 |
| Copilot | Consumer surface | Consumer surface | Consumer surface |
| Google AI | Consumer surface | Consumer surface | Consumer surface |
| Google AI Mode | Not asked | Not asked | Consumer surface |
How to read this report
- Every figure is read from the published monthly editions (July 2026, August 2026 and September 2026) exactly as they were frozen. Nothing in this report re-scores a ranking.
- The editions measured 183 categories in July, 191 categories in August and 240 categories in September. Movement compares only categories present in both editions, and every figure is a share or a rank, never a raw count across months.
- The July 2026 edition used earlier model versions (gpt-4.1-mini for ChatGPT, claude-sonnet-4-6 for Claude on 73% of its answers, gemini-2.5-flash for Gemini and grok-3 for Grok), replaced from August 2026. August to September is the closest like-for-like step; in it, Google AI Mode joined and Grok moved from grok-4.5 to grok-4.3.
- Google AI Mode joined with the September 2026 edition, so that edition asked up to 9 AI models where earlier editions asked up to 8.
- In the August 2026 edition, Google AI named a ranked brand in only 7 of 191 categories, so that month's agreement figures rest on the AI models that did.
- In the September 2026 edition, Gemini named a ranked brand in only 49 of 240 categories and Google AI named no ranked brand in any category, so that month's agreement figures rest on the AI models that did.
- The July 2026 country boards were a single-day sample (22 July 2026), so regional figures for July are indicative.
- A country board flips when its No. 1 is named differently from the global No. 1, as on every edition page, so a local arm of a global firm counts as a different No. 1.
- The citation archive begins with the August 2026 edition, so source figures cover August and September only.
- DeepSeek returned no citations in August 2026 or September 2026, so it has no source mix here.
- Google AI returned 3 citations in August 2026 and no citations in September 2026, under the 20-citation floor, so it has no source mix here.
- Mention roles for September 2026 are not shown: the labels the rankings use cover 0% of that edition's answers, under the 90% floor. Labels from an earlier classifier are kept out, because it sorted mentions differently.
- Rankings in the July, August and September editions carry a disclosed recompute under a dated methodology change, noted on every record. The answers themselves are as captured.
- Citations from Google AI (the AI Overviews panel) are captured from the October 2026 edition onward.
- Rankings are computed from AI answers alone. Rankings are never for sale.
July 2026 edition → August 2026 edition → September 2026 edition → Read the methodology →