Could OpenAlex become a single foundational source for evidence synthesis?

By Aaron Tay, Head, Data Services

Could OpenAlex be the starting point for systematic reviews, with specialist databases and supplementary searches used to fill its gaps?

Recent developments in OpenAlex make that question worth revisiting. The newly launched OpenAlex Query Language, or OQL, offers more expressive searching, while study-design classifications and access to open full text extend its potential role beyond initial discovery.

The attraction is straightforward. Traditionally, evidence synthesis involves searching several databases (e.g. Scopus, Web of Science, Embase, PubMed, PsycINFO) in parallel which involves translating strategies, managing and deduping overlapping results before doing the screening.

If only one could avoid all that by searching one single comprehensive source! OpenAlex's Walden update brought its index to 477 million records in early 2026, and the total has since exceeded 510 million. But there is a catch. Around 190 million records belong to a separate expansion corpus, consisting largely of datasets and repository records, which is not searched by default. It is a open question if this additional records help with recall without creating too much noise. But even the core index of over 320 million records is enormously larger than various traditional databases. But is size all that is needed?

There is already evidence that OpenAlex (or its predecessor Microsoft Academic Graph (MAG) can support living-review updates), but this is only for updating completed systematic reviews but does not establish the effectiveness of using OpenAlex for the initial Boolean search. EPPI-Reviewer’s auto-update service learns from papers judged relevant and recommends new OpenAlex records for screening, rather than matching predefined terms joined by AND or OR.  

But this is not the first time we have dreamt of this. Google Scholar’s enormous coverage once encouraged hopes that it could serve as a single source for systematic reviews. But studies in the 2010s showed why this was a pipe dream and a large index containing the needed literature is only part of what is needed.

Today, researchers are starting to study the same question for OpenAlex, a particularly interesting 2025 paper by Claire Stansfield showed OpenAlex might indeed be a promising search tools

The study did not support using OpenAlex alone, partly because of limitations in its Boolean search operators. Subsequent updates, including proximity searching and the new expressive Open Query Language, address several of those limitations. Early indications are promising, but whether these changes make OpenAlex sufficient as a single source remains the question.

The Google Scholar experience

Let's first look at Google Scholar to understand why despite it's massive index, it failed to be the one search source for evidence synthesis.

In 2013, Jean-François Gehanno's "Is the coverage of google scholar enough to be used alone for systematic reviews", first raised hopes that Google Scholar alone would be enough for systematic reviews. They managed to locate all 738 included references from 29 completed systematic reviews in Google Scholar. Other studies showed similar results, for most intents and purposes, papers included in systematic reviews were overwhelming likely to be indexed in Google Scholar. So why can't Google Scholar be used alone for systematic reviews?

Unfortunately, these tests demonstrated only known-item coverage, not effective subject searching: they showed "success" by looked up papers by their known titles. While this proves a included paper was in the index, a real systematic review instead needs a subject search that retrieves relevant articles without knowing their identities, while keeping the result set manageable enough to screen.

Google Scholar also had an additional quirk. It only showed roughly the top 1,000 results. This meant even if the item you needed was indexed in Google Scholar, it would do you no good if it was ranked below the top 1000!

Indeed, later studies by Boeker, Vach & Motschall (2013) and Giustini & Boulos (2013) all made the same point, showing that in practical terms, when you searched Google Scholar with an actual search strategy you were usually unable to retrieve all the relevant papers in the top 1000.

Bramer, Giustini & Kramer (2016) probably provided the strongest evidence against the use of Google Scholar alone for systematic reviews

In a prospective study, across 120 systematic reviews, they compared searches in Embase, MEDLINE and Google Scholar, distinguishing between whether included studies were present and whether the original subject searches had retrieved them.

They found that coverage of GS alone was high (97.2 %) often as much as Embase and MEDLINE combined (97.5 %). However, only 46.4 % of the included papers were among the downloadable first 1000 references in Google Scholar which meant that practically speaking you would miss more than half of relevant papers!

The lesson is that a potential single foundational source needs to meet several requirements:

The test is not simply whether a database contains the relevant studies, but whether a reproducible subject search can retrieve them. That requires expressive search logic—strict and nested Boolean operators, field searching, phrases, proximity, truncation and controlled vocabulary—along with support for long queries and access to sufficiently deep result sets.

