The short answer: are your dialogue tags getting in the way
Most dialogue-heavy drafts read cleanest when tags stay near 1 to 4 per 100 words, said supplies more than half of every tag, adverbial tags stay near zero, and at least a third of your dialogue units carry no tag at all. If your audit shows 5 or more tags per 100 words, or said below 40 percent of tags, or three or more -ly tags in one scene, the tags are competing with the conversation instead of disappearing behind it. Paste 150 to 300 words, compare your rate against the stated convention for your genre, then cut adverbial tags first, repeated colorful verbs second, and redundant tags third. The tool on this page only counts the words you supply with fixed deterministic rules and states its norm bands openly as editorial conventions, not market data or accuracy promises.
That paragraph is the whole method in miniature. Everything below explains how the counting works, shows the arithmetic on real numbers, walks through hypothetical scenes, handles the tricky cases, and adjusts the advice by genre. Read the sections you need and skip the rest.
What a dialogue tag actually is
A dialogue tag is a short clause that names the speaker with a speech verb. Mara said. Jonas asked. She whispered. He replied. The tag exists to answer one question: who is talking right now. Nothing more. When tags do that job quietly, readers forget they exist. When tags add adverbs, pile up on every line, or swap in a fresh colorful verb each time, readers start noticing the machinery instead of hearing the voices.
A beat is different. A beat is an action or a small piece of grounding written as its own sentence. Mara folded the letter. Jonas looked away. Rain crossed the window. Beats also identify the speaker, because the action sits beside that speaker's line, but they add movement, setting, or emotional information the tag cannot carry. Strong scenes lean on beats for texture and reserve tags for orientation.
An action used as a tag is a third category and a common source of clutter. She grinned the words. He laughed his reply. Grinning and laughing are not speech verbs, so they cannot tag dialogue grammatically, and editors flag them. Rewrite them as beats: She grinned. Then give the line its own sentence. The audit on this page counts a list of nearly sixty speech and sound verbs as tags, including edge cases like laughed, sighed, grinned, nodded, and shrugged, precisely so these disguised tags surface in your inventory instead of hiding.
How this audit counts your sample
The tool performs pure string analysis on the text you paste. It does not send your words anywhere, does not consult any database, and does not generate prose for you. It assembles counts from fixed word lists and arithmetic, then frames those counts against openly stated editorial conventions. Here is each measurement in order.
First it splits your sample into words on whitespace and counts them. A 200-word paste is 200 words regardless of genre, formatting, or quality. Every rate below divides by this count, which is why very short samples produce jumpy rates. One tag in 20 words is 5 per 100. The same single tag in 250 words is 0.4 per 100. The tool warns you below 40 words that the reading is unstable and asks for a longer passage.
Second it scans the lowercased text against a fixed inventory of speech and sound verbs and counts every match. Said, asked, replied, whispered, shouted, admitted, teased, warned, and the rest each add one to the tag count. A 200-word sample with 9 verb matches has 9 tags. The matching is case-insensitive and whole-word, so Said and SAID count, but words containing said as a substring do not. Spans inside straight or curly double quotation marks are blanked before counting, because attribution lives outside the quotes: promised inside a spoken line is story content, while said after the closing quote is the tag.
Third it counts said separately and divides by total tags for said share. Six said out of 9 tags is 67 percent. Said share matters because said is the only tag most readers skip without registering. Asked is nearly as quiet. Every other verb draws some attention, and attention spent on tags is attention taken from the exchange itself.
Fourth it hunts adverbial tags with two patterns: a tag verb followed by an -ly word, and an -ly word followed by a tag verb. Said softly. Whispered angrily. Quietly admitted. Each match adds one to the adverbial count. These patterns are deliberately narrow. A free-floating -ly word elsewhere in a sentence does not count. Only an adverb directly attached to a speech verb counts, because that is the construction that tells an emotion the line should already show.
Fifth it builds a verb inventory. Every matched verb gets a tally, sorted most frequent first, capped at twelve rows, each with a count and a share of total tags. Repeats surface instantly. Whispered three times in 200 words reads as a tic. Said eight times reads as discipline. The table flags any non-said verb appearing more than once as repeated and suggests varying or cutting it.
