Which AI Genealogy Tools Can Actually Find Your Ancestors' Records?

You ask a chatbot to find your great-grandfather's baptism. It answers in seconds, and the answer looks exactly like a citation: a parish, a register, a volume, a page number. You order the film, or you write to the archive, and the page does not say what the chatbot said it said. Sometimes the volume does not exist.

Or you go the other way and search one of the big record sites directly. It finds nothing, and you conclude the record does not survive, when the truth is that the site never held it.

Both mistakes have the same cause. Every AI tool that promises to find records can only find what it is able to see, and the tools see very different things.

The short answer. "Finding records with AI" means four different jobs, and no single tool does all of them. Full-text search, which FamilySearch made part of its standard search tools in a post dated 30 August 2025 and which Ancestry described in one dated 8 May 2026, reads the handwriting inside one organisation's own images and nothing else. AI record hints, such as FamilySearch's AI Research Assistant and MyHeritage's AI Record Finder, match your tree against one company's index. A general chatbot such as ChatGPT, Gemini or Claude can reason about where a record might be kept, but when it answers from memory it can invent the record itself: Walters and Wilder, writing in Scientific Reports in 2023, found that 55 per cent of the citations GPT-3.5 produced, and 18 per cent of GPT-4's, did not exist. The fourth job, working out which record your tree is missing and where that record is likely to be kept, is the one the first three were not built to start from. It is the job KleioBase, an AI research companion that turns your records into one connected family archive, is built around: it names the specific record a person should appear in, then searches the public web for the archive most likely to hold it.

"Find records with AI" means four different jobs

The easiest way to see the difference is to ask of each tool one question: what can it actually look at?

The table is the whole argument in miniature. The first two rows are powerful and each is fenced by one company's holdings. The third row has no fence and no floor. The fourth row starts somewhere different: not from a name you type in, but from the specific record your tree says should exist and does not.

Full-text search reads one archive's images, and only that archive's

This is the most important development in record-finding in years, and it is worth being precise about what it does.

FamilySearch first released Full-Text Search in its Labs area at RootsTech 2024. In a post dated 30 August 2025 it announced that the feature had become part of its standard search tools, and described it as letting you search nearly 2 billion images using transcripts generated by artificial intelligence. Unlike an ordinary indexed search, which only looks in the fields a volunteer chose to type up, it scans the whole transcript, so a name can match anywhere in a document: a witness to a deed, a neighbour named in a boundary, a godparent in the margin. It also produces AI-generated summaries of records, including the names and relationships in them. It is free, from an organisation that says on its own blog that it provides family history services for free.

Ancestry has built the same thing for its own collections. In a post dated 8 May 2026, Ancestry explained that its Applied AI Science team built its full-text search after experimenting with several new large language models, and that a result points you directly to where a name appears on the page. The same post says Ancestry is prioritising another 300 million record images for the searchable set, and a footnote notes that access requires an All Access subscription.

The fence is the collection. FamilySearch is careful to say that Full-Text Search is not yet available for all of its image collections. And neither tool reads anything its owner does not hold. The certificate in your grandmother's drawer, the letter a parish priest photocopied and posted to you, a register an archive digitised on its own website and never shared: none of those are inside either corpus, so neither search can return them. A null result means "not in the transcripts of these images", which is a much weaker statement than "not in the record", and it is exactly the kind of result worth writing down as negative evidence rather than treating as a conclusion.

The two tools are also not interchangeable, because they hold different images. If you have searched one and found nothing, you have searched one.

Record hints match your tree against one company's index

The second family of tools does not ask you for keywords. It looks at what you already know and goes looking for matches.

FamilySearch's version is its AI Research Assistant. A post on its blog dated 22 December 2025 describes AI-discovered hints on the signed-in home page that pick out records where a potential parent, spouse or child appears who is not yet in your tree. FamilySearch adds, in the same post, that all record hints, including the AI ones, need careful review.

MyHeritage went further towards conversation. It announced AI Record Finder on 26 December 2023 as a chat-based way to find historical records about an ancestor, running on technology from OpenAI and living under a Chat tab in MyHeritage's own record search. In March 2026 it added Scribe AI, which transcribes and translates an image you upload, explains its context, and suggests next steps.

