Choose the unit before interpreting the count

Word, Character, and Line Diff Modes Explained

The same drafts can produce different change groupings when the comparison unit changes. Learn what each mode preserves and what its numbers actually describe.

Open Text Compare to inspect a supplied pair in Words, Characters, or Lines. The platform selector supplies context only; it does not change tokenization or apply native platform length rules.

Start with an exact comparison rather than a meaning test

A text diff first divides each source into units and then compares the resulting sequences. The unit might be a word segment, a user-perceived character cluster, or a line token. Those choices change which pieces can align, but they do not turn the operation into a judgment about meaning. “The workshop is free” and “There is no workshop fee” may communicate a similar idea while producing several differences. Conversely, adding a single word such as not can reverse an instruction. Use the display to locate supplied changes, then read the surrounding text before deciding whether the revision is appropriate.

Keep the direction of the pair clear. Original draft is the baseline; Revised draft is the proposal. Added text belongs to the revision, removed text belongs to the original, and unchanged text occurs in both aligned sequences. Swap drafts reverses that relationship and invalidates the prior results. The tool does not infer which source is newer from its wording or filename, and it does not label either draft a winner. If you accidentally paste the newer copy on the original side, the evidence can still be technically valid while your explanation of what was added or deleted becomes misleading.

The review retains enough text to reconstruct both validated sources: combine unchanged and removed chunks for the original, and unchanged and added chunks for the revision. This preservation is more useful than a similarity score when a reviewer needs to check an exact deletion. It also creates a disclosure responsibility because the deleted material remains available in reports. Mode selection changes how changes are grouped, not whether an old name or discarded sentence is retained. Decide whether the task is sentence review, character inspection, or line organization before comparing, and choose the export only after understanding what the complete report contains.

Read Words as all English-locale segments

Words uses Intl.Segmenter with the locale fixed to en and word granularity. It includes all returned segments, not just segments classified as words. Spaces, punctuation, and other separators therefore remain part of the sequence under comparison. A change from one space to two spaces is not silently ignored, and replacing a comma with a semicolon can appear even if every alphabetic word remains the same. This is an exact-text mode organized around word segmentation, not a stripped-down vocabulary comparison. Do not interpret its source-unit totals as an ordinary editorial word count or as a count of meaningful ideas.

Consider the original “Doors open Friday.” and the revision “Doors open early Friday!” A reviewer wants to see the added timing qualifier and the punctuation change in context. Words is often a useful first pass because it can keep familiar word-sized pieces readable while still preserving the surrounding separators. The precise change grouping belongs to the token sequence and alignment method, not a promise that every human edit appears as one colored block. When a term repeats several times, the library may align a different occurrence than the one you mentally associated with the revision. Read both sources if the highlighted grouping looks surprising.

English is the segmentation locale even when the supplied text contains another language, mixed scripts, hashtags, or identifiers. This is not language detection or a claim that every segment follows the writer’s preferred linguistic boundary. Browser ICU data can affect token boundaries, so a unit count or grouping should be understood with that implementation context. The exact source strings still matter more than a comparison of unit totals from different environments. For a disputed short spelling or symbol, inspect Characters; for a paragraph order question, inspect Lines. Changing mode is a deliberate second view, not a correction of the text or an automatic interpretation of its grammar.

Separate grapheme clusters from Unicode codepoint totals

Characters uses Intl.Segmenter with grapheme granularity and the same fixed en locale. A grapheme cluster is a user-perceived character unit rather than an individual Unicode codepoint or a JavaScript UTF-16 code unit. Combining accents and some emoji sequences can place several codepoints inside one cluster. This makes the mode useful for inspecting a short visible change without deliberately splitting every combined symbol into its underlying codepoints. It still compares the exact cluster strings, however. A cluster that looks like another cluster is not automatically considered equal merely because a font happens to draw them in a similar way.

For example, an accented e can be represented as one precomposed codepoint or as the letter e followed by a combining acute accent. Each representation can form one grapheme cluster, while their codepoint totals differ. The tool does not apply Unicode normalization to make those representations equal. If those are the only supplied sources, the comparison can show a removed cluster and an added cluster even though the visible letters look alike. That is expected exact-text behavior. Confirm the desired representation from the approved source instead of concluding that the comparison is broken or assuming it has repaired the spelling for you.