Sixth it estimates beats versus tags. It divides your sample into units, using line breaks when you paste multiple paragraphs and sentence boundaries when you paste one block. A unit counts as tagged when it contains a tag verb outside any quoted span, so Mara asked counts even as its own short sentence while promised inside a spoken line does not. Everything else counts as untagged. Untagged units divided by total units gives beat share. A scene with 12 units and 4 tagged has 67 percent beat share. High beat share means orientation is handled and tags can thin out.
Seventh it computes tags per 100 words as tag count divided by word count times 100, rounded to one decimal, and compares the result against the genre band you selected. General fiction expects 2.0 to 4.0. Romance expects 2.5 to 4.5. Mystery and thriller expect 1.5 to 3.5. Science fiction and fantasy expect 2.0 to 4.0. Literary expects 1.0 to 2.5. These bands are editorial conventions stated openly in the tool, not measurements of any market or claims about reader behavior. They exist so your number has a stated reference instead of floating alone.
Eighth it assigns a verdict. Inside the band with healthy said share and few adverbials earns a pass. Just over the band, or said below 40 percent, or three or more adverbials earns a warning with targeted cuts. Far over the band, or adverbial density above 1.5 per 100, earns a fail with a three-step cutting order. Under 40 words earns an informational verdict asking for a longer sample before any revision.
Worked numbers: three samples done by hand
Follow these calculations once and the tool output will never feel mysterious. Each example shows the raw counts, the division, and the verdict logic.
Worked sample one: a 200-word passage with 9 tags, 6 of them said, 2 adverbial, across 12 units of which 4 carry tags. Tags per 100 equals 9 divided by 200 times 100, which is 4.5. Said share equals 6 divided by 9 times 100, which is 67 percent. Beat share equals 8 untagged units divided by 12 total times 100, which is 67 percent. Adverbial density equals 2 divided by 200 times 100, which is 1.0 per 100. Against general fiction at 2.0 to 4.0, the 4.5 rate sits just over the top, so the verdict is a warning. The prescription is concrete: delete one tag per 200 words and convert both adverbials to beats, which lands the passage at 8 tags per 200 words, exactly 4.0 per 100, inside the band with said share rising to 75 percent.
Worked sample two: a 180-word literary passage with 3 tags, all said, zero adverbial, across 10 units of which 2 carry tags. Tags per 100 equals 3 divided by 180 times 100, which is 1.7. Said share is 100 percent. Beat share equals 8 divided by 10 times 100, which is 80 percent. Against literary at 1.0 to 2.5, the 1.7 rate sits comfortably inside, so the verdict is a pass. Nothing needs cutting. The only maintenance advice is to protect the pattern by resisting any new -ly tag during revision.
Worked sample three: a 150-word passage with 11 tags, 2 of them said, 5 adverbial, across 11 units of which 9 carry tags. Tags per 100 equals 11 divided by 150 times 100, which is 7.3. Said share equals 2 divided by 11 times 100, which is 18 percent. Beat share equals 2 divided by 11 times 100, which is 18 percent. Adverbial density equals 5 divided by 150 times 100, which is 3.3 per 100. Against any band on the page this fails: the rate doubles the top, said is a small minority, and adverbials alone exceed the density that triggers a fail. The cutting order is fixed. Remove all five adverbials first, swap the repeated colorful verbs toward said, then drop every second redundant tag. Cutting six tags lands at 5 per 150 words, which is 3.3 per 100, inside general fiction and romance and close for thriller.
Notice what the math rewards. Said share rises when you replace verbs rather than adding words. The per-100 rate falls when you delete tags or add untagged beats and narration around the same tags. Beat share rises when orientation moves from tags to actions. All three levers are under your control in revision, and the recommendations block on this page always lists them in that priority order.
Illustrative hypothetical example one: the tag-heavy reunion
This first scenario is an illustrative hypothetical constructed to show a common failure pattern. Imagine a writer named Dana pasting 210 words of a reunion scene between two estranged sisters. The audit reports 12 tags, 4 of them said, 4 adverbial including said tearfully, whispered bitterly, cried suddenly, and admitted quietly, with hissed appearing twice and snapped once, across 13 units of which 9 carry tags.