Google Scholar famously fails on almost all of this criteria.

To be fair, some more recent studies have challenged this conclusion. Ungar, Muhanna and Chaushu (2023), for example, examined 10 meta-analyses in otolaryngology and head-and-neck surgery and reported recovering 107 of 108 included papers (99%) by searching up to the first 500 Google Scholar results, supplemented by reference-list searching. This is encouraging, although the findings concern a relatively specialised field and cannot necessarily be generalised to more complex review questions.

More controversially, Winn et al. (2024) argued that Google Scholar could serve as the primary database for non-medical systematic reviews. They reported recovering 1,116 included articles across 16 reviews, retrieving 50,793 records compared with 70,011 deduplicated records from conventional databases.

However, Lipke, Monnin and Fonseca (2025) challenged their methodology. Their rebuttal raised concerns about the distinction between finding previously known articles and conducting genuine subject searches. They also reported correspondence indicating that searches had involved looking up the titles of previously included articles. Winn and colleagues published a reply in April 2026 which did not address this point.

For me, the crucial distinction remains the same: being able to locate known articles in Google Scholar does not establish that a subject search can reliably discover those articles without knowing their identities in advance.

Can OpenAlex do better?

Can OpenAlex do better? After all, it does support more boolean searching functions and there is no arbitrary 1,000 record view/export limits like in Google Scholar.

Stansfield et al. (2025)'s, Analyzing the Utility of OpenAlex to Identify Studies for Systematic Reviews: Methods and a Case Study put this to the test, and also addressed several of these requirements. It distinguished between content coverage, meaning whether relevant records were present, and content capture, meaning whether the available search methods retrieved them.

The study used a systematic map of research on implementing diabetic eye screening in Australia, Canada, Ireland, New Zealand and the United Kingdom as a gold standard. The map followed systematic review methods, although it did not include formal synthesis or critical appraisal. The original search covered 26 resources, including databases, academic search engines and websites. Screening 12,293 records, it produced 131 included records.

Those 131 records became the benchmark for the OpenAlex evaluation. Throughout the following comparisons, recall means relative recall against that benchmark. It does not establish the proportion of all eligible evidence retrieved, and the unit being counted is the record rather than necessarily a unique study.

Of the 131 benchmark records, 128 records were indexed in OpenAlex giving 97.7% coverage. The three missing records were a journal article originally found through CINAHL Plus, a report found through Google Scholar and a master’s thesis also found through Google Scholar.

Coverage was therefore high, although some of the material found through specialist and supplementary searching remained absent.

Interestingly, the researchers also checked 34 records that had passed title-and-abstract screening in the original map but whose full texts could not be obtained. Only nine (26.4%) were present in OpenAlex. These were not confirmed eligible studies, so the figure is not directly comparable to the 97.7% coverage of included records. Nevertheless, it raises an interesting question about how representative coverage estimates based only on known included studies really are.

The next and more important question was whether a subject search could retrieve what OpenAlex contained.

The researchers translated the original Embase strategy into the OpenAlex website. This required substantial adjustments: proximity expressions became Boolean AND combinations, additional word variants replaced wildcards, and OpenAlex-generated keywords substituted for author keywords. Restrictions on search length and Boolean combinations meant splitting the strategy into nine searches.

The July 2024 search retrieved 126 of the 131 benchmark records (remember only 128 were indexed).

Two missed benchmark records were present in OpenAlex but were not retrieved by the July 2024 Boolean searches. Neither contained information about the relevant country setting in its title or abstract, and their author-country affiliation fields were blank. The authors attributed these retrieval failures to the absence of geographical information needed by their country-restricted search strategy.

While the relative recall seems reasonable, particularly since OpenAlex could not search MeSH as a controlled vocabulary, the screening workload was substantially larger.

The July 2024 search initially returned 21,747 records, reduced to 21,016 after deduplication and removal of records published before 2003. This compares with 12,293 records screened in the original multi-resource search, representing approximately 71% more records to screen.

Even the smaller March 2025 result set of 16,573 records would have involved approximately 35% more screening than the original multi-resource search.

When the researchers repeated the OpenAlex Boolean searches in March 2025, the result set fell substantially from 21,747 to 16,573 records. The authors linked changes in retrieval to the removal of closed-access abstracts from Elsevier and SpringNature.

