Why AI Search Engines Disagree With Each Other

Ask Google Search, Google's AI Overviews and Gemini the same question, and you might expect them to cite roughly the same sources. A 2026 study of 11,500 real queries found something surprisingly different: the overlap between the sources they retrieved was less than 20%.

The study was led by researchers at the New Jersey Institute of Technology, Indiana University, and Nanyang Technological University, and it was accepted for the 2026 ACM SIGIR conference, one of the more selective venues in information retrieval. It compared the results returned by Google's traditional search engine against its own AI Overviews and against Gemini Flash 2.5, using average Jaccard similarity — a standard measure of how much two sets of items overlap — to quantify how often the three systems pulled from the same pool of sources for identical queries. The average overlap came in under 0.2, meaning the systems frequently drew from substantially different sets of sources for the same query — even when, in the case of AI Overviews and traditional Search, the query came from the same company's own infrastructure. It's worth being precise about what that measures: it says the systems retrieved different sources, not that they disagreed on the underlying facts.

That gap is worth sitting with, because it cuts against a fairly intuitive prediction from computer science research on algorithmic decision-making. When many systems increasingly share the same underlying components — training data, base models, retrieval infrastructure — you'd expect their outputs to converge, not diverge. That prediction has a name and a body of formal research behind it. AI-mediated search appears to behave differently, at least on this measure, and working out why is more useful than treating the finding as just another data point about "AI changing search."

What the Theory Predicted

The formal concern here is called algorithmic monoculture. It was named and modelled by Jon Kleinberg and Manish Raghavan, two computer scientists at Cornell, in a 2021 paper published in the Proceedings of the National Academy of Sciences. Their original context was hiring and lending: when many employers or banks all screen applicants using the same algorithm, even a highly accurate one, the pool of decisions made collectively can get worse, not better. Errors and blind spots stop being independent across decision-makers and start correlating. An applicant rejected by one bank's model gets rejected by every bank running that model.

Kleinberg and Raghavan showed this isn't a hypothetical edge case. It holds even under normal operating conditions, without any assumption of a shock or a bug, purely as a mathematical consequence of many agents converging on one scoring function. They compared it to agricultural monoculture: a single high-yield crop variety planted everywhere is efficient right up until a pathogen that exploits its specific weakness shows up, at which point the uniformity itself becomes the vulnerability.

A year later, a team at Stanford's Center for Research on Foundation Models — Rishi Bommasani, Kathleen Creel, Ananya Kumar, Dan Jurafsky, and Percy Liang — extended the idea in a paper presented at NeurIPS 2022. They introduced what they called the component-sharing hypothesis: if algorithmic systems are increasingly built on the same training data or the same pretrained foundation models, they will tend to produce more homogeneous outcomes for the same individuals or groups, even if the systems are operated by different companies with no direct coordination. They tested this empirically on fairness benchmarks and found that shared training data reliably increased outcome homogenization. Their finding on shared foundation models specifically was more mixed — the degree of homogenization depended heavily on how each downstream system adapted the shared model — but the general logic was clear enough, and directly relevant to a search landscape now built largely on a handful of large language models licensed, fine-tuned, and wrapped by dozens of different products. Neither paper studied search directly; applying their framework here is an extension of the theory to a new domain, not a finding either research group reported themselves.

If that logic carried over cleanly to search, you'd expect something close to the opposite of what the NJIT-led study found. Generative search products can combine foundation models with retrieval systems, proprietary indexes, ranking or reranking layers, and product-specific instructions, while different products may share some underlying models or training data. The component-sharing hypothesis would predict that this shared substrate pulls outputs toward convergence: the same handful of sources with strong topical authority getting cited across products, an information ecosystem funnelling toward a narrower set of winners. But sharing a foundation model is not the same as sharing a search product — a point the rest of the evidence bears out.

What Actually Happens

The SIGIR study provides a particularly useful empirical benchmark here, and its headline figure points toward fragmentation rather than convergence. The paper reports several other findings that sharpen the picture. Traditional Google search results were significantly more likely to draw from institutional sources — government and education domains. The generative systems, by contrast, were significantly more likely to surface Google-owned content instead.

A third finding is worth isolating on its own. Websites that had technically blocked Google's dedicated AI crawler were cited by AI Overviews significantly less often than websites that hadn't. This held even when the underlying page content was identical in every other respect, and even when Google's regular search crawler still had full access to the same page. The divergence here isn't a judgment call the model made about quality. It's a mechanical consequence of a site owner's robots.txt configuration.

Independent researchers have found comparable fragmentation in other markets. A separate large-scale preprint study of Chinese-language generative search, led by Zhen, Liu, Zhang, and Niu, covering four major platforms across both web and app interfaces, ran 614 queries and drew on more than 160,000 cleaned citation records out of an original pool of over 214,000. It found that the same platform's web and app versions frequently surfaced different source sets for identical queries. That means the divergence isn't only cross-company; it shows up inside a single company's own products depending on which interface is used. A related, smaller preprint working from a public citation dataset across ChatGPT, Google's Gemini-based results, and Perplexity described a sharp divergence between citation breadth and citation depth among the three platforms: some cite widely but shallowly, others narrowly but with heavier reliance on what they do cite.

A number of industry analytics vendors — companies that sell tools for tracking a brand's visibility across AI answer engines — have also reported low cross-platform overlap in their own data, describing gaps that are broadly consistent with the academic findings. These figures are directionally interesting, but they should not carry the same evidentiary weight as the academic research. Vendor methodologies vary, are not always disclosed in enough detail for independent auditing, and the companies publishing them have a commercial interest in convincing potential customers that cross-platform visibility tracking is necessary. None of them carries the peer-review scrutiny the SIGIR paper went through. They don't independently prove the academic finding; at most, they suggest the general shape of the finding isn't an artifact of one research group's particular query set.

Why the Prediction Missed

The gap between what monoculture theory predicts and what's actually observed comes down to where the shared component sits in the pipeline, and how much gets layered on top of it before an answer reaches a user. A foundation model is one input among several, not the whole system.

On top of that shared model sits a retrieval system that decides which documents even enter consideration. Product-specific instructions and fine-tuning choices then shape what the model treats as a good source. A reranking step follows that. And, as the SIGIR study's crawler finding shows directly, there is a harder constraint underneath all of it: whether a given product's crawler was even allowed onto a particular site in the first place.

Two products can share the same underlying language model and still diverge sharply as a result. Their retrieval indexes may be built differently. Their crawlers may be blocked from different sets of sites. Their fine-tuning may nudge the model toward different notions of what counts as a trustworthy source. Bommasani and colleagues' own mixed findings on foundation-model sharing anticipated exactly this outcome: model sharing homogenizes results only when the downstream adaptation is similar too. In AI search, the downstream adaptation can therefore be an important source of product differentiation, and potentially of divergence.

This retrieve-then-generate architecture, in which a system pulls candidate documents before a model writes an answer from them, is the same basic pipeline that other technical breakdowns of modern question-answering systems have described in more mechanical detail. A system built on a shared foundation model can still diverge substantially from another system built on the same model, depending on what its retrieval layer is even allowed to see. Neither Google nor the other companies involved publish the full internal details of these pipelines, so this explanation is best read as the most plausible mechanism consistent with the evidence, not a confirmed account of any single company's architecture.

Personalization adds a second source of divergence that has nothing to do with which model is doing the generating. Google says that AI Mode personalisation can use previous searches and other saved Search activity to tailor responses. When personalisation is enabled, two people submitting the same query to the same product can receive different results or recommendations based on their prior context. This within-platform fragmentation compounds the cross-platform fragmentation the academic studies measured.