University of Toronto researchers found that AI answer engines cite very different websites than Google Search. In a test of 1,000 ranking queries, GPT-4o's cited domains overlapped with Google's top ten by 4.0% on average, and Perplexity Sonar Pro's by 15.2%. The paper, by Mahe Chen, Xiaoxuan Wang, Kaiwen Chen and Nick Koudas, appeared in the workshop proceedings of the EDBT/ICDT 2026 conference.
By EPR Research. Originally published June 17, 2026. Updated October 19, 2026.
Who conducted the Toronto study, and where was it published?
The study was written by four researchers in the Department of Computer Science at the University of Toronto: Mahe Chen, Xiaoxuan Wang, Kaiwen Chen and Nick Koudas, who is a professor there. It is titled "Navigating the Shift: A Comparative Analysis of Web Search and Generative AI Response Generation." It was published in the Proceedings of the Workshops of the EDBT/ICDT 2026 Joint Conference, held March 24 to 27, 2026 in Tampere, Finland, and is available as a free preprint on arXiv under a Creative Commons Attribution 4.0 license.
The authors state that they did not use generative AI tools to write or edit the paper. It builds on the team's 2025 paper on generative engine optimization, and the authors cite earlier work including the original GEO paper from ACM SIGKDD 2024. For background on the field, see Generative Engine Optimization.
How did the researchers test the engines?
The team compared five systems on identical prompts: Google Search, GPT-4o with web access, Claude 4.5 Sonnet with web access, Gemini 2.5 Flash with Google Search grounding, and Perplexity Sonar Pro in web search mode. All systems received the same prompts without personalization, and results were collected in the same time window to limit drift.
| Test | Queries | What it measured |
| Ranking queries | 1,000 | Domain overlap with Google across ten consumer topics |
| Entity comparisons | 216 (108 popular, 108 niche) | Whether niche queries change overlap |
| Consumer electronics, by intent | 300 (100 each informational, consideration, transactional) | Type of source cited: brand, earned or social |
| Freshness | 100 each in consumer electronics and automotive | Age of cited pages |
| Pre-training case study | Hundreds of SUV and niche ranking queries | Memory versus retrieved evidence, GPT-4o only |
The ten consumer topics were smartphones, athletic shoes, skin care, electric cars, streaming services, laptops, airlines, hotels, credit cards and smartwatches. The 1,000 ranking queries came from 100 fixed templates, such as "Rank the best {topic} from 1 to 10."
How much do AI engines overlap with Google?
Overlap was low for every engine. For each query, the researchers compared the domains an engine cited with the domains in Google's top ten and averaged the overlap across all queries. GPT-4o shared the fewest domains with Google, and its median overlap was 0.0%, meaning that for more than half of the queries it cited no domain that Google's top ten also contained.
| Engine | Mean overlap with Google | Median overlap |
| GPT-4o | 4.0% | 0.0% |
| Gemini 2.5 Flash | 11.1% | 8.5% |
| Claude 4.5 Sonnet | 12.6% | 8.7% |
| Perplexity Sonar Pro | 15.2% | 14.3% |
All pairwise differences between systems were statistically significant (p below 0.001) under a 10,000-iteration paired bootstrap. Niche queries raised overlap by only 3 to 4 percentage points for most engines. GPT-4o moved from 1.3% to 1.9%, a change the authors found not statistically significant. The authors read this as engines relying on a concentrated, shared set of review sites and discussion threads for niche topics, not on Google's ranking logic.
What kinds of sources do AI engines cite?
AI engines cite more earned media and less social content than Google. The researchers classified every cited source as brand (company-owned), earned (independent media and review outlets) or social (community and user-generated platforms), using GPT-4o for labeling with a manual spot check.
| System | Earned | Social | Brand |
| Google | 41% | 34% | 26% |
| Claude 4.5 Sonnet | 65% | 1% | 34% |
| GPT-4o | 57% | 8% | 36% |
| Perplexity Sonar Pro | 50% | 11% | 39% |
| Gemini 2.5 Flash | 46% | 7% | 46% |
The figures cover all 300 consumer-electronics queries and are read from the paper's source-category chart, so they are rounded and may not sum to 100. Source mix also shifted with intent. On consideration queries such as "best budget noise-canceling headphones under $200," AI engines cited earned media for 59% to 86% of sources, against a social-heavy Google profile. On transactional queries, brand citations rose to 52% to 68% across the AI engines. Claude returned no links for most informational and transactional queries until it was explicitly prompted to search.