More importantly, three previously retrieved relevant records were no longer found. These records were still indexed in OpenAlex but had lost their abstracts: two were published in Elsevier journals and one in a Springer Nature journal. Their abstracts remained available in PubMed.

Boolean boolean search with citation network search in OpenAlex.

The researchers also tested whether OpenAlex's citation network and related-works functionality could recover relevant studies missed by Boolean searching. Using the records retrieved by the Boolean searches as seeds, they conducted one round of backward citation searching, forward citation searching and related-works searching.

In July 2024, this network search returned 3,970 records and recovered one of the two relevant studies missed by the Boolean search. When repeated in March 2025, it returned 4,276 records and recovered four of the five missed studies, including three that had become unretrievable through Boolean searching following the removal of their abstracts. This illustrates how citation-based discovery can complement text-based retrieval when metadata is incomplete.

Consequently, despite the reduction in Boolean search recall between July 2024 and March 2025, combining Boolean and network searching retrieved the same 127 of the 128 original relevant studies present in OpenAlex (99.2%) at both timepoints.

The remaining record was a PsyArXiv preprint without backward- or forward-citation information in OpenAlex. Its suggested related records were not relevant based on their titles. Network searching therefore did not guarantee complete retrieval.

The researchers also tested a different approach: combining the original PubMed Boolean search with citation-network searching in OpenAlex. This yielded just 7,185 records for screening and retrieved 122 of the 131 benchmark records (93.1% relative recall). This was considerably more manageable than the OpenAlex Boolean searches, but came at the cost of lower recall.

This is an interesting alternative, suggesting that the most effective workflow might not involve using OpenAlex for everything, but combining a more targeted database search with OpenAlex's citation-network functionality.

One of the most interesting findings was that OpenAlex contributed evidence beyond the original benchmark.

After removing duplicates against the original review database and excluding records published from 2024 onwards, the researchers had 17,519 records to investigate. They developed six machine-learning classifiers in EPPI-Reviewer and used the best-performing one to prioritise screening.

The team assessed the highest-ranked 2,318 records and identified two additional eligible records: a report from the First Nations of Quebec and Labrador Health and Social Services Commission and an article from the International Journal of Ophthalmic Practice. Neither had been found through the original 26-resource search, and neither was recovered by the graph searches used in the evaluation.

This is a useful reminder that a conventional multi-source search is itself an imperfect reference standard. Recovering all of its included records would not necessarily mean finding all eligible evidence.

Improvements to OpenAlex in 2026 could change things

For Stansfield et al. (2025), OpenAlex’s broad coverage came with a larger screening workload. Limited search functionality, particularly the absence of proximity operators, forced the researchers to combine terms with AND, making retrieval less efficient. As with Google Scholar, coverage alone did not ensure practical searching.

Against this background, the arrival in 2026 of the new expressive OQL (OpenAlex Query Language) is significant.

It has support for nested Boolean expressions, fielded searching, phrases, wildcards and most importantly proximity expressions. It also allows switching between exact and stemmed matching to retain and share the search logic. These capabilities address several difficulties encountered in the earlier evaluation. In particular, OpenAlex now supported wildcard truncation and proximity operators, important features needed for evidence synthesis noted as missing in past reviews of OpenAlex.  Stansfield et al. (2025) described above also called out the lack of proximity operators.

A query can be run through the OpenAlex website or API. For longer, more complex searches, the API supports HTTP POST requests, which transmit the complete query in a JSON request body rather than embedding it in the URL. This bypasses the URL-length restrictions that previously required lengthy Boolean strategies to be split into multiple searches.

Another particularly useful new capability is automatic translation of search strategies from other databases. OpenAlex can now translate strategies written for PubMed, Scopus, Web of Science, Ovid Embase, PsycINFO and other databases into OQL. Unlike an LLM-based translation, it uses deterministic rules and identifies where the translation changes the search's meaning. This could substantially reduce the work of adapting complex strategies, although differences in searchable fields, indexing and controlled vocabularies mean that the translated strategy still requires validation.

Another important development is the improvement in meta-data with better keywords, topics and types. In particular, in October 2026, OpenAlex introduced a substantially revised keyword system and made "Title, abstract and keywords" its default search field.