Run the arithmetic. Tags per 100 equals 12 divided by 210 times 100, which is 5.7. Said share equals 4 divided by 12 times 100, which is 33 percent. Beat share equals 4 divided by 13 times 100, which is 31 percent. Adverbial density equals 4 divided by 210 times 100, which is 1.9 per 100. Against general fiction at 2.0 to 4.0 this fails on rate and on adverbial density together.
Dana's revision follows the tool's order. First the four adverbials become beats. Said tearfully becomes a beat about the shaking cup plus the bare line. Whispered bitterly becomes a short action showing the bitterness, tagged with said or left untagged. Cried suddenly loses suddenly entirely, because crying already carries urgency. Admitted quietly becomes a beat about the lowered voice. Second, hissed twice becomes said once and an untagged line once, since two speakers need fewer tags than Dana assumed. Snapped stays once because a single sharp verb in a 200-word scene adds snap without becoming a tic. Third, two redundant tags on clearly alternating lines come off. The result is 6 tags in roughly 215 words, which is 2.8 per 100, with said at 5 of 6, zero adverbials, and beat share near 60 percent. Same conversation, same information, dramatically less machinery.
Illustrative hypothetical example two: the clean two-hander that still warns
This second scenario is a hypothetical built to show a subtler outcome. Imagine a writer named Theo pasting 190 words of a fast two-person argument in a car. The audit reports 8 tags, 7 of them said, zero adverbial, across 11 units of which 7 carry tags. Tags per 100 equals 8 divided by 190 times 100, which is 4.2. Said share equals 7 divided by 8 times 100, which is 88 percent. Beat share equals 4 divided by 11 times 100, which is 36 percent. Against general fiction at 2.0 to 4.0 the verdict is a warning despite the excellent said share and zero adverbials, because the rate sits just over the band.
Theo's fix is the smallest possible edit. In a two-speaker scene with no action, every third or fourth tag can come off once speakers are established. Deleting two said tags from alternating lines drops the count to 6 in 190 words, which is 3.2 per 100, inside the band, with said share unchanged at high levels and beat share rising past 50 percent. If Theo worries about clarity, one deleted tag returns as a beat: a hand tightening on the wheel, a glance at the mirror. The lesson is that even disciplined said-only dialogue can warn when tags appear on lines that do not need them, and the cure is deletion rather than substitution.
Why said disappears and everything else announces itself
Readers process said the way they process punctuation. Eye-tracking and editorial experience alike suggest the word resolves who is speaking and then vanishes from awareness. Asked behaves similarly for questions. Every other speech verb carries meaning beyond identification. Whispered specifies volume. Snapped specifies temper. Admitted specifies reluctance. That extra meaning forces a micro-decision: does this specification match the line, add to it, or contradict it. One such decision per page is texture. Ten per scene is fatigue.
This is why said share is the second number to check after the per-100 rate. A passage at 3.0 per 100 with 80 percent said reads quieter than a passage at 2.5 per 100 with 20 percent said, because the second passage asks the reader to evaluate a fresh verb choice every few lines. The audit flags any non-said verb appearing more than once so you can see which choices repeat. A single teased in a warm scene is characterization. Teased four times is a label the writer keeps reapplying.
Asked deserves a special note. It is nearly as invisible as said when attached to genuine questions, and the audit treats it as quiet in the table. But asked attached to non-questions draws attention, and demanded or questioned in place of asked adds confrontation the line may not earn. Keep asked for questions, said for everything else, and beats for everything said cannot do.
The adverb problem in detail
Adverbial tags fail for a structural reason. Said softly, whispered angrily, and asked curiously each split one emotional job across two words that do not trust each other. The verb claims a manner and the adverb grades it, while the dialogue line itself already performs a tone. Three overlapping signals force the reader to reconcile them. Most of the time the line alone would have sufficed.
Revision follows a fixed substitution. Read the line aloud with no tag. If the emotion is audible, delete the adverb and keep said, or delete the tag entirely. If the emotion is not audible, keep the line and add a beat that shows the cause: hands, breath, posture, an object handled, a glance held or avoided. Her voice dropped to barely a whisper carries more than whispered softly because it stages the sound instead of grading it. He folded the napkin in half before answering carries more than admitted reluctantly because it shows reluctance through behavior.
