Last month, we were joined by Tricia McCraney, Head of Account Management for North America at Virtusales, developers of BiblioSuite. In this session, Tricia shared up-to-date best practices on how to use your metadata to stay discoverable to AI search engines, as well as how you can keep yourself informed as AI continues to change book discoverability.
In this post, we go over a few key takeaways from this session, some follow up questions that we didn’t have the chance to address in the session, and, to see how publishers have been adapting to AI, we take a look at the data in BiblioShare to see how many titles carry details about AI use.
Three key takeaways
Keywords and phrases have always played an important role in increasing book discoverability. With the introduction of AI into our search habits, everything from how we think about search terms to how search engines produce results have been changing.
A few key things to remember when accounting for these changes:
Keywords are unique metadata
- Avoid repetition: For the most effective results, it’s important that your keywords do not include any information that can be found anywhere else in your ONIX Product record.
- Use phrases: It’s also important to support keywords with phrases. Tricia recommends building phrases using 2 to 5 words and that they be up to 50 characters long.
On using varying metadata across different platforms
- Core metadata needs to be consistent so AI can recognize your title as the same product across different platforms. (Remember: no keyword stuffing in titles and subtitles!)
- Descriptive data, however, can vary by recipient or trading partner, but ideally should aim to maintain some consistency.
Populate your backlist metadata
To keep your backlist titles discoverable as metadata requirements change, you should revisit your backlist titles and make sure you:
- Add subjects and descriptive data where none exists
- Add a synopsis, extract, or excerpt
AI according to BiblioShare
Recently, it has come to our attention that publishers have been requesting to add data to their metadata indicating where AI was or was not used to help with the various aspects of the book creation.
With AI moving as fast as it is, we were curious to see if publishers have been able to start providing information about AI usage. So, we took a quick look at BiblioShare to see how many titles currently carry details about AI. And while we didn’t find much as of right now, we expect that to change shortly and we wanted to take this opportunity to go over all the areas of metadata where you can disclose the use of AI.
This summary is pulled from an excellent resource released by EDItEUR called Aspects of AI in ONIX that is making the rounds in the industry. There are a few areas that might support AI disclosures — whether that be the use or non-use of AI.
Unnamed person(s)
One area in the metadata you can look into is the unnamed person element in the Contributor composite. There are values that allow you to say a synthesized voice or AI was used.
05 Synthesized voice – male
06 Synthesized voice – female
07 Synthesized voice – unspecified
08 Synthesized voice – based on real voice actor
09 AI (Artificial intelligence)
While it may seem so at first glance, this does not only pertain to audiobooks. This is specifically in reference to value 09, which says the unnamed person was actually an AI. You would use code 09 when the creator (of text, of images, etc.) is a generative AI model or technique. Note: This can also be combined with the role ‘assisted by’.
Currently, there are no records in our data that have these values applied.
Website Role
The next area in ONIX where AI is having a moment is Website Role. There are two new codes in the Website Role: 55 and 56 that can be provided in the <Website> composite nested in the <Publisher> composite. Here is where you can declare whether AI was used or not. This is the spot where you should be looking to deal with the requests that are coming from places like Overdrive or Spotify.
The headings for this are actually stated in a negative way, i.e., a third-party or first-party declaration of the ABSENCE of AI-generated content.
If you are working with a third-party certifier to validate that no AI was used in the creation of the product, you can use code 55 and provide a link to the certificate on the certifier’s website.
While third-party certification is still relatively new, the suggestion coming from EDItEUR is for companies to create an AI policy statement. You can then post your policy on your website and link to your own policy statement. This is where code value 56 (first-party declaration of ABSENCE of AI-generated content) would be used.
Again, there are no titles in the BookNet Canada BiblioShare data set with either of these values yet.
Usage type and Usage status
And finally, the last area we looked at was Usage type and Usage status. These are the places where you can tell AI models whether you are open to or opposed to having your data used for data mining purposes.
11: Text and data mining
Make use of the content of the product (text, images, audio, etc.) or the product metadata or supporting resources for extraction of useful (and possibly new) information through automated computer analysis, or for training of tools for such analysis (including training of generative AI models). By convention, use this in conjunction with <EpubUsageStatus> codes 01 – permitted unlimited or 03- prohibited in <EpubUsageStatus>. It’s important to note that code 03 should be regarded as ‘prohibited to the full extent allowed by law’, or otherwise expressly reserved by the rightsholder, as in some jurisdictions, text and data mining (TDM) may be subject to a copyright exception (e.g., for not-for-profit purposes), subject to optional reservation, or allowed under ‘fair use’ doctrine.