These tools are genuinely useful, and they share the same limit as full-text search, one level up. A hint can only point at a record in the index it is matching against. MyHeritage's chat cannot hint at a record that only FamilySearch holds, FamilySearch's assistant cannot hint at one that only Ancestry holds, and none of them can hint at a record that nobody has put online. The suggestion is also only as good as the tree it was matched against: a hint built on a wrong birth year will confidently lead you to the wrong man, which is why FamilySearch's own caution is worth taking literally.

A chatbot with no search makes records up, and it has been measured

A general chatbot is the opposite case. It has no collection at all, which makes it more flexible and more dangerous.

When a chatbot answers from memory, it is not looking anything up. It is producing text that resembles an answer, and a citation is just a kind of text with a very recognisable shape. That is how you get the register, volume and page number in the opening of this article.

The best measurement of how often this happens comes from outside genealogy. William Walters and Esther Wilder, publishing in Scientific Reports in 2023, had ChatGPT write short literature reviews on 42 topics and then checked all 636 citations in the 84 resulting papers against real databases. 55 per cent of the citations produced by GPT-3.5 were fabricated, and 18 per cent of those produced by GPT-4. Even among the citations that did point at a real work, 43 per cent of GPT-3.5's and 24 per cent of GPT-4's contained substantive errors.

One thing not to claim, because it has not been measured. We searched in September 2026 for a comparable study of chatbots inventing genealogical record citations specifically, and found blog posts and forum warnings but no measured study. Walters and Wilder tested scholarly references, not parish registers, and newer models with web search switched on are a different case from the ones they tested. The direction is not in dispute. The rate for your kind of question is not known.

Web search changes the picture without removing the problem. A chatbot that searches can read pages a search engine has indexed: an archive's catalogue, a finding aid, a society's guide to a county's records, a transcribed volume someone published. Search alone does not run a query inside a record database that sits behind a login. Agent modes are the exception to watch: OpenAI's announcement of ChatGPT agent, dated 17 July 2025, describes a browser the agent operates and says you can log in to websites by taking the browser over. That is a different tool with different risks, and nothing in this article tests it. So a chatbot with search is good at the question "where would this kind of record be kept?" and poor at the question "does this record exist?", and it will not always tell you which of the two it is answering.

Used for the first question, it is a real help. Used for the second, it is where the invented baptism comes from. How accurate AI is at reading old handwriting is a separate question from any of this, and a more encouraging one.

None of them start from what your tree is missing

Look back at the table and notice what every row has in common. Each tool starts from an input you give it: a keyword, a tree, a question. None of them first works out which record should exist and does not.

That step is where experienced researchers spend their time, and it is where most people skip straight past. Before you search anywhere, it pays to be able to say, for a specific person: this man was alive and living in a known place at the date of a census, so he should be on it; this woman had four children baptised, so there should be a marriage before the first one; this family was in the country during a war, so there should be a draft or service record. A search that begins from that target is a different activity from a search that begins from a name, because it tells you what to look for, roughly where, and roughly when.

Take four ancestors from four different record systems.

A household in Ohio. A man born in the 1860s and living in Ohio in 1900 should appear in the 1900 United States census, and if he does not, that is a target rather than a mystery. The 1890 census is the trap on the same line. The National Archives records that a fire in the Commerce Department building in Washington on 10 January 1921 destroyed most of it, and that the surviving fragments name more than 6,160 people, so an ancestor's absence from 1890 proves nothing.

An Irish birth. For an ancestor born in Ireland after civil registration began, the irishgenealogy.ie site gives access to General Register Office civil records, and to church records that in many cases pre-date civil registration. Knowing which of the two you need changes where you look.

A Swedish parish. The Swedish National Archives, Riksarkivet, has made its digitised records free of charge since February 2018. That is a national archive's own site, with its own search, and a researcher who only ever searches the big platforms may never open it.

A Mexico City marriage. FamilySearch holds a collection titled "Mexico, Distrito Federal, Civil Registration, 1832-2006". Knowing the collection exists, and which years it spans, is a different piece of knowledge from knowing whether your ancestor's entry has been transcribed.

In every one of those four cases, the useful first answer is not a record. It is the name of the record, the place it should be, and the institution that holds it. That is the job the fourth row of the table is for, and it is the reason a clearly defined gap tends to break a wall that searching the name never did. The same logic, applied to the people around your ancestor rather than to your ancestor alone, is what makes FAN club research work.

