Per-Document Pricing: Examples & Companies

What is it

Per-Document Pricing is a billing unit where customers are charged per document processed or generated — common in AI writing, SEO, and document-intelligence tools. It is the most deliverable-shaped unit in usage pricing: the thing being counted is the thing the buyer actually wanted — a finished article, a processed page, a translated contract, or a knowledge artifact — which makes the bill legible in a way tokens and credits rarely are.

The catch is that “document” is not one unit but four, and the meaning sets the price. To a writing tool it is a deliverable (Byword, Surfer SEO, Scalenut, Frase, Writesonic — monthly article allowances). To a knowledge platform it is capacity — a stock you keep indexed (Dify, Nomic, Mem, Sana AI). To a processing engine it is throughput — files run through a pipeline (Unstructured, DeepL, Unbabel). To a legal-AI vendor it is a professional outcome (EvenUp, Ironclad AI). Deciding which of the four a given product actually sells is the core judgment the right-usage-metric guide walks through.

How it works

The standard structure is a monthly allowance with optional overage: bill = tier fee + max(0, documents − included) × per-document rate. The levers:

Worked example — document as deliverable. On Byword, the Standard plan is $299/mo for 80 included articles, an effective $3.74 per article if you use them all; articles beyond the allotment bill at $3.00 each, and unused credits accumulate. A team producing 120 articles pays $299 + (120 − 80) × $3.00 = $419 that month. Compare that to Byword’s $1,999/mo Unlimited plan, which removes the meter entirely and lets you plug in your own Claude and Gemini keys at roughly $0.10 per article of raw model cost — a completely different economic object at high volume.

Worked example — document as processing throughput. On Unstructured, the meter is $0.03 per page for any file type and any pipeline (Fast, Hi-Res, VLM, or Auto), after a one-time 15,000-page free allowance. Processing 100,000 pages in a month costs $0.03 × 100,000 = $3,000, full stop — no seats, no platform fee. The hidden multiplier here is not tier but reprocessing: re-running a corpus after a chunking-strategy change pays the per-page rate again.

Worked example — document as capacity. On Dify, documents don’t bill individually at all: the $59 Professional workspace includes 500 knowledge documents as standing capacity alongside 5,000 message credits and 5GB of storage. The unit is the same noun, but it meters a stock (what you can keep indexed) rather than a flow (what you produce this month).

Companies using this

14 in-corpus companies meter documents, split across the four meanings above: writing/SEO (Byword, Surfer SEO, Scalenut, Frase, Writesonic), document processing (Unstructured, DeepL, Unbabel), legal AI (EvenUp, Ironclad AI), and knowledge infrastructure (Dify, Nomic, Mem, Sana AI). The table below sorts by pricing model, billing units, and free-tier availability.

Patterns observed

The deliverable cluster prices countability and hides variance. One article is one unit regardless of length, so vendors trade size-sensitivity for a number buyers can verify themselves and absorb long-document cost inside the allowance. That size-blindness is the writing cluster’s defining choice — and the reason the tier grid, not the unit, does the real price discrimination.

The value metric keeps moving under the same noun. Frase has re-cut its document meter four times — document credits (2020) → articles (2023) → search queries (2024) → SEO documents (2025) — and even shipped pure pay-as-you-go “Rank-Ready AI Documents” at $3.50 each (dropping to $1.67 at volume) for a few months in 2025 before folding it back into subscriptions. Scalenut has cycled through four packagings in five years, and Writesonic migrated from generation credits to words to credits to GEO coverage. The document survives every rewrite as the headline unit because it is the one thing a content buyer can count — but what it represents drifts with the market, and by 2026 all three had pushed the surrounding meters toward tracked AI-visibility prompts.

Discount through the denominator, soften the cap at the top. Scalenut’s standing annual promo doubles the article quota at the same sticker (5 → 10 GEO articles/mo) — repricing through the allowance instead of the price. And at the top of the ladder the ceiling goes soft: Surfer SEO’s “unlimited*” documents on Peace of Mind convert the meter into a fair-use promise exactly where heavy users would otherwise churn to overage-free competitors.

Legal documents run on a different logic. EvenUp and Ironclad AI price the document as a professional-services artifact, not a content asset — the rate reflects value delivered (attorney hours displaced, contract risk reduced) rather than compute consumed. That is why their per-document numbers sit orders of magnitude above the writing cluster, and why both are organized around the matter, case, or contract rather than the page.