The summary’s original and revised character totals count Unicode codepoints, regardless of the selected mode. Characters mode units count grapheme clusters, so these two sets of numbers can differ without contradiction. Neither is a byte count, a screen-width measurement, or a native social platform allowance. A multi-codepoint emoji can occupy one comparison unit while contributing several summary characters and several UTF-8 bytes. Case, punctuation, and variation in the actual sequence remain significant. Use the totals as clearly defined descriptions of the supplied sources, and use the actual highlighted text to decide whether a visible name, symbol, or punctuation change is intended.

Understand line tokens and their trailing LF

Lines divides the normalized source into tokens that retain a trailing LF where present. A source consisting of A followed by LF is one token containing both characters, not the letter token plus an invented final empty line token. A source containing A, LF, and B has two tokens: A with its LF, then B. A source containing only LF has one newline token, while two consecutive LF characters form two such tokens. The empty string has zero tokens. These rules describe the comparison sequence, not every possible line-count convention used by an editor or a destination platform.

Because a trailing LF belongs to its line token, adding or removing that LF can make the last line token different even when its visible letters remain unchanged. This is useful when exact line structure matters, but it can look larger than a one-character edit. Recompare in Characters if you need to isolate that detail. Source CRLF and standalone CR are normalized to LF before the comparison, so two drafts differing only in those line-ending conventions can compare equal. This is the declared normalization exception; the tool otherwise preserves the case, spaces, punctuation, and Unicode sequences supplied in the two validated sources.

Use Lines for a description organized into separate instructions, a numbered production brief, or a body draft with stable paragraph boundaries. A moved paragraph may be represented as removed lines and added lines rather than as a special move label. There is no semantic matching of paragraphs and no outline analysis. Also separate source lines from visual wrapping: a long line can wrap across the result panel because of available width without gaining an LF. A changed browser width is not an edit to the source. Inspect the original line structure when the question concerns a deliberate blank line or an instruction moved to another part of the draft.

Inspect the same pair at the right scale

Choose Words, Characters, or Lines

Start with the review question, select a unit, and compare the current drafts again before relying on the result or its exports.

Know which whitespace and Unicode differences remain

Outside line-ending normalization, the comparison does not trim drafts, collapse repeated spaces, fold letter case, replace punctuation, or normalize Unicode forms. A leading space, a trailing tab, or a nonbreaking space can remain a real difference. That fidelity helps when a handoff must preserve exact wording, but it does not guarantee every difference is obvious in the chosen font. If two drafts look alike yet produce a change, inspect the approved source with an appropriate editor when necessary. The diff is not a dedicated hidden-character inventory and does not automatically label every invisible character with its Unicode name.

Tabs, Unicode text, and bidirectional controls are retained as text. Their retention is not a sanitization or privacy guarantee, and the utility does not certify how another application will display them. Supplied HTML-looking text is displayed as text rather than interpreted as user HTML; supplied addresses are not automatically made into links. This prevents the comparison view from becoming a rendering or browsing action, but it does not verify the contents of those strings. If a pasted passage contains an instruction, address, or embedded-looking markup, the tool compares the characters without deciding whether the material is appropriate to publish or trustworthy to follow.

Validation rejects NUL and unpaired surrogate values rather than silently deleting them. Each source must also fit all three input limits: 10,000 Unicode codepoints, 1,000 LF-delimited lines, and 64 KiB of UTF-8 text. Those limits measure different properties, so satisfying one does not waive the others. They are tool limits, not platform posting limits. If validation fails, repair the source deliberately or compare clearly identified smaller parts. Do not describe a shortened excerpt as the complete draft. An empty side is allowed for a full addition or removal, but at least one side must contain text before Compare drafts can run.

Interpret summary units without inventing an edit score

Original units and revised units use the selected token mode. Unchanged, added, and removed units describe the alignment of those same tokens. Original units equal unchanged plus removed units, and revised units equal unchanged plus added units. These identities explain the report rather than evaluate it. A line-mode replacement may remove one long line and add one long line, while a character-mode view of the same pair may isolate only a short phrase. Comparing raw counts across modes is therefore not a sensible way to decide which revision changed more meaning or required more editorial work. The denominator itself has changed.

Change runs count nonmatching chunks in the merged change list, not independent human edits. A replacement can involve a removed chunk and an added chunk, while one chunk can contain several related wording decisions. Adjacent chunks with the same status are merged by the library. Repeated text can also make several valid-looking alignments possible. Do not tell an approver that a report contains exactly three editorial issues merely because it shows three nonmatching runs. Use a separate decision list for changed facts, disclosure concerns, wording questions, and formatting choices, then associate those decisions with the passages that the comparison actually preserves.

