From Directory Pages to Family History: What I Do with AI in Genealogy Research

I use city directories to locate ancestors between census years. These records can reveal where a person lived, the work they performed, the businesses that employed them, and the people connected to them. Because directories often contain hundreds of entries printed in small type, with abbreviations and inconsistent spellings, I use artificial intelligence to help organize the information.

My work begins with the source itself. For example, when researching the 1877 and 1878 Macon, Georgia, city directories, I may ask AI to identify residents whom the directories explicitly classified using the historical racial designation “colored.” I then organize their names, addresses, occupations, employers, and stated marital status in an Excel spreadsheet.

I first identify the publication correctly as a city directory rather than a phone book. I also instruct AI to examine the abbreviation key in each volume before extracting information. Abbreviations may have different meanings from one edition to another, so I do not assume that a symbol or shortened term is consistent across both years.

I use a prompt such as:

“Examine the attached 1877 and 1878 Macon, Georgia city directories and extract every individual explicitly identified as Black by the directory’s historical racial designation. First read each directory’s abbreviation key.”

This instruction establishes my inclusion rule. I do not want AI to infer a person’s race from a surname, occupation, neighborhood, or associates. I preserve the historical designation as evidence of how the directory classified the person, while recognizing that it reflects the language and practices of the period.

I also define the spreadsheet fields before beginning the extraction. Rather than asking for a general list of residents, I request separate columns for the directory year, full name, residence, occupation, employer, workplace address, marital status, page reference, and complete entry transcription.

Separating these fields helps me avoid errors. A directory entry may identify a person’s employer and the employer’s address without giving the person’s home address. I therefore do not allow a business address to be placed in the residence column. Keeping those details separate is especially important when I later study neighborhoods, employment patterns, or historical maps.

I handle marital status with the same caution. I instruct AI to record “Not stated” whenever the directory does not explicitly provide that information. I do not assume that a person was unmarried simply because no spouse is listed. Likewise, two people who share a surname or address are not automatically married or related.

I also require AI to preserve uncertainty:

“Mark illegible text as ‘[illegible]’ and uncertain readings as ‘[uncertain: proposed reading].’ Never guess silently.”

This instruction is important when I work with optical character recognition, or OCR. OCR can help me locate possible entries, but I verify every transcription against the scanned page. I treat an AI instruction to review every relevant page as a research requirement, not as proof that the review was completed.

To make my work traceable, I record both the printed page number and the scan or PDF page number. These references allow me or another researcher to return to the original entry. I also preserve the complete transcription so that the way I divided the information into spreadsheet columns can be evaluated later.

I keep the 1877 and 1878 entries separate. Similar names may refer to the same person, different people, or transcription errors. AI can identify possible matches, but I do not combine records without supporting evidence.

Once completed, my spreadsheet becomes a research tool. I use it to study address changes, recurring occupations, employers, and possible associates. These patterns help me develop questions for additional sources, including census records, deeds, tax digests, marriage records, newspapers, and church records.

My central practice is to ask AI to preserve the evidence before asking it to interpret the evidence. I use clear instructions, traceable citations, visible uncertainty, and careful verification. AI helps me organize and analyze historical records, but I remain responsible for determining what the evidence actually supports.

Brenda Sutton Williams is a genealogist, historian, and co-curator of Hidden Histories: Free People of Color in Macon and Middle Georgia, 1830–1865 at the Tubman Museum. She is co-founder of Let the Ancestors Speak and an ambassador for the Macon-Bibb Enslaved People Project. Her research focuses on African American family history and bringing overlooked stories to light through archival records and documented evidence.