OpenAlex reports substantial improvements in keyword quality, although these evaluations rely partly on AI judgements. The keywords are also automatically generated rather than author-supplied or manually curated

This means that potentially records can be retrieved through assigned keywords even when the search terms do not occur in their titles or abstracts (important with the reduction in abstracts from Springer-Nature and Elsevier).

However, OpenAlex still does not offer the equivalent of searching a controlled vocabulary such as MeSH in PubMed or Emtree in Embase. The OpenAlex team has announced plans to link its keywords to external vocabularies, including MeSH, but these should not be confused with existing controlled-vocabulary search capabilities. Whether the new keywords improve recall sufficiently without greatly increasing irrelevant results remains to be tested.

OpenAlex now assigns study-design classifications, including randomised controlled trials, clinical trials, observational studies, systematic reviews and meta-analyses. In September 2026, OpenAlex reported classifications for 21 million works and an impressive 99.7% precision for its randomised-trial tags. It also reported that 306,000 tagged randomised trials appeared in journals outside MEDLINE's coverage.

However, these are provider-reported figures, with the main precision evaluation relying on an LLM judge (Claude Opus 5.5), supplemented by a smaller full-text evaluation. Furthermore, the classification system is deliberately optimised for precision rather than recall. OpenAlex reports identifying only 69% of randomised trials in MEDLINE under its classification criteria. Works without abstracts cannot receive these automated study-design tags.

This makes the classifications potentially useful for filtering and prioritisation, but they should not be treated as a complete substitute for validated study-design search filters in systematic reviews.

Beyond discovery, OpenAlex is also expanding access to full text. It now provides cached access to more than 50 million open-access PDFs, with approximately 43 million also available as machine-readable XML. This potentially supports downstream evidence synthesis tasks such as screening, data extraction and text mining. However, full-text coverage remains incomplete, and the availability of a bibliographic record does not guarantee access to the underlying document.

Beyond Boolean: OpenAlex's new semantic search

OpenAlex has also introduced semantic search, based on neural vector embeddings. As of September 2026 , it is using Qwen3-Embedding-0.6B, a multilingual embedding model, allowing searches to identify conceptually related papers across languages.

Unlike Boolean searching, which depends on matching specified terms and expressions, semantic search can retrieve papers that are conceptually related to a query even when they use different terminology. Researchers can enter a natural-language description of their research question or even an entire abstract to find related work.

This opens up another possibility for evidence synthesis. Rather than relying exclusively on increasingly elaborate Boolean strategies, researchers could use semantic search as a complementary discovery method, potentially identifying relevant studies missed because of differences in terminology. This would be particularly interesting alongside the citation-network searching evaluated by Stansfield et al. (2025).

However, there is an important limitation. OpenAlex currently returns a maximum of 50 results for each semantic query. Such a small retrieval window makes it unsuitable as the sole search method for a comprehensive systematic review. Like citation searching, it seems more promising as a supplementary approach, although its ability to recover additional eligible studies has yet to be properly evaluated.

Could OpenAlex become a foundational source for evidence synthesis?

Stansfield et al. (2025) provide encouraging evidence that OpenAlex can retrieve a high proportion of the records identified through conventional multi-resource searching. Their evaluation also demonstrates the difference between having relevant literature in an index and being able to retrieve it efficiently.

Since that study, OpenAlex has introduced several capabilities that could change the balance. OQL supports more sophisticated Boolean strategies, including proximity searching, while automatically assigned keywords, study-design classifications and access to open full text expand its potential applications. The ability to translate strategies from other databases into OQL is another important development.

However, none of these improvements as promising as they seem, on its own, establishes that OpenAlex can replace conventional database searching without more testing. The larger corpus includes records with incomplete metadata, controlled-vocabulary searching remains limited, and the new search functionality has yet to be systematically evaluated against the strategies and benchmarks used in earlier studies.

Perhaps the more realistic question is not whether OpenAlex can become the only source for systematic reviews, but whether it can become the foundational one: a broad, open starting point, supplemented by specialist databases, trial registries, citation searching and targeted searches where necessary.

If such an approach can achieve comparable retrieval effectiveness with less searching, translation, deduplication and screening effort, the implications for evidence synthesis workflows could be significant. But demonstrating that requires evaluations of actual subject searches, not simply comparisons of database size or known-item coverage.