July 2026
Prepared by ASI AI Committee
IndexerLabs advertises “accurate, elegant, publication-ready indexes comparable to professional work”, produced using a custom model trained on 1,000+ indexes harvested from the open web. The IndexerLabs website offers indexes starting at $99/book in less than 2 hours and alleges that a comparable human index would cost $3,000/book and take 7–10 days. (As professional indexers, this strikes us as an excessively short timeframe and an excessively high price: while the cost and time do vary depending on the text, professional indexes for trade books typically take 1–2 weeks and run $700–1500, and professional indexes for scholarly books typically take 2–4 weeks and run $1500–2500.) Neither of the people behind IndexerLabs are professional indexers.
IndexerLabs’ AI was created by taking existing generative AI models, which have been trained on massive amounts of textual content harvested from the Internet, and further training those models on book indexes available on the open Internet. However, many published indexes are poor by today’s standards because they were not created by trained professional indexers or were created prior to the development of modern-day indexing practices. This creates two problems. First, the lack of quality in the training set influences all types of AI-based indexing—whether with general-use chatbot prompts or with an AI-based tool that has had additional training—because nothing in its training tells the AI which indexes are well-done and which are poorly done. Second, when an AI-generated index is compared to the book’s original index, if the original is of poor quality, then the AI will appear to perform better than it actually does.
This can be seen in the two examples provided by IndexerLabs, both of which show the AI-generated index side-by-side with the book’s original index. A quick review of both original published indexes shows that neither follows modern indexing standards. The published index for The Oxford History of the French Revolution predates modern indexing practice, and the published index for Neomania omits subheadings, an odd choice that was either a poor decision on the part of the publisher or of the index creator. Neither published index is structured effectively, lacking both a metatopic entry and effective use of cross-references.
For this review, we evaluated the IndexerLabs-generated index (hereafter, “the generated index”) for The Oxford History of the French Revolution. We performed three analyses. First, we evaluated the generated entries for one chapter of the book against an index for the same chapter that was created by a trained indexer following modern indexing practice. Second, we evaluated the generated index as a whole against modern indexing practice. Third, we performed an accuracy check of the generated index. We found it performed poorly in all analyses.
Comparison to modern professionally-created chapter index
The generated index entries for the 22-page chapter contained the same number of main headings as the professionally-created index—142—although the specific headings were not identical; the generated index contained some topics the professional indexer chose to exclude and excluded some they chose to include. The average number of references per page were also similar: 10 for the generated index, 9.5 for the professionally-generated index (see table below).
The two indexes differed primarily in subheadings, cross-references, names and titles of works, and handling of the chapter metatopic. The generated index under-indexed titles of works, picking up only 14 of 29 indexable works (48.3%). It indexed 36 names, compared to the 42 names indexed by the professional index: the generated index omitted 12 names indexed by the professional indexer (28.5% of the total names) and included three names that were passing mentions and should not have been indexed.
The generated index also contained far fewer subheadings (37 versus the professional index’s 71) and cross-references (zero for the chapter, and in fact only 10 for the entire index, compared to the professionally-generated index’s 7 cross-references for just the single chapter). The relative absence of subheadings and cross-references suggests broader structure and navigability issues with the generated index as a whole (more on that in the next section). The generated index also contained a small number of sub-subheadings, which were inconsistently used (7 sub-subheadings were present for 3 subheadings). In modern indexing practice, sub-subheadings are typically used very sparingly and only when a better solution is not available.
The chapter metatopic, i.e., the topic of the whole 22-page chapter, was the Enlightenment. The professional indexer handled the metatopic by including an entry with a chapter page span at the main heading, plus appropriate subheadings and cross-references in order to capture the entire topic discussion appropriately. In contrast, the generated index, in its entry on Enlightenment, included only two of the chapter’s pages, omitting over 90% of the chapter’s content from the metatopic entry. There was no equivalent metatopic entry elsewhere.