Counterexamples & variants

DeepL shows per-document pricing coexisting with a finer meter inside one product. Its Pro apps bill per translated document folded into per-seat tiers — Individual at $8.74/user/mo (3 file translations/mo), Team at $28.74 (20/mo), Business at $57.49 (100/mo) — while its API is pure per-character ($5.49/mo base + $25.00 per 1,000,000 characters). A user translating thousands of short snippets prefers character billing; one translating complete contracts prefers document billing. “Per-document” here is a UX simplification that turns an opaque character count into a countable thing buyers already grasp. Unbabel sits at the opaque extreme: its enterprise LangOps translation is quote-only, while its self-serve Widn.ai sibling caps file uploads per month (5 Free, 25 on the $19 Plus, 300 on the $90 Pro) rather than pricing per document at all.

Nomic buries the unit deeper still. Documents are what its platform embeds and maps, but the AEC Platform bill is structured around seats ($40/user/mo, 25-seat minimum, $1,000/mo commitment) plus a pooled $20-per-seat AI-usage balance shared org-wide. The document count governs how much embedding capacity a team consumes but never surfaces as a billing line — proof that “per-document” can name a workload type without being the invoice unit. (Only Nomic’s separate Atlas app exposes a per-token embedding rate, $1 per 10M text tokens.)

Dify is the capacity variant worth naming explicitly: when documents meter a knowledge base rather than a pipeline, the right comparison set is vector-storage pricing, not article allowances. A 500-document workspace is not producing 500 documents a month — it is maintaining a base that holds 500, a standing limit rather than a consumption meter. Mem makes the same move at the consumer end: its 25-note / 25-PDF-page caps exist only on the $0 Free tier and vanish on the flat $12/mo Pro plan, so the meter is a free-tier throttle, not a paid-tier charge.

Unstructured flattens the model in the opposite direction. Most document-AI APIs charge more for higher-fidelity processing — premium OCR, layout models, VLM parsing. Unstructured collapses every strategy into one flat $0.03/page, so the bill is a pure function of pages pushed, not which model touched them — the exact inverse of the writing cluster’s size-blindness. Together with EvenUp’s ~$300+ legal demand letters (reported $500–$800+ once add-ons stacked), it brackets the roughly 100,000× per-document price range this one unit spans across the corpus — the gap between counting compute and counting value.

What this means for buyers vs vendors

For buyers

First identify which of the four meanings applies, because the buying question changes with it. For the deliverable cluster, compute the effective per-document price at your volume — allowance fee ÷ documents actually used, plus marginal overage — rather than comparing stickers; check rollover (Byword’s credits accumulate; most expire), check what one document includes, and treat “unlimited” tiers as fair-use contracts to read, not ceilings to ignore. The introduction to usage-based pricing walks through the overage-crossover arithmetic.

For legal documents, measure the rate against the alternative — attorney hourly rates, contract-review labor, the cost of not pursuing a claim — not against other document tools. EvenUp’s December 2025 shift from per-document to case-based pricing is the signal to watch: when per-artifact charges accumulate into an unpredictable per-case total, the vendor itself moves the unit, so judge whether the price fits the value per matter in your workflow. For knowledge infrastructure (Dify, Nomic, Sana AI, Mem), the document limit is a capacity constraint: model whether it fits your base twelve months out, since Dify’s 50 → 500 → 1,000 ladder and Sana AI’s 1,000 → 10,000 jump both create upgrade pressure as the corpus grows. Treat these like vector-storage pricing.

For vendors

The document is trusted because buyers count their own output — protect that by keeping one document = one unit and absorbing variance in the allowance price, not surprise multipliers. Publish the overage rate and let it fall with tier the way Byword does ($3.50 → $2.50 across its ladder); the prepaid-credits guide covers why visible marginal pricing beats opaque packs. If your documents are stocks not flows, price capacity explicitly instead of dressing storage limits as usage meters. And when the buyer’s value metric moves — as it did from articles toward AI-visibility coverage across the SEO cluster in 2025–2026 — move the headline unit before a competitor does.

For legal platforms, per-document works because the professional value per document is high and understood: Ironclad AI justifies a five-figure platform fee (reported $40,000–$80,000/yr for small/mid-market) against contract-review labor, not content-generation cost. The durable move is to align the billing unit with the buyer’s budgeting unit — case, matter, or contract. And for translation, offer both denominators as DeepL does: file-thinking human users get per-document, snippet-processing developers get per-character, and neither is forced into a count that obscures their real cost structure.

Explore this theme in the knowledge graph →