Frequency guidance is simple. Zero adverbial tags per scene is the target. One per chapter can survive when the line genuinely needs grading and no beat fits. Three or more per scene always warns in this audit, and density above 1.5 per 100 always fails, because at that concentration the pattern is visible to any reader. The default sample on this page carries several adverbials deliberately so first-time users see the flag fire and learn the substitution on their own prose next.
Beats versus tags: how much of each
Think of orientation as a budget. Readers need to know the speaker roughly every third or fourth line in a two-person scene, more often with three or more speakers or with long paragraphs between lines. Tags and beats both spend from that budget. Tags spend cheaply and add nothing else. Beats spend the same orientation and add setting, body, or subtext. So the efficient scene uses tags for pure traffic control and beats everywhere orientation can also do descriptive work.
Beat share between 40 and 70 percent suits most commercial scenes. Below 30 percent the scene usually feels airless, all voices and no bodies. Above 80 percent the scene can feel over-staged, with every line wrapped in gesture. Literary scenes often sit higher on beat share with lower tag rates overall, which is why that genre band runs 1.0 to 2.5. Romance tolerates more tags because rapid banter needs traffic control, hence 2.5 to 4.5. Neither band is a law. Both are stated starting points this tool shows you before you decide.
Placement matters as much as proportion. A beat before a line colors how the line lands. A beat after a line comments on it. A tag in the middle of a line creates a pause the line may not want. When revising, move orientation to the position that serves the line: beats forward for setup, beats after for reaction, tags wherever they interrupt least, usually at the end of short lines. Read the revised scene aloud. Any tag you stumble over is a tag to cut or convert.
Sound verbs, action verbs, and the verbs that cannot tag
The audit inventory includes nearly sixty verbs, and not all of them are equal. Pure speech verbs like said, asked, replied, answered, added, and continued identify speech cleanly. Manner-of-speech verbs like whispered, murmured, muttered, shouted, and yelled add volume information and draw moderate attention. Attitude verbs like snapped, demanded, insisted, protested, and retorted add conflict information and draw strong attention. Disclosure verbs like admitted, confessed, denied, revealed, and recalled add knowledge information and can feel explanatory when repeated.
Then there are verbs that cannot grammatically tag dialogue at all: laughed, sighed, grinned, nodded, shrugged, and similar actions. She laughed the words is ungrammatical because laughing is not speaking. Editors correct these to beats on sight. The audit counts them as tags on purpose, not because they are valid tags, but because flagging them shows you exactly which lines need restructuring into a beat plus a line. If your inventory shows grinned or laughed, rewrite those two or three lines first. The fix is mechanical: action as its own sentence, dialogue as its own sentence, tag only if the speaker is unclear.
Watch also for elaborate verbs that appear once each: ejaculated in older prose aside, modern clutter verbs like vocalized, verbalized, opined, articulated, and orated sit outside the tool inventory entirely, which is itself a signal. One each across a chapter still registers as showing off because each forces a fresh evaluation. Said would have vanished. When your draft shows six different verbs at one count each, the revision is consolidation toward said, not further variety.
Edge cases and failure modes you should know
Short samples mislead. Below 40 words the tool returns an informational verdict instead of a judgment, because a single tag swings the rate by several points. A three-line exchange with one said is either fine or terrible depending on context the counter cannot see. Always audit at least a full scene fragment, ideally 150 to 300 words with at least two speakers and some narration between lines.
All-dialogue transcripts break the beats math. If you paste bare alternating lines with no narration, every unit carries a tag and beat share reads near zero. That reading is accurate for the pasted text but not for your scene, because your scene presumably contains grounding the paste omitted. Repaste with the surrounding paragraphs included. The audit needs the narration to credit your beats.
Multi-speaker scenes need more tags than the band assumes. With three or four voices, orientation costs rise and a rate near the top of the band or slightly above can be correct. The tool does not count speakers, so apply judgment: if every tag in a four-person scene is said and the rate reads 4.6 against a 4.0 top, clarity may justify the excess. Cut adverbials and colorful verbs first, then reassess whether the remaining said tags are truly redundant before deleting for the number.
