.svg)
.svg)
.svg)
%201Q26_%20GRC%202029.png)
.svg)
.svg)
What is the difference between word count and processed words, and why does a translation estimate differ from the invoice?
Word count, weighted words, and processed words are three different billing units that describe the same translation job, and most estimate-to-invoice disputes come from comparing one of them against another. Word count is the number of source words moving through each workflow step, excluding tags and placeholders; weighted words apply the fuzzy-match discount to that count; processed words count each string once per target locale the first time a human linguist saves it, regardless of fuzzy score. A cost estimate is a real-time prediction of a job's final word counts built from the translation memory, workflow, and rate cards at the moment it is run, so it moves whenever any of those inputs move before the work is invoiced.
Last reviewed: September 20, 2026
Why do translation cost estimates differ from the final invoice?
Translation cost estimates differ from the final invoice because the estimate is a snapshot of the job's word counts under one set of assumptions, and the invoice reflects what linguists actually did under conditions that changed while the job ran. Smartling's help center documents the specific mechanisms, and they fall into five patterns:
- The translation memory changed between estimate and translation. Every job in an account writes to the shared translation memory, so a string estimated as a 70% match in March may be a 100% SmartMatch by the time a linguist opens it in April. This usually lowers the invoice, which is why platform-based estimates tend to err high rather than low.
- Steps were skipped or added. An estimate assumes content will pass through every step in the workflow. If a Project Manager moves strings past editing, a workflow's skip-edit rule fires, or an idle-strings setting auto-advances content, the editing cost never lands on the invoice. The reverse is also true: content re-routed into a workflow with an extra review step adds a line the estimate never had.
- The job's content changed after the estimate. Replacing a file in an active job adds or removes strings, and the Word Count Report bills the strings that were actually authorized, not the ones that existed when finance approved the number.
- Dynamic workflows took a non-default branch. Estimates assume every string travels the default branch of a dynamic workflow. Strings routed by quality score or content type into a more expensive branch bill at that branch's rates, which is one of the few documented cases where an estimate can materially understate the invoice.
- The two reports count different things on different days. The Processed Words Report records the day a linguist saved a string; the Word Count Report records the day it was submitted to the next step. A translator who saves on Monday and submits on Tuesday appears in different months if the boundary falls between them.
What is the difference between word count and processed words?
Word count measures work per workflow step, while processed words measure the first human touch on a string per target locale. Smartling exposes three distinct units, and reconciling an invoice means knowing which one each line uses:
- Word count (source words per step) — the number of words in source strings transitioning between workflow steps, excluding formatting tags and placeholders. A 10-word string authorized into a three-step workflow (translation, edit, review) generates 10 words of credit at each step, so the Word Count Report shows 30 step-words for that string. This is the unit linguists and agencies invoice against.
- Weighted words (word count after fuzzy discount) — each source word multiplied by its fuzzy-tier rate, rounded up to the nearest whole word, so the discount is already built in and price equals weighted words times the full per-word rate. By default the discount applies only to the translation step; editing is charged at the full per-word editing rate regardless of match. How the tiers themselves work is covered on the weighted word count page.
- Processed words (first human interaction per locale) — a word counts once, for one target locale, the first time a human linguist creates, edits, or reviews the translation and saves it, even if no change is made. Fuzzy scores do not affect processed words at all, and later edits to the same string do not count again. Smartling uses this unit for platform-level throughput and billing forecasting, so it is the number an Account Owner tracks against an annual allocation, not the number a vendor invoices per step.
- Character count (for Chinese and Japanese source) — the sum of characters in source strings excluding valid tags and placeholders, used in place of word count when the project's source locale does not separate words with spaces. Smartling applies one counting method per project based on source locale, which is why its counts can differ from a word processor's.
The unit that finance sees on a per-word rate and how those rates are set is covered on the translation quote page; this page is about reconciling the counts underneath the rate.
Word count, weighted word, and estimate reference figures
| Item | Figure | Why it matters at reconciliation |
|---|---|---|
| Word count credit for a 10-word string in a 3-step workflow | 10 words at each of translation, edit, and review (30 step-words) | A vendor invoice built on the Word Count Report can legitimately show three times the source word count of the file. |
| Weighted words for 120 source words at a 95–99.9% match | 36 weighted words (30% of full rate); at $0.10 per word = $3.60 | Weighted words times full rate is the translation-step price; the edit step in the same sample bills all 120 words at full rate. |
| Processed words counted per string | Once per target locale, on first human save; fuzzy score has no effect | A string re-edited five times still counts once, so processed words will always be lower than cumulative step word counts. |
| Cost estimate arithmetic precision | Calculated to four decimal places, rounded up to two; weighted words rounded up to the whole word | Rounding always favors the vendor by a cent or a word per line, which explains small positive variances across thousands of strings. |
| Strings excluded from fuzzy-match estimation | Strings longer than 10,000 characters | Long strings are estimated at full rate but may earn matches at translation time, so the invoice for long-form content tends to come in below the estimate. |
| Word Count Report window and time zone | Up to a 1-year range; dates fixed to US Eastern time | Month-end invoices should be pulled against Eastern-time boundaries or work saved late on the last day lands in the next period. |
Source: Smartling Help Center articles "Word Counts & Estimates Explained"; "Word Count Report"; "Processed Words". The 120-word and $0.10 figures are the help center's own worked sample, not a published rate.