The tool supplies no similarity percentage, quality grade, or preferred version. It does not measure readability, validate facts, detect plagiarism, judge tone, or automatically rewrite either source. A large unchanged portion does not prove that the remaining difference is safe, and a large change list does not prove that the revision is bad. For an announcement, changing a single date may require more scrutiny than rewriting three explanatory sentences. Read the highlighted text against the purpose of the draft and the approved reference material. The summary helps you navigate a comparison; the responsibility for accepting a change remains with the reviewer.

Use a bounded comparison as one review operation

The engine compares token arrays through the pinned jsdiff diffArrays implementation in a dedicated same-origin worker. The worker runs the comparison without putting private input into its URL. This is a local exact-text operation, not a remote language service or an API request to a social platform. The library comparison is bounded to 1,500 milliseconds and a maximum edit length of 6,000. Those are comparison safeguards, not promises that every valid pair must complete successfully. A pair may satisfy source-size validation but still exceed the permitted comparison work, particularly when the two token sequences differ substantially.

When the bounded comparison cannot produce a result, the tool reports an error rather than providing a partial success report. Cancel operation terminates an active worker comparison, and a timeout also terminates the comparison worker. Try a different mode when it answers the same legitimate question with larger units, or split the material at meaningful boundaries and identify those parts externally. Do not assume changing modes makes a refused comparison equivalent to the earlier intended review. A line-level result can answer where paragraphs changed while leaving a separate character-level question unresolved. Record the remaining inspection honestly instead of treating any completed run as interchangeable evidence.

Any input change invalidates the previous results and exports. This includes a mode or platform-context change as well as a wording edit, so select Compare drafts again for the current pair. Load example provides supplied sample drafts for exploring the controls; Clear drafts starts a new pair; Swap drafts changes direction. None of those actions creates an approval record. The tool compares one current pair, not an automatic revision history or a three-way merge. If multiple contributors have made separate proposals, compare each defined pair deliberately and keep the relationship between versions in your own notes rather than expecting the engine to reconcile them.

Preserve the right evidence for reopening and sharing

Save project downloads editable JSON with the supported kind, version, platform, mode, original, and revised fields. Open project validates that project schema, with a 1 MiB JSON file limit. Two empty drafts are valid in an editable project even though they cannot produce a comparison. Version one accepts pasted source text and project JSON, not source TXT import. There is no autosave, account, or background storage. Save an intentional project when you need to reopen the pair, and keep filenames and review decisions outside the source text. A review report or its attachment is not automatically the project file expected by Open project.

Copy revised and Download TXT contain only the revised source. Copy review, Download CSV, and Download PDF are report outputs that include both source versions and deleted material. The PDF uses explicitly bounded visible excerpts and embeds the full original, revised text, and changes in review JSON and CSV attachments. Removing a passage from Revised draft does not redact it from those reports. Before sharing, decide whether the recipient needs final wording or review history, then check the complete artifact’s disclosure scope. A short displayed excerpt and a small change count do not mean the full exported file contains little sensitive material.

The entered text is processed in the browser without source upload, network lookup, or API access; existing site analytics are separate. Clipboard copies, downloaded files, and an open browser tab still need appropriate handling. For the human decision process, continue with reviewing social media revisions. For the next exact pair, return to Text Compare and choose the mode before interpreting its totals. A useful review records the source pair, selected unit, understood changes, and remaining checks, rather than converting a technically correct alignment into an unsupported claim about quality, privacy, or publishing readiness.

Questions about Word, Character, and Line Diff

Why do Characters units differ from the character totals?

Characters mode compares grapheme clusters, while the summary character totals count Unicode codepoints. A combining sequence or emoji can contain several codepoints in one grapheme cluster.

Does Words ignore punctuation, spaces, or letter case?

No. Words includes all English-locale word-segmentation segments, including separators. Case and Unicode normalization differences are preserved; CRLF and standalone CR are normalized to LF before comparison.

Does a final newline create an extra empty line token?

No. Line tokens retain their trailing LF without an artificial final empty token. The empty source has zero tokens. A final LF can still change its preceding line token, and reports retain the exact validated sources and deleted text.

Continue with another text comparison guide

Compare supplied drafts locally in the Social Media Text Compare & Diff Checker, or choose another guide below. Find more practical resources in all free tools.

Keep the method beside the decision

Organize the work after the text review

Explore SMTasker for your broader social workflow once your team has understood the differences and approved the wording through its own review process.