Heavily styled dialects and sparse punctuation can confuse sentence splitting. The unit splitter falls back to line breaks when present, which is why pasting with paragraph breaks gives a more reliable beat share than pasting one wall of text. Preserve your line breaks when you paste. A few seconds of formatting visibly improves the measurement quality.
Internal monologue in quotation marks counts as dialogue to the counter. If your sample mixes spoken lines with quoted thoughts, the tag rate reflects both. That is usually acceptable, since quoted thought needs orientation too, but know that a passage heavy with quoted thought and said tags may read as cluttered for a different reason: the thoughts might work better in italics or free indirect style with no tags at all.
The inventory is fixed and finite at nearly sixty verbs. Coined verbs, archaic verbs, and verbs outside that list do not count as tags. If you tag lines with an exotic verb the tool misses, your rate understates the problem. Read the inventory after each audit: if it reports fewer tags than you see on the page, the missing ones are verbs outside the list, and they are almost certainly worth cutting precisely because they are exotic. Quote style matters too: the counter blanks straight and curly double-quoted spans before counting, so paste with double quotes for the most accurate reading, since single-quote dialogue and em-dash styles get a rougher count.
Finally, remember what the tool cannot do. It cannot hear rhythm, cannot judge whether a colorful verb is earned, and cannot tell a deliberate stylistic choice from a tic. It assembles counts with fixed rules and states its conventions openly. Your judgment makes the final call. Use the numbers to find candidates, then revise by ear.
Genre-specific guidance
General fiction at 2.0 to 4.0 per 100 rewards balance. Keep said near two-thirds of tags, beats near half of units, and adverbials at zero. Most commercial scenes land here without strain once the first draft's colorful verbs consolidate. If you write across genres, calibrate on this band first, then adjust for the room you are entering.
Romance at 2.5 to 4.5 per 100 tolerates the most tags because banter timing depends on rapid attribution. Said plus asked can cover 80 percent or more of tags without monotony, since the voices carry variety. Reserve attitude verbs like teased, challenged, and protested for turning points rather than every volley. One murmured per scene can survive in intimate moments; three becomes parody. Beats should carry sensory detail: touch, proximity, breath, shared tasks. Let the higher tag ceiling fund pace, not decoration.
Mystery and thriller at 1.5 to 3.5 per 100 want lean attribution because pace depends on velocity. Interrogations and confrontations tempt writers into snapped, demanded, and insisted on every line, which slows the scene by grading conflict the lines already perform. Keep said dominant, use beats for procedural texture like files opened and doors checked, and spend colorful verbs only on genuine power shifts. Short lines need fewer tags than you think; the urgency of the exchange orients the reader.
Science fiction and fantasy at 2.0 to 4.0 per 100 face a special pressure: invented titles, ranks, and customs tempt writers into explained, reported, ordered, and declared as tags that smuggle exposition. Restructure those lines so the worldbuilding lives in beats and narration while tags stay quiet. Council scenes with many speakers may sit at the top of the band legitimately; fund the extra said tags by cutting every adverbial and every disclosure verb that merely restates the line.
Literary at 1.0 to 2.5 per 100 expects the most invisible attribution. Untagged lines, free indirect movement, and beats carrying psychological weight replace much of the tagging commercial scenes need. A literary passage at 1.7 per 100 with 80 percent beat share is exactly on target. Adverbials read loudest here: a single said wistfully can draw a workshop comment. Convert every graded tag to an image or gesture. Let said appear where orientation truly fails without it, and nowhere else.
Middle grade and young adult writers borrowing this page should aim at the top of general fiction or the romance band, since younger readers benefit from clearer attribution, but keep the same hierarchy: said first, beats second, colorful verbs rarely, adverbials never. Children's dialogue with a fresh verb every line exhausts young readers fastest of all.
A practical revision pass in four sweeps
Run this sequence on any scene the audit flags. It takes about fifteen minutes for a thousand words and converges reliably.
Sweep one removes adverbials. Search your scene for ly words near speech verbs, or simply revise every line the audit's adverbial count implicates. For each, either delete the adverb and keep said, or replace the whole tag with a beat. Do not swap one adverb for another. Softly to quietly is not revision. Either the line carries the tone alone or a beat stages it.