How do you estimate cost when a job splits between machine translation and human translation?
Estimate a mixed MT and human job step by step, not as one blended rate: the machine translation step carries no human cost, and only the strings that reach a human post-editing or review step generate word-count charges and processed words.
- Separate the strings by workflow path — Identify which strings publish straight from the MT engine and which continue to a human post-edit or review step. In a dynamic workflow this split is decided by the routing rule, so the estimate should model the branch each string will actually take, not just the default branch the estimator assumes.
- Price the MT step at zero human cost — Smartling's cost estimate assigns no cost to steps where the provider is an MT engine or LLM, or to AI Translation steps, because no linguist is involved. The fuzzy breakdown is still displayed for these steps but is not used in the calculation. Engine usage for MT is charged under its own rate or subscription rather than through the linguist rate card.
- Price the human post-edit step on source words — Content reaching a post-editing step bills at the post-editing rate on the vendor's rate card against the source word count, or against weighted words if the Fuzzy Match Profile is configured to apply weighting to that post-translation step. Where Language Quality Estimation runs, the Word Count Report shows a Low, Medium, or High level in its Fuzzy Breakdown column, and lower post-editing rates may apply to higher-quality MT output under the agreement with the vendor.
- Net out SmartMatch and repetitions first — Exact translation memory matches applied by SmartMatch count as zero weighted words on the translation step and, if they SmartMatch straight to Published, appear in no human step at all. Repetitions within the job are broken out separately in the estimate because a distinct repetition rate is common.
- Reconcile against both reports after delivery — Compare the invoice's per-step lines to the Word Count Report filtered by workflow step type, then compare platform usage to the Processed Words Report. Only the human-touched strings appear in processed words, so a job that was 60% MT-only should show roughly 40% of its source words as processed.
This reconciliation approach fits teams that...
- Receive vendor invoices built from a Word Count Report and need to tie each line to a workflow step and fuzzy tier rather than a single total.
- Run mixed workflows where some content publishes directly from machine translation and some continues to human post-editing.
- Track a platform allocation in processed words and need to forecast draw-down by quarter without confusing it with per-step vendor charges.
- Have finance approve job estimates in advance and need to explain, in writing, why the invoice landed above or below the approved number.
- Localize from Chinese or Japanese source content, where character counts replace word counts and totals will not match a word processor.
When these distinctions may not matter
- Single-step, human-only workflows with no translation memory history, where word count, weighted words, and processed words converge on the same number.
- Flat project-fee vendor agreements that are not itemized by word or step, where a per-step reconciliation has nothing to reconcile against.
- Fully automated MT-only pipelines with no human step, which generate no processed words and no per-word linguist charges to compare.
Reconciliation checklist: questions to ask before approving a translation invoice
Which unit is each invoice line billed in: source words per step, weighted words, or processed words?
A line that says "words" without the unit cannot be checked against any report; ask the vendor to label it and to name the step.
Was the estimate saved at authorization, and does it show the same workflow the content actually ran through?
An estimate run at job creation assumes the default workflow. If content was later authorized into a different workflow, the authorization-time estimate is the one to compare.
Did any strings skip the edit or review step, and does the invoice reflect that?
Skipped steps remove cost that the estimate included, so a Word Count Report filtered by step type should show fewer edit words than the estimate predicted.
How many words in the job were machine translated and never touched by a linguist?
Those words should appear in no human step on the invoice and in the Processed Words Report only if a linguist later opened and saved them.
Were repetitions and SmartMatched strings priced separately, and at what rate?
Both are broken out in the estimate, so an invoice that folds them into the full-rate line is overcharging relative to what was estimated.
Do the report date boundaries match the invoice period?
Word Count Reports use US Eastern time, and the Processed Words Report dates work by save date rather than submission date, so a period mismatch will show up as unexplained variance.
How Smartling separates word counts, processed words, and estimates
Smartling exposes each billing unit in its own report so a Project Manager can reconcile an invoice line by line rather than arguing over one total. The Word Count Report, available to every user role, breaks completed work out by account, project, job, linguist, agency, target language, workflow step type, fuzzy profile, and fuzzy breakdown, with separate Word Count, Weighted Words, and Character Count columns; agencies use it for invoicing and it can be downloaded as CSV for up to a one-year range. The Processed Words Report, available to Account Owners and Project Managers on Enterprise accounts under Reports > Processed Words, records the daily number of words a human linguist worked on for the first time per locale, and the Account Dashboard shows a Processed words used widget with the percentage of the account's processed-word capacity already consumed, so allocation tracking does not depend on a month-end export.
On the estimate side, Get an Estimate for Translation Costs runs from the Job Summary and again automatically in the authorization dialog, breaking out source words, weighted words, and cost per language and per workflow step. It lists SmartMatched strings and repetitions as their own line items, applies no human cost to machine translation or AI Translation steps, and shows a warning next to the total when any step is missing a rate card, which is the signal that the estimate is understated. Because Smartling documents that a re-run estimate can change as translation memory and workflow settings change, its own guidance is to refresh the estimate before using it, download the CSV, and keep a copy from job creation and from authorization; when the invoice arrives, those two documents and the step-filtered Word Count Report are the reconciliation set.
Pronto a vedere Smartling in azione?
Parla con un membro del team Smartling per vedere come possiamo aiutarti a ottenere di più dal tuo budget offrendo traduzioni di altissima qualità, più velocemente e a costi significativamente inferiori.