OCR, short for optical character recognition, is software that finds text inside an image and converts it into characters a computer can search, store and classify. A shop sign, a scanned invoice and a banner with a price painted across it all hold words that exist only as pixels. The technology exists because most automated systems in advertising, from keyword classifiers to policy filters, reason about text, while a growing share of the material they must judge arrives as pictures.

In advertising the term surfaces in four places: creative review, brand safety and contextual classification, commerce data built from receipt scans, and visual search. Nielsen also used the letters for a measurement product until 2015.

How the pipeline works

A recognition job runs in stages. According to Wikipedia's overview of the field, pre-processing usually includes de-skewing a tilted scan, despeckling and binarisation, which reduces a colour or greyscale image to black and white so that text separates from background. Layout analysis then identifies columns, paragraphs and captions as distinct blocks, followed by line and word detection.

Recognition follows one of two classic methods. Matrix matching compares each isolated glyph pixel by pixel against stored shapes and struggles with unfamiliar fonts. Feature extraction breaks glyphs into lines, loops and intersections and compares those features with abstract character prototypes, the approach behind most modern software. Newer engines skip single characters altogether: version 4 of Tesseract added a neural network engine based on long short-term memory (LSTM) that reads whole lines.

Post-processing repairs the output. Constraining results to a lexicon raises accuracy but can mangle proper nouns missing from the word list. Common output formats include ALTO, an XML schema maintained by the US Library of Congress, as well as hOCR and PAGE XML.

Cloud services sell the pipeline as an API. Google's Cloud Vision offers two OCR features. TEXT_DETECTION handles sparse text such as a road sign and returns the full string plus individual words with bounding boxes. DOCUMENT_TEXT_DETECTION is tuned for dense text and returns a hierarchy of pages, blocks, paragraphs, words and symbols. Billing is per feature per image, with the first 1,000 units each month free, according to Google's pricing page; third-party listings put text detection at $1.50 per 1,000 units between 1,001 and 5 million. Synchronous calls accept up to 16 images and asynchronous batches up to 2,000. At that rate, reading the text in one million creatives costs roughly $1,500 before storage and compute, cheap enough to run by default inside review systems.

Where it sits in the advertising chain

No OpenRTB field carries raw extracted text. OCR output travels as derived labels instead: content categories, suