Indexia Review

July 2026

Prepared by ASI AI Committee

Indexia is an AI-based software product. Indexia Technology, the organization behind Indexia, claims that Indexia produces publication-ready indexes 95% compliant with industry standards. The founders and developers, Ben Vagle and Will Dinneen, are not professional indexers.

We previously reviewed Indexia’s predecessor, AI-Indexing, from the point of view of whether it would be a useful tool for a professional indexer. In contrast, this review of Indexia focuses on the quality of its output, i.e. the indexes it produces. Although Indexia advertises itself as able to create publication-ready indexes directly, it also allows users to modify the generated indexes.^[i] Consequently, when planning our review, we debated whether we should modify a generated index. If we generated an index and then edited it to fully meet modern indexing standards, the index would reflect the competencies of the indexer rather than those of Indexia. If we generated an index and did not edit it at all, we would open ourselves to the charge that we were not reviewing Indexia as it is intended to be used. Ultimately, we decided to review an actual published index created with Indexia, to better reflect how Indexia is actually used.

The index we selected for review is for a trade book (i.e., not a dense or challenging scholarly book) aimed at a general audience on how to give effective virtual presentations. The book was published in December 2025 with the index presumably created earlier that year. We selected this book for two reasons: first, the author provided a testimonial which was published on Indexia’s website, indicating that the index was created with Indexia and also endorsed by both the author and Indexia Technologies; second, the book was available for download in PDF format, which facilitates indexing. We do not identify the book by name here as our purpose is to evaluate the index, not the book or its author.

We performed four analyses: first, we compared a professional index created for a section of the book to the corresponding subset of the generated index (i.e., all entries with page locators falling within the page range of that section). We then compared the generated index’s performance to that of indexes generated via directly prompting an AI chatbot earlier that year. Three professional indexers then evaluated the AI-generated index as a whole against current indexing best practices, with attention to potential issues. Finally, we performed an accuracy test of the index’s locators (page references).

Comparison to professionally created index

The professional index was created by an indexer who did not look at the generated index prior to indexing. They selected a 43-page section of the book for indexing and indexed it as they normally would. We then compared the two indexes on the number of main headings, subheadings, names, and cross-references, as well as on the average number of entries per page.

Compared to the professionally created index, Indexia severely under-indexed the text (see table below). The generated index contained fewer than a third as many main headings and about 40% as many subheadings as the professionally created index. It omitted nearly 90% of personal names that should have been indexed and included only 11% as many cross-references as the professional index, indicating a failure to provide the reader with appropriate navigation. Overall, the generated index had just over a third as many entries per page as the professional index (34%), and an average of 1.8 entries per page compared to the professional index’s average of 5.3 entries per page. This is very inadequate coverage of content compared to the professional index.

Professionally created section index statistics	Indexia section index statistics
●    94 main headings ●    106 subheadings ●    9 names indexed ●    18 cross-references ●    Average entries per page: 5.3	●    27 main headings (28.7% of professional index) ●    42 subheadings (39.6% of professional index) ●    1 name indexed (11.1% of professional index) ●    2 cross-references (11.1% of professional index) ●    Average entries per page: 1.8 (34.0% of professional index)

Comparison to indexes generated via direct prompting

We then compared the generated index’s performance to that of AI-generated indexes generated via direct prompting earlier that year. The quality of Indexia’s output is similar to that of ChatGPT and Claude; it falls in between the two chatbots’ performance but closer to ChatGPT, which performed the most poorly (see graph below).

 

Review of the generated index as a whole

In addition to severely under-indexing, the generated index contained numerous violations of good indexing practice. Following is a partial non-exhaustive list with selected examples:

• Absence of an appropriate metatopic entry or index structure. Professional indexes typically contain an entry for the book’s overall topic that includes cross-references to the book’s main subtopics (or sections). Other large topics are usually handled similarly, such that all information about a given topic can be reached either under the corresponding entry or via subheadings and/or cross-references. Locators are broken down appropriately to provide easy navigation for the user. In contrast, the metatopic entries in the generated index—”virtual presenting” and “presentation”—each have a single page range spanning the entire book (minus the introduction). Such entries are unhelpful to the reader and an inappropriate way to handle large topics.
• Under-indexing, or failing to index topics that are significantly discussed in the book. For example: posture, sitting, standing, body language, jokes, and laughter are all discussed in the text, but lack corresponding entries in the generated index. Similarly, much of the book is about delivery and presence, yet the entry on delivery and presence has locators that cover only four pages. The author also analyzes numerous video clips which are accompanied in the text by QR codes, but these clips are not represented in the index. As one indexer who reviewed the index commented, the index seems to miss the point of many sections.
• Uneven indexing. For example, the generated index includes uptalk but not monotone, quicktalk, or vocal fry, all of which were all discussed in the same chunk of text.
• Failure to appropriately connect related entries. For example, gestures are a form of body language and should be included in, or cross-referenced from, the body language entry. Similarly, the entries for webcams and cameras should have been connected to each other.
• Excessive numbers of locators at both 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 has eight main headings or subheadings that contain more than six locators, with the greatest number of locators being 15.
• Orphaned subheadings (main entries with a single subheading beneath them) which should have been condensed into main headings (the generated index has four of these).
• Cross-reference errors: Of the seven cross-references in the generated index, one points to a nonexistent entry, and a further four (two pairs) were circular/open, i.e. they pointed to each other, without a specific reason for doing so.
• Errors in pluralization, e.g. “script”, “presentation”, “typeface”, “camera”, all of which should have been pluralized. In addition, the generated index contained both an entry for “audiences” and an entry for “audience”.
• Unruly locators at main headings that have subheadings, where the unruly locators should have been assigned to subheadings of their own: the generated index contained 14 entries with subheadings but also with more than one locator at the main heading level, most of which should have been assigned to subheadings, plus 12 entries with subheadings with one locator at the main heading, at least some of which should have been assigned to a subheading.
• Inclusion of material that should have been omitted, e.g. passing references to a conference that did not need to be indexed.

The cumulative usability cost of these flaws is likely to be high. For example, it takes a reader extra time to sort through excessive and unruly locators, flipping through the text to check each one for relevance. It is frustrating to follow a cross-reference and discover it points to nowhere, or to a page they have already seen in the original entry, or to remember that there was something in the book about posture, with charts, but find nothing about it in the index. The end result is that the reader is likely to lose trust in the index. Furthermore, if this happens to a potential reader who is checking the index to get a sense of the book, they may lose trust in the book itself and choose not to read (or buy) it.

It is important to note here that, although the generated index suffers from under-indexing, the solution is not to over-index: more entries does not equal better indexing if the added entries are not valuable as access points. Quality matters more than quantity.

Accuracy check

To investigate the accuracy of locators, we used the copyeditor’s rule of thumb, a standard test for indexing accuracy, which is 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.^[ii]

For this purpose, we define an entry as a heading plus optional subheading/sub-subheading and a single locator. Inaccurate locators are those where the concept referenced is not present on that page, either as 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 late). The generated index failed the accuracy test at the first step.  We checked 10% of entries  (58 entries and found 29 locator inaccuracies (50% of entries checked), rendering it unnecessary to check a second 10%.

While doing the accuracy check, we also uncovered numerous additional issues that a professional indexer would either not have made or would certainly have remediated before delivering to the client. These included impositions, where Indexia substituted its own terminology for that of the author; entries where material was indexed that should not have been, either because it was a passing mention (not significantly discussed) or was part of the author’s acknowledgements, which is not traditionally indexed; and other bad entries, such as wording that misrepresented the author’s discussion of the term or concept. Of the 58 index entries checked we found only 12 to be adequate (20.7%), and numerous entries had not just one but multiple problems. In other words, all entries should have been checked before publication, nearly 80% of them likely should have been remediated, and the extensive under-indexing should have been addressed by going back through the book and ensuring the missing material was captured.

Conclusion

The index produced by Indexia was inadequate by every measure: it missed two-thirds of the indexable material (and almost 90% of the names), failed to use cross-references effectively to guide the reader to subtopics and related topics, failed a standard locator accuracy test, and contained a host of errors, omissions, and violations of professional indexing best practice. While the author may have made some changes to the index before publication, we find it highly unlikely that they introduced these issues, given that the issues observed are so similar to those found in indexes generated via direct AI prompting during a similar timeframe.

In short, Indexia’s current output appears to parallel indexes generated by directly prompting an AI without any intervening software layer. Those indexes are likewise poor: our cumulative review of our testing and analysis of indexes generated with AI prompting from early 2025–early 2026 finds that AIs fail to create an adequate index on prompting and that their performance at this task has not significantly improved over time.

We do not recommend that Indexia be used to generate indexes. While the developers’ lack of indexing training and experience may have contributed to the software’s shortcomings, based on our testing it is likely that AI simply is not up to the task of creating high-quality back-of-book indexes at this time, with or without an added layer of software.

^[i] By default, there is no professional indexer involved to catch and remediate poor indexing choices by either Indexia or the author. We will add here that should an author seek out a professional indexer to remediate an AI-generated index, the indexer will likely decline to do so: the issues inherent in AI-generated indexes—including those we have seen produced by Indexia—are so deep and pervasive that it will be faster and cheaper to index the material from scratch.

^[ii] Nancy C. Mulvany, Indexing Books, 2nd ed. (Chicago: University of Chicago Press, 2005), Kindle

Edition.