Do AI engines cite fresher content than Google?
Yes, on the median, and the gap was larger in automotive. The researchers extracted a publication date from page metadata and visible text, then measured the age of each cited URL. Gemini was left out of this test. Coverage matters, because ages are computed only for pages where a date could be found.
| Engine | Median age, consumer electronics | Median age, automotive |
| Claude 4.5 Sonnet | 62.3 days | 148.0 days |
| GPT-4o | 79.8 days | 162.2 days |
| Perplexity Sonar Pro | 90.4 days | 216.6 days |
| Google Search | 130.4 days | 492.9 days |
Google's automotive median of about 16 months compares with roughly five months for Claude. Date coverage ranged from 0.426 to 0.930, so the authors also report a coverage-adjusted score. On that score GPT-4o ranked first in consumer electronics, narrowly ahead of Claude, and Claude ranked first in automotive.
When does an AI model rely on memory instead of search?
The model leans on memory for well-known brands and on retrieved pages for obscure ones. In a controlled case study of GPT-4o alone, the researchers retrieved web snippets for each query and then changed the evidence in three ways: shuffling the snippets, instructing the model to use only the snippets, and swapping two entity names throughout. They measured how far rankings moved over 10 runs per condition.
| Entities | Rank shift after snippet shuffle (normal) | Rank shift after name swap | Agreement between one-shot and pairwise rankings (Kendall's tau) |
| Popular | 2.30 | 2.60 | 0.911 |
| Niche | 4.15 | 4.63 | 0.556 |
For popular entities such as SUV brands, rankings barely changed, and on average 16% of the entities in the final rankings did not appear in any retrieved snippet. Miss rates ran from 0.03 for Honda and 0.06 for Toyota to 0.58 for Cadillac and 0.73 for Infiniti. For niche entities, such as family law firms in Toronto, rankings moved much more, and under a strict instruction to use only the snippets, the shuffle effect fell to 0.46. The authors describe a "knowledge-seeking mode" in which retrieved evidence directly shapes the answer when the model has little prior knowledge.
What are the limits of the study?
The study is a workshop paper with a defined scope, and several limits matter when applying it.
- It tests consumer topics, mostly electronics, automotive and ten ranking categories, and the authors do not claim the results hold for B2B or professional services queries.
- The pre-training experiment uses a single model, GPT-4o, so its memory-versus-retrieval findings may not apply to other engines.
- Source types were labeled by GPT-4o, checked by a manual spot check of a random subset.
- Engines were tested in their configurations at the time, and answer engines change often.
- The paper measures what engines cite. It does not test which tactics make a page more likely to be cited.
What does this mean for communications teams?
The findings suggest that a team should check each engine's sources for its own category and not rely on Google rankings as a proxy. This reading is EPR's, not the authors'. Three practical points follow from the data.
- Rank in Google is a weak guide to citation in an AI engine, since overlap with Google's top ten ranged from 4.0% to 15.2%.
- Earned media and fresh content were cited more often by the AI engines than by Google, which supports keeping independent coverage current.
- For lesser-known brands, retrieved pages shape the answer more than they do for household names, so the content an engine can retrieve matters most when a brand is not already well known to the model.
EPR's own tracking of engine sources is in the AI Platform Citation Source Index, and the term is explained in answer engines.
Sources
- Chen, M., Wang, X., Chen, K. and Koudas, N., "Navigating the Shift: A Comparative Analysis of Web Search and Generative AI Response Generation," Proceedings of the Workshops of the EDBT/ICDT 2026 Joint Conference, March 24 to 27, 2026, Tampere, Finland. arXiv:2601.16858.