13: TDM for AI to enhance the private purchase experience
Where text and data mining is prohibited or otherwise expressly reserved by the rightsholder, make use of the content of the product (text, images, audio, etc.) or the product metadata or supporting resources to train or enable a generative AI-enhanced purchase experience for a customer, with no use outside of a private AI instance controlled by the reseller — for example, to provide AI-assisted book selection.
14: TDM for AI to enhance the private reading experience
Where text and data mining is prohibited or otherwise expressly reserved by the rights holder, make use of the content of the product (text, images, audio, etc.) or the product metadata or supporting resources to train or enable a generative AI-enhanced reading experience for the purchaser, with no use outside of a private AI instance controlled by the reseller or embedded in the product itself — for example, to provide AI-enhanced reading.
Usage Type value 11 is the only value being used out of all of these possibilities at this point. But we can expect this all to change very quickly.
Your questions answered
Below are some of the questions that came up during the session and Tricia’s answers:
Are publishers providing new metadata on older collections or titles? And if they do, are they using AI tools to help them?
It’s normal practice to create good, comprehensive metadata for new releases, so revisiting backlists and filling metadata gaps can be a good internal project. Examples of where publishers might create new metadata for those older titles: adding Thema subject categories where none exist; creating short descriptions or series descriptions. We know through conversations with our customers that some publishers are using off-the-shelf AI tools to generate subject codes and keywords. However, consumer tools are not as focused on publishing as AI offered within systems like BiblioSuite. Biblio offers in-system AI tools that allow our customers to generate these same metadata points from existing metadata and using manuscript ingestion (if desired).
How can we boost discoverability while respecting copyright and intellectual property rights?
A key consideration for publishers in relation to copyright and intellectual property rights, is whether or not the content of the published book is available to AI. We addressed how to explicitly state that you are opting out of AI text and data mining for digital products. The broader question is whether or not a publisher will allow manuscript ingestion for use by Artificial Intelligence to create metadata. One way of avoiding manuscript ingestion, while working to boost discoverability, is to create more and better metadata that can be used by AI. For example, rather than ingesting the complete manuscript, a publisher could create and have AI ingest a synopsis, extract, or excerpt that is geared to use by AI and not public-facing.
What might creating and running a compliance report on backlist metadata entail?
A compliance report is simply a report containing your metadata that can be evaluated and measured in some way. Publishers should be able to create this view of data with any of the reporting options available through their title management system, or even through spreadsheets managing metadata in-house. A compliance report will ideally give you a view of the presence or absence of data (to check where you need to create metadata, such as missing subject codes). It can be helpful to sort or link the data in your report by Work (i.e., to view and assess all formats or editions of a title together); Author or Contributor (i.e., to check for continuity of information across an author’s titles); Series (i.e., to confirm all members of a series are present, numbered (where appropriate), and using the same Series Name).
How long should an excerpt be?
An excerpt that is an intended to form part of your book’s descriptive copy should be a few paragraphs in length, typically shorter than what publishers provide for look-inside-the-book.
How long should a keyword phrase be?
It’s best to check with your trading partners on this, but we can suggest aiming for a 2 to 5 word phrase that is anywhere from 30 to 50 characters in length. For example: “enemies to lovers hockey” for Heated Rivalry or “Canadian winter apocalypse thriller” for Waubgeshig Rice’s novel Moon Over the Crusted Snow.
Is there a difference between the metadata for a series and titles from a specific imprint?
ONIX 3.0 offers the <Collection> composite, which allows books to be grouped together in ways that communicate commonality. As such, the titles published under an imprint can be grouped together as a Collection in ONIX using Collection Type 10, indicating a Publisher Collection.
<Collection>
<CollectionType>10</CollectionType> <!-- Publisher collection -->
<TitleDetail>
<TitleType>01</TitleType>
<TitleElement>
<TitleElementLevel>02</TitleElementLevel> <!-- Collection level -->
<TitleText>Your Imprint Name Here</TitleText>
</TitleElement>
</TitleDetail>
</Collection> Should you be including/duplicating BISAC codes in keywords?
No, please don’t include BISAC codes, such as FIC027020, as keywords. Keywords should communicate data that is not part of your structured metadata anywhere else in your ONIX record, including BISAC subjects. However, you might find some overlap in keyword phrases and BISAC subjects, such as when “enemies to lovers hockey romance” is used and the book has a BISAC subject code of FICTION / Romance / Contemporary.
What fields in ONIX are best for anniversaries when it’s not a new edition?
The descriptive text fields are best for this. Try adding it at the start of your main description, or as a Promotional Headline.
Browse our collection of blog posts to learn more about the latest on standards and metadata. Want more from Tech Forum? Sign up for our newsletter, subscribe to our YouTube channel, and keep your eyes peeled on our website to register for upcoming webinars and rewatch old favourites.