Overall, while the generated index for the comparison chapter included the same number of main headings as the professionally-created chapter index, it significantly under-indexed names and titles of works, failed to include subheadings and cross-references for navigability, and failed to include an appropriate entry for the chapter’s metatopic.
| Professionally created chapter index | IndexerLabs chapter index |
| ● 142 main headings
● 73 subheadings ● 42 names indexed ● 7 cross-references ● Average entries per page: 9.5 |
● 142 main headings (100% of professional index)
● 37 subheadings (50.7% of professional index) ● 36 names indexed (85.7% of professional index) ● 0 cross-references (0% of professional index) ● Average entries per page: 10 (105% of professional index) |
Review of the generated index as a whole
For our next analysis, we reviewed the generated index for the whole book and found it contained numerous violations of good indexing practice, such as:
- Failure to index large topics effectively: As mentioned above, the generated index included only 2 pages of the 22-page chapter on the Enlightenment under its Enlightenment entry. Other large topics were handled similarly poorly: for example, there is a 23-page chapter on the Directory (capital D), of which the index covers only 13 pages, all under the incorrect term directory (lower-case D), although all the pages in this chapter cover material either directly about or actively related to the Directory. There is no entry for “French Revolution”, the topic of the entire 425-page book. Other broad topics such as “popular democracy”, which are significant in their own right and also relate to other topics in the book, have minimal locators (page numbers) under their main entry and no cross-references. Similarly, the entry for clergy should be related to the entries for parish clergy and refractory clergy (better yet, there should be a main heading “clergy” that guides the reader to all information about the clergy).
- Failure to index all indexable information about smaller topics: For example, there is content about Calonne that is present in the text but not included in the index entry for Calonne. There is indexable information about a clerical oath on sixteen different pages, but only ten of these pages are indexed under entries relating to the clerical oath (and they are indexed inconsistently between entries; see below regarding double-posting.)
- Failure to include locators in multiple locations appropriately: For example, the entry for “peasant resistance” contains locators that differ from the locators under “peasantry, resistance to the Republic”, but these are the same topic so the locators should be identical. Similarly, the entries regarding the clergy’s oath fail to include the same locators at all locations: “clergy (First Estate): Civil Constitution of the Clergy and clerical oath” includes pages 137, 140–141, and 144; “clerical oath” includes pages 144–147, 304, 334, and 397; and “refractory clergy: clerical oath and origin” includes pages 144–145, 147. The entry for “Civil Constitution of the Clergy” also contains a number of locators that are not included under the corresponding subheading under “clergy (First Estate)”.
- Failure to use parallel structures for important topics: For example, the clergy entry includes a gloss of “(First Estate)”, but the nobility is not glossed as the second estate, and the peasantry and bourgeoisie entries are not glossed as the third (and there is no entry for merchants, who were also part of the third estate). A similar issue occurs with entries for estates: there is an entry for the Third Estate, but no entry for the second or first.
- Failure to use cross-references appropriately: The generated index contained a total of 10 cross-references, two of which pointed to nonexistent target entries. The remaining 8 cross-references consisted of four pairs of reciprocal see also cross-references (dog, See also cat; cat, See also dog). While reciprocal see also cross-references are not always a mistake, crafting entries so that only one cross-reference is needed typically results in a clearer delimitation of topics and a more accessible structure for the reader. The absence of cross-references for their most common purposes—to point the reader to subtopics and to point the reader from a term the reader might look up to the author’s term—is notable. For example, although there are entries beginning with “church”, none of them point the reader to the main entry for the Catholic church.
- Excessive numbers of locators at main headings and subheadings. Typical indexing practice allows for no more than six locators before the topic is either broken into subheadings or divided into separate main entries, to which the reader is directed with a cross-reference; the generated index contained 150 main headings or subheadings with more than six locators (9% of the total headings in the index), with the greatest number of locators being a remarkable 26.
- Failure to use subheadings where needed: the generated index contained 55 entries with subheadings that had more than one locator at the main heading, most of which should have been moved to subheadings, and 21 entries with subheadings with one locator at the main heading, at least some of which should have been moved to a subheading. Fifteen subheadings with sub-subheadings also had locators of their own, most or all of which should have been moved to subheadings.
- Subheading page ranges not matching topic page ranges: For example, there is a discussion of Cahiers de doléances that begins on p. 96 and continues to p. 97, but the corresponding index entry places p. 96 at the main heading and p. 97 at a subheading, even though the two pages are a continuous discussion. Similarly, the entry “Puisaye, Joseph, Comte de, 308-314” erroneously used a single page range for intermittent (non-continuous) mentions of Puisaye; the correct entry would have been “Puisaye, Joseph, Comte de, 308, 309, 311-312, 314”.
- Flipped and/or mischaracterized relationships: For example, the entry “Robespierre, Maximilien: on freedom of the press, 155” correctly identifies that Robespierre is on the page but misrepresents the content as Robespierre’s words on freedom of the press when in fact the quote on the page is from the Assembly’s decree, and the mention of Robespierre only states that he opposed the Assembly. Similarly, the entry “Austria: internal administration and unrest, 378” implies that the discussion is about Austria’s internal administration and unrest, but in fact this passage is about a coalition of other nations that experienced internal administration and unrest issues due to Austria’s actions.
The above issues can be summed up as failures to pick up topics, failures to appropriately characterize the beginning, ending, and nature of discussions, and failures to provide appropriately labeled and organized access for the reader.
Accuracy check
To investigate the accuracy of locators, we used the copyeditor’s rule of thumb, a standard test for indexing accuracy, defined as follows:
A copyeditor’s rule of thumb for checking accuracy in an index is to randomly look up 10 percent of the entries in the referenced pages. If only one or two inaccurate entries are discovered, then the editor should check another 10 percent of the entries. If no inaccuracies are found in the second group, one can hope that the inaccuracies of the first group are anomalous. But if more inaccuracies are discovered, the entire index must be checked…..If an index is full of inaccurate reference locators, it is not usable.[i]
For this purpose, we define “entry” as a heading, plus optional subheading/sub-subheading, and a single locator. For AI-generated indexes, we define inaccurate locators to be ones where the concept referenced (in the main heading, subheading, or both) is not present on that page, either by the indexed term or a synonym, or where the provided page range does not accurately represent the boundaries of the discussion (i.e., it starts too early or ends too early or too late). The IndexerLabs index failed the copyeditor’s rule of thumb almost immediately—within the first ten randomly selected entries checked. This means that all entries in the index would have to be checked and any errors remediated before the index should be passed to a client.
Due to the length of the index and how quickly it failed, we elected not to check a full 10% of the entries and instead checked a total of 102 randomly selected entries (2% of the entire index). Those 102 entries contained 23 inaccurate locators (22.5% of total entries checked) as well as an additional 34 entries (33.3%) that were problematic for other reasons, such as imposing the AI’s own terminology or formatting over the author’s terminology or formatting, being part of an entry with too many locators and thus needing subheadings or mischaracterizing the nature of the discussion. Only 45 (44.1%) of the 102 entries checked were adequate, meaning that over half the entries in the index would likely require remediation.
A note on elegance
Although IndexerLabs does not provide a definition for its claim to elegance, the criteria for the ASI indexing awards describe elegance as follows: “Succinctness; the right word in the right place—even if the word isn’t found in the text; a certain ‘charm’; visual appeal; a sense that the index contains exactly what it needs to, no more, no less; simplicity; grace. Elegance is the quality that makes an exceptional index more than the sum of its parts.” In addition, the ASI training course includes a reading on “creating elegant subheadings”.[ii]
An elegant index is intuitive and easy to use. It goes beyond the basic standards of accuracy and completeness necessary for a usable index. An index that fails at minimal usability cannot be elegant, which already excludes the generated index reviewed here. More specifically, the generated index fails at including exactly what is needed, as it misses large parts of major discussions but includes passing mentions. It lacks the parallel structure that provides a sense of graceful order to the reader and the clear cross-references that simplify that structure. Instead of including words not in the text as double-posted synonyms or as cross-referenced guides to the author’s preferred term, the software imposes its own terminology: not “the right word in the right place.” Finally, the presence of long locator strings and unruly locators damages the visual appeal of the index, producing a haphazard look. Agee and Towery note that “Elegance manifests itself in the balance of art and science in an index.”[iii] We propose that such a balance requires a human understanding of each book as a unique human-to-human communication between author and reader.
Conclusion
The generated index under-indexed titles of works; missed indexable names and included passing mentions; failed to use cross-references effectively to guide the reader to subtopics and related topics; failed a standard locator accuracy test; and contained various violations of professional indexing best practice, mainly related to failing to include all indexable information about a topic, failing to appropriately characterize the location and nature of discussions, and failing to divide and organize the information in ways that provide effective access to the reader. In short, it is neither accurate, nor elegant, nor publication-ready. We do not recommend that IndexerLabs be used to create indexes.
[i] Nancy C. Mulvany, Indexing Books, 2nd ed. (Chicago: University of Chicago Press, 2005), Kindle
edition.
[ii] Agee, V. and Towery, M. 2010. “Creating Elegant Subheadings: Or, Everything You Ever Wanted to Know About Subheadings and Were Afraid to Ask.” pp. 1-35 in Index It Right! American Society for Indexing.
[iii] Ibid.