Sweep two consolidates verbs. Open the inventory and list every non-said verb appearing more than once. Change all but one occurrence of each to said or to no tag. Then list the one-offs and change at least half of them to said. Keep at most one or two colorful verbs per scene, placed where their meaning genuinely exceeds identification: a single snapped at the true breaking point, a single admitted where reluctance is plot information.
Sweep three deletes redundant tags. In two-person exchanges, remove tags from every line where the speaker is obvious, keeping orientation roughly every third line or wherever a beat already orients. Read the scene with the tags removed. Any line whose speaker you cannot name gets a tag or beat back. The rest stay clean. Expect to delete a quarter to a third of tags in a flagged scene.
Sweep four rebuilds beats. Wherever a tag came off, consider whether a beat should replace it. Not every gap needs filling; silence between lines can pace a scene well. Add beats where bodies have frozen, where setting has vanished for a page, or where emotion needs staging the line cannot perform. Prefer concrete actions over summarizing ones. She refilled both cups beats she was stalling nervously, and it gives the other speaker something to react to.
Then repaste the revised scene and reaudit. The per-100 rate should fall by one to three points, said share should rise past 60 percent, adverbials should read zero, and beat share should climb toward half. If the numbers move but the scene reads flat aloud, restore one beat of texture before cutting further. Numbers guide; the ear decides.
Frequently encountered confusions, answered at depth
Writers ask whether asked counts against them. It does count as a tag in the arithmetic, and rightly so, because it still orients. But it costs almost no attention on genuine questions, so a high asked count alongside said is healthy in interrogation or classroom scenes. The problem pattern is asked doing emotional work: demanded and questioned where asked would vanish. If your inventory shows demanded three times, try asked twice and a beat once before keeping any of them.
Writers ask whether whispered and shouted are acceptable scene-setting. They are, in strict moderation, because volume is information tags can legitimately add. A whispered conspiracy and a shouted warning each earn their verb once. The failure mode is volume as a permanent trait: characters who whisper every line for a chapter, or arguments where every line is shouted, snapped, or yelled. Vary volume through beats after establishing it once. She glanced at the sleeping child sets whisper-level without retagging every line.
Writers ask how to handle accents, stutters, and trailing lines without tags multiplying. The answer is to stage the speech pattern once in narration or a beat, then render subsequent lines cleanly. Tagging every stuttered line with stammered multiplies a moderate tic into a severe one. One beat establishes the pattern; readers sustain it across untagged lines. The same holds for interrupted lines: one interrupted plus an em dash teaches the rhythm, and further interruptions need no grading.
Writers ask about tag placement variety as a substitute for verb variety. Moving said to the front, middle, or end of lines does add rhythm without adding attention, and it is a legitimate tool. But placement variety cannot rescue a scene whose rate is double the band. Fix the count first with deletion, then adjust surviving tags for rhythm. A clean scene with all tags at line ends still reads better than a crowded scene with artfully distributed clutter.
Writers ask whether dialogue-only scenes on the page, like text exchanges or transcripts, follow the same bands. They do not, quite. Without narration there are no beats, so tags carry all orientation and rates legitimately run higher. The practical fix is to add beats rather than cut tags: timestamps, physical context, reactions between messages. If the form forbids beats, accept a higher rate and keep every tag said. The audit's bands assume conventional prose scenes; label your exception and move on.
Writers ask how this audit relates to filtering and filler-word checks. Tags overlap with filter words when verbs like wondered, recalled, and noted narrate cognition alongside speech. She wondered whether he would come is not dialogue attribution at all, yet the counter may tally wondered if the sentence carries dialogue punctuation. Read flagged lines in context. If the verb narrates thought rather than tagging speech, revise the sentence rather than blaming the tag count.
Where this work continues
Planning a cleaner scene is only half the job; executing the revision line by line is the rest. Do that revision work in Final Edit at /final-edit, where your draft can be tightened with the audit results beside it. Bring the three numbers this page gives you — your tags per 100, your said share, and your adverbial count — and work through the four sweeps above directly on the manuscript. The tool identifies the candidates; Final Edit is where each cut, swap, and rebuilt beat actually lands in your chapter.