Nobody publishes how often these tools miss

One gap runs under this whole article, and it should be stated rather than smoothed over.

None of the three platforms publishes a recall figure: how many of the records that do exist in its collection its AI search or hints actually find, or how often a hint points at the wrong person. We searched for one in September 2026, for FamilySearch's Full-Text Search, MyHeritage's record matching and Ancestry's hints, and found help pages, launch posts, forum threads and one blogger's hand count of Ancestry hints from 2016, but no rate published by any of the companies. So when a tool returns nothing, you cannot know whether you have hit the edge of the collection, the edge of the transcription, or a miss.

That is not an accusation. Machine transcription of historical handwriting is hard, the collections are enormous, and a single honest figure would be difficult to define. It is a reason to treat every empty result from every tool in this article as provisional, and to record what you searched so that a later search can pick up where it stopped.

What KleioBase does to find missing records, and what it does not

Disclosure, since we sell a tool.

KleioBase is an AI research companion that turns scattered historical records into one connected family archive. It links every person, date and place across the records you add or import, flags likely duplicates and conflicting evidence, and works out what is missing. Finding records is part of that, and it works from the fourth row of the table rather than the first three.

It starts with the gap. On the Researcher plan and above, a person's profile can show an Expected records panel: concrete targets worked out from where and when that person lived, such as "1900 United States census - Ohio", alongside their approximate age at the time. Census targets are generated for the United States, England and Wales, Scotland, Ireland, Canada, the Russian Empire, Austria-Hungary and the German Empire. Wartime targets cover the American Civil War, both World Wars and the Commonwealth equivalents, and missing birth, marriage, death and burial records show up too. The 1890 United States census still appears, labelled "fragments only", so it never sends you hunting for a census that mostly burned, and the same label marks the German Empire and 1897 Russian Empire censuses, whose name-level returns mostly were not kept. Working these out costs no AI credits.

Then the Research Companion takes the target somewhere. You can attach a research suggestion to a conversation and ask how to close it, and the Companion can search the public web and read pages to find the archives, collections and finding aids most likely to hold the missing record. Every web source it uses comes back as a plain link, kept visually separate from the citation chips that point into your own records, so you can always tell which is which. Web search is available on every plan, including the free one.

Now the honest limits. Three of them.

It points to records; it does not retrieve them. The Companion reads public web pages. It cannot sign in to FamilySearch, Ancestry, MyHeritage or any other account-based site and search inside their collections, so when the record you need is in a FamilySearch collection, Full-Text Search is still how you reach it.

Its suggestions are leads, not findings. A web search can be wrong about which archive holds which years, and each answer is capped at four searches and eight page reads, so a thin corner of the web gets a thin answer. Follow the links it gives you and check the holdings yourself.

The targets have edges. Census targets exist only for the eight countries listed above, and wartime targets only for the United States, Britain, Ireland and Canada, so a Swedish or Mexican ancestor gets birth, marriage, death and burial targets and nothing else. And a target is only as good as the facts it was worked out from: a wrong birth year produces a confident target for the wrong census.

Where to start with the gap you already know about

Pick one ancestor and one missing record, and work through the four rows in order.

Write down the target first: the person, the record you expect, the place, and a date range. If you cannot write that sentence, you are not ready to search yet, and the family tree that shows what is missing is a better place to begin.

Then run FamilySearch's Full-Text Search on more than the name. Search the place, the neighbour, the witness, the parish. If you hold an Ancestry All Access subscription, run the same search there, because the images are different.

Next, look at the hints each platform offers for that person, and review each one against the evidence you already have before you accept it.

Then ask the question none of those tools answers directly: if this record exists but is not online, who holds it? That is where a chatbot with web search, or the Companion, earns its place, and where you should demand a link for every claim it makes. An archive's catalogue page is a finding. A register number with no link is a guess.

Finally, write down every search that came up empty, with the collection and the years it covered. The next time the collections grow, you will know exactly which searches are worth running again.

The clerks who made these records were writing for a parish, a court or a census office, not for a search engine. Most of what they wrote is still sitting in the institution that commissioned it, in a volume no machine has read yet. The fastest route to it is rarely a better search box. It is knowing which volume to ask for.