Google Vision OCR vs Document AI: Which One Actually Reads a Receipt?

Last updated July 2026.

Google Cloud Vision API does OCR. Google Document AI does OCR plus structured parsing and entity extraction. In practice that means Vision hands you every word on a receipt and where it sat on the page, while Document AI's Expense parser hands you the merchant, the date, the tax, the total, and the line items. Vision has no receipt parser and no invoice parser. If you are trying to pull a total off a receipt with Vision, you are using a product that was never built to do it, and Google's own documentation says so.

What does Google Cloud Vision OCR actually return?

Text and coordinates. TEXT_DETECTION returns the full extracted string plus individual words with their bounding boxes. DOCUMENT_TEXT_DETECTION, the mode meant for dense documents, returns a richer structure organized into pages, blocks, paragraphs, words, and symbols, each with a bounding box.

Read that list again, because the thing that matters is what is missing from it. The hierarchy is geometric and typographic. It describes how the ink was arranged. There is no merchant key, no date key, no tax key, no total key, and no line items array, because Vision is not looking for those things. A receipt with nine dollar amounts on it produces nine numbers with coordinates, and deciding which one is the total is entirely your problem.

Does Google Cloud Vision have a receipt or invoice parser?

No. Vision's feature list is text detection, document text detection, landmark detection, logo detection, label detection, image properties, object localization, crop hints, web entities, SafeSearch, and face detection. There is no receipt parser, no invoice parser, no expense parser, and no field extraction of any kind anywhere in the product. This is the single fact that decides most Vision-versus-Document-AI evaluations, and it is the one people discover three days into a prototype.

What Google itself says about the difference

You do not have to take our word for it. Google's own Cloud Vision OCR documentation carries a notice telling readers that if they are detecting text in scanned documents, they should try Document AI for optical character recognition, structured form parsing, and entity extraction. Google is explicitly routing document-extraction users away from Vision and toward Document AI. When a vendor tells you its own product is the wrong one for your job, believe it.

Vision vs Document AI: pricing and output

These figures come from Google's own pricing pages in July 2026. Confirm before you commit, because cloud prices move.

The structure of that table is the whole story. Raw OCR costs $1.50 per 1,000 pages whether you buy it from Vision or from Document AI. The parsers that return named fields cost $10 per 1,000. That roughly seven-fold gap is not a charge for better character recognition. It is the price of meaning: you are paying for the model that knows a receipt has a total and an invoice has a due date.

Why the cheap option is usually the expensive one

A unit is one feature applied to one image, and each page of a PDF counts as one unit, so Vision looks like a bargain on a spreadsheet. The bargain evaporates the moment you write down what has to happen after the API responds.

With raw text, you write the layer that decides which number is the total rather than the subtotal, the tax, the tip, or the change due. You write the rules that find the transaction date among several dates. You write the logic that groups words into line items. Then you maintain all of it, because merchants redesign receipts and every redesign is a bug. Teams routinely spend weeks on this and end up with something that works on the merchants they tested and fails quietly on the ones they did not.

Eight dollars and fifty cents per thousand pages is what Document AI charges to skip that entire project. Measured against a single engineer-week, it is not a close call.

Does Google Cloud Vision or Document AI export to Excel or CSV?

Neither one does. Both return JSON, and there is no CSV or Excel output anywhere in either product. Vision adds a wrinkle for PDFs and TIFFs: the result is not even returned to you inline. Vision writes a JSON file into a Cloud Storage bucket that you have to create and own first. So the path from a scanned invoice to a spreadsheet runs through an image, a JSON file in a bucket, code you wrote to fetch it, code you wrote to interpret it, and finally a row. Four of those five steps are yours.

Is Google Cloud Vision free?

The first 1,000 units per feature per month are free, and that allowance recurs monthly rather than expiring like a trial credit. But you still need a Google Cloud project with billing enabled before the API will answer at all, so a payment method is required even to use the free units. Add the gcloud CLI, service-account credentials, and a client library, and the free tier is free in dollars and costly in hours.

When Vision is genuinely the right tool

Vision is excellent when text is the deliverable. Digitizing books, reading signs or license plates, transcribing handwriting, indexing scanned archives so they become searchable, or pulling text out of images in dozens of languages: Vision is fast, accurate, cheap at scale, and hard to beat. If your job is finished the moment the characters are recognized, it is a first-rate choice.

It is also a sound component when you already own the semantic layer. If you have built an extraction pipeline and only need strong OCR underneath it, Vision is well-priced infrastructure. The mistake is not using Vision. The mistake is expecting Vision to understand a document.

What to use if you just need receipt data in a spreadsheet

If nobody on the team is going to build and maintain a parsing layer, both Google products are the wrong shape, because both end in JSON inside a cloud project. Receipt OCR software reads receipts into labeled fields in a browser, and the receipt to Excel converter lands them in a sheet. For vendor bills, invoice OCR software pulls header fields and line items, and if the source is a PDF full of tables rather than a receipt, a dedicated PDF to Excel converter is the faster path. Large batches go through the bulk receipt scanner.

Developers who want parsed fields as JSON, without writing the layout heuristics themselves, can call the receipt OCR API or the invoice OCR API. If you are comparing clouds, the Google Vision OCR alternative and Google Document AI alternative pages go deeper on each, the Azure Document Intelligence alternative and AWS Textract alternative cover the other two big clouds, and the OCR software pillar maps the whole category.

The short version

Vision gives you words. Document AI gives you fields. Vision has no receipt parser and never has, Google's own docs point document users to Document AI, and the seven-fold price gap between raw OCR and the Expense parser is exactly the value of not building the parsing layer yourself. Choose Vision when text is the product. Choose Document AI when data is. Choose neither if you were only ever after a spreadsheet.