Where Title Search Errors Come From

Most title search errors are not errors of law. They are errors of reading and indexing — a surname keyed as written by the clerk, a marginal satisfaction read as a full release, a bearing copied one digit off — and each produces a clean-looking result that surfaces years later as a claim. This is a diagnostic pass through the five places those failures cluster, what to check in each, and how to keep a record trail you can stand behind.

Each of the five failure modes below yields the same outcome: a search that looks complete and a defect that emerges at the next survey, refinance, or claim. The control is the same in every case. Keep searchable text beside a re-verifiable image of the original, and check the name, the date, the scope word, and every bearing against that image before the abstract goes out.

The material itself is the constraint. Reading and searching deed books, grantor/grantee indexes, and metes-and-bounds descriptions means working through series long enough to cross many clerks and several generations of handwriting.

What the claims data can and can't tell you about title search errors

There is no national statistic for "title search error," and any figure presented as one deserves suspicion. What exists is claims data. The 2024 ALTA/Milliman analysis drew on 127,228 claims from underwriters representing roughly 70% of 2022 annual premium volume, on residential policies issued 2013–2022, with industry claims-related losses of about $2.8 billion for the period. By claim count, Examination and Opinion Irregularities accounted for 11.9%, and Endorsements/Title Plant/Search & Abstract for 7.8%.

Add those and you get 19.7% — but that addition is yours and mine, not a published causal figure, and claim counts are not loss dollars. The 2025 ten-year update, covering 161,934 claims on policies issued through the end of 2023, likewise does not isolate misread names, missed marginal releases, or transposed bearings as separately measured causes. The data is member-reported and unaudited, with a reporting lag.

The honest read: examination and search-and-abstract failures are a substantial share of claim counts, the specific transcription and indexing escapes below are not separately quantified anywhere public, and the exposure per file is large enough that you should manage them as if they were.

Failure mode 1: The name key in the index

A grantor/grantee index is a name index. The grantor (direct) side is searched backward through prior owners; the grantee (inverse) side forward through conveyances. A tract index organizes by parcel instead. The consequence of a name index is mechanical: it sorts on the written key. If a clerk wrote Reinholt where the instrument says Reinhold, or an abstractor keyed Mc Donald with a space, the exact-name search returns nothing and the searcher moves on.

The doctrine most often invoked here does less than people assume. Idem sonans presumes identity despite a misspelling where the spelling conveys substantially the same sound. It is a jurisdiction-bound rule of identification and notice, not a search operation, and its application to public indexes is contested. California's Orr v. Byers declined to charge a good-faith purchaser with notice of an abstract of judgment containing a misspelled name, and ALTA's own Title News discussion noted long ago that some courts hold idem sonans inapplicable to public records imparting notice, such as a judgment index. Other states are more notice-friendly. Local counsel and the governing statute control.

The practical posture: assume the doctrine will not rescue you.

The search plan that survives review

Run known spelling variants, initials, reversed given/surname order, spacing and hyphen forms, and Anglicized or phonetic equivalents.
Search entity forms as well as personal forms — Smith Bros., Smith Brothers, Smith Bros Construction Co.
Use phonetic or fuzzy matching to widen recall. Soundex codes surnames by sound rather than spelling, which is the right tool for recall and the wrong tool for identity. It returns candidates.
Resolve every candidate against the source image, not the index abstract, and record the hit/no-hit decision.

Names are where machine-assisted reading is most useful and most hazardous at once, because a plausible surname is easy to generate and hard to catch. If you rely on machine transcripts for name recall, prioritize verification of names, dates, and other high-stakes tokens against the page.

Failure mode 2: The instrument that was recorded but misindexed

Suppose the mortgage was delivered and recorded but indexed under the wrong name or the wrong book. Does it impart constructive notice? Nationally, unresolved. A D.C. bankruptcy decision, Albert v. Green Tree Servicing (In re El Erian), 512 B.R. 391 (Bankr. D.D.C. 2014), is reported in secondary case analysis as treating delivery for recording as notice because the statute made indexing no condition of recording. Orr points the other way in California. And the Florida Bar Journal's analysis of constructive notice and title-search fallibility shows how rigidly the doctrine can bind even where the search failed for indexing or description reasons — Florida Statutes §695.01 makes an unrecorded conveyance ineffective against a later purchaser without notice, but resolves nothing about indexing.

The legal outcome varies. The economic exposure does not. Search as though the index error will be charged to you: run the parcel side as well as the name side wherever a tract index or plat-based lookup exists, check book-and-page sequence around known instruments for gaps, and read deed-book images through the adjoining entries rather than jumping straight to the abstracted hit.

Failure mode 3: The marginal notation read as a full release

Older mortgage records were frequently discharged on the face of the record. Kentucky's KRS 382.290 required a blank space to be left in the record for exactly that purpose. South Carolina's Code Title 29, §§29-3-330 and 29-3-350 expressly recognize face-of-record and marginal cancellation methods and distinguish release from satisfaction procedures.

Two errors follow. First, the notation is missed entirely — cramped marginal script in a different hand and ink from the body, sometimes running into the gutter, often skipped by both human abstractors and automated text capture. Second, it is over-read: a partial release of one parcel out of five, or a release of a specific tract, gets recorded in a database as "released," and the remaining encumbrance disappears from the file. New York County's judgment and lien practice similarly distinguishes a satisfaction from a release for mechanic's liens.

The check is unglamorous. Open the image of the original mortgage page and its index entry, read the marginal notation in full, identify the property described in it, and reconcile it against any separate release instrument. Never rely on a normalized status label.

Failure mode 4: Statutory lien dockets searched like a grantor index

Judgments and mechanic's liens live in their own dockets under their own conventions, and one exact debtor-name search does not clear them.

Wisconsin §806.19 gives a judgment creditor a ten-year statutory lien on the debtor's real property by entry in the judgment and lien docket — so the effective search window, plus renewals or extensions where the state allows them, is set by statute, not by the transaction date. South Dakota's chapter 44-9 requires an owner's name, filing date, amount, property description, and release information for mechanic's liens. New York permits searching by name or by block and lot.

The plan therefore needs all of it: name variants including entity forms, property identifiers where the docket supports them, the full statutory duration with renewals, and image review of each candidate. Where the underlying series is a handwritten or typed court docket rather than a recorded conveyance, the reading problem is closer to transcribing historical case files and dockets than to abstracting a deed.

Failure mode 5: The bearing that moves the corner

A metes-and-bounds call is an instruction: point of beginning, bearing, distance, endpoint or monument, or a reference to an adjoining call. Change N 45° E to N 54° E and every subsequent point walks off line. Swap E for W and the figure inverts. Drop a call and the description loses a side.

Errors in these descriptions are a known and recurring source of trouble. Legal commentary on how property problems grow by metes and bounds treats miscopied directional bearings and distances as a standard defect class, and surveyors report errors of closure within record descriptions as among the most common issues they encounter. There is no published national rate for transcription-induced bearing errors specifically. What is clear is that the failure is latent. The traverse fails to close, or it plots a gap or overlap, and nobody knows until a survey, subdivision, or dispute forces the question.

Two habits contain it. Transcribe the description verbatim — degree symbols, abbreviations, and apparent inconsistencies intact. Do not tidy it. Then plot it, or have it plotted, and treat non-closure as a transcription hypothesis before treating it as a boundary hypothesis.

The common root: a derivative treated as the record

Every failure above shares a structure. Someone converted a page into text or an index entry, and someone downstream treated that conversion as the record.

FADGI's digitization guidelines frame optical character recognition as the conversion of a raster image of text into searchable data — the image remains the digitized record. That framing is the whole discipline in one sentence, and it holds for handwriting as much as print.

The dangerous conversions are the ones that read cleanly. Generic OCR pipelines are engineered for modern type and normalize toward it, "correcting" archaic orthography and misreading uneven ink and show-through as character evidence; reported error rates around 20% on historical material are common in the retrieval literature, and those errors at least tend to look like errors. Generative models fail differently and worse for this work: an illegible surname or a smudged bearing gets completed from language priors into something fluent and wrong. A 2025 benchmark on OCR hallucination under noisy image conditions documents the mechanism. A 2025 study of LLM transcription of historical handwriting reported encouraging accuracy on some tasks, but on a single non-deed corpus, in prose rather than dense tabular or formulaic legal text, and still with human verification assumed. Neither establishes anything about American deed books. Treat generative output as an assistive draft, never as the evidentiary record, and understand why fluent transcription errors are the hardest kind to catch.

Specialist handwritten text recognition (HTR) is the right class of tool, with one practical catch in a title plant: the established systems expect you to train on your own material first. Transkribus's data-preparation guidance sets a floor of 25 transcribed pages before training, and the comparative literature suggests useful handwriting models often need thousands of words, sometimes around 10,000 or more per hand. A long deed series contains many clerks and several eras of penmanship. That is not one training event. It is a program.

Reading deed-book hands without keying ground truth first

This is the stage Leo is built for. Leo's transcription model, ATR-1, is zero-shot: it reads Latin-script material — English deed books, but equally French notarial records, German or Dutch registers, whatever language is written in that alphabet — without per-office model training or ground-truth keying. It transcribes what is on the page. Abbreviations, contractions, strikethroughs, marginal additions, and archaic spelling stay as written rather than being smoothed into modern prose, which is the property you need when the disputed fact is a surname's spelling or a bearing's second digit.

The workflow matters as much as the model. Each page image sits beside its transcription, so verification is a glance rather than a document hunt. Retranscription writes to a new tab instead of overwriting, so the base reading stays intact and the file shows what was read and what was changed. Per-document metadata fields — including Collection, Box, Folder, and Identifier — hold book and page citations. Search runs across all transcriptions with adjustable fuzzy sensitivity, which is a recall instrument for name variants, not an identity determination. Export goes to PDF, Word, HTML, or TEI XML, and a read-only share link lets a reviewer or counsel see image and text together without an account.

On accuracy, the published benchmark comes from manuscript material, not deed books. On a randomized 97-image sample of early-modern English manuscripts from the Folger Shakespeare Library at ATR-1's release, Leo scored roughly 5% character error rate — 61% fewer errors than the next-best model tested (Transkribus/Text Titan I ~13%; Claude Opus ~23.3%; Gemini 2.5 Pro ~24.8%; GPT-4.1 ~56.7%). Extrapolate cautiously. Nineteenth-century county clerical hands are their own problem, and you should test on your own volumes.

One limitation stated plainly, because it bears directly on this material: pages where a heavy pre-printed form dominates and the manuscript entries are dense are the known weak spot, and the model can favor the printed headings over the written content. On those pages in particular, read the image. Where output looks structurally wrong, Leo's failure detection hides it, retries, and refunds the credit rather than shipping suspect text — and the errors that do get through are wrong characters and words you can catch against the image, not confident inventions.

What the file has to prove

An abstract is an argument, and its strength is the trail behind it. Whatever tools you use, the file should show which indexes you searched, which name and entity variants you ran, the date range and statutory window applied, the hit and no-hit decisions you made, and the image behind each one — because ALTA's Best Practices FAQ sets no national retention period and leaves the documented policy to you.

None of the five failure modes above are exotic. They are what happens when a name index is asked to do a name's work, when a marginal note is read at speed, when a bearing is copied by eye at the end of a long day. The examiners who avoid escapes are not the ones who read faster. They are the ones who keep the image within reach and never let a derivative stand in for the record.

Frequently Asked Questions

What are the most common title search errors and what causes them?

Most title search errors are errors of reading and indexing rather than errors of law. They cluster in five places: a surname keyed as the clerk wrote it, so an exact-name search returns nothing; an instrument recorded but indexed under the wrong name or book; a marginal notation missed entirely or over-read as a full release when it released one parcel out of five; statutory judgment and mechanic's lien dockets searched like a grantor index; and a metes-and-bounds bearing copied one digit off, which walks every later point off line. Each produces a clean-looking result that surfaces later at a survey, refinance, or claim.

How often do title search errors cause title insurance claims?

There is no national statistic for "title search error," and any figure presented as one deserves suspicion. What exists is claims data. The 2024 ALTA/Milliman analysis drew on 127,228 claims from underwriters representing roughly 70% of 2022 annual premium volume, on residential policies issued 2013–2022, with industry claims-related losses of about $2.8 billion. By claim count, Examination and Opinion Irregularities accounted for 11.9% and Endorsements/Title Plant/Search & Abstract for 7.8%. Those are claim counts, not loss dollars, and neither that analysis nor the 2025 ten-year update isolates misread names, missed marginal releases, or transposed bearings as separately measured causes.

Will idem sonans protect a search that missed a misspelled name?

Assume it will not. Idem sonans presumes identity despite a misspelling where the spelling conveys substantially the same sound, but it is a jurisdiction-bound rule of identification and notice, not a search operation, and its application to public indexes is contested. In California, Orr v. Byers declined to charge a good-faith purchaser with notice of an abstract of judgment containing a misspelled name, and some courts hold the doctrine inapplicable to public records that impart notice, such as a judgment index. Other states are more notice-friendly. Local counsel and the governing statute control, so run the variants yourself.

How do you catch a transcription error in a metes-and-bounds description?

Plot it. A metes-and-bounds call is an instruction — point of beginning, bearing, distance, endpoint or monument — so changing N 45° E to N 54° E sends every subsequent point off line, swapping E for W inverts the figure, and dropping a call loses a side. The failure is latent until a survey, subdivision, or dispute forces the question. Two habits contain it: transcribe the description verbatim, with degree symbols, abbreviations, and apparent inconsistencies intact rather than tidied, then plot it and treat non-closure as a transcription hypothesis before treating it as a boundary hypothesis.

Can OCR or AI transcription be trusted for deed books?

Treat any machine transcript as an assistive draft, not the evidentiary record. Generic OCR is engineered for modern type and normalizes toward it; reported error rates around 20% on historical material are common, and those errors at least look like errors. Generative models fail worse here: an illegible surname or smudged bearing gets completed from language priors into something fluent and wrong. Specialist handwritten text recognition is the right class of tool. Leo's ATR-1 is zero-shot and needs no per-office training; on a randomized 97-image sample of early-modern English manuscripts it scored roughly 5% character error rate, 61% fewer errors than the next-best model tested. Deed-book hands are their own problem — test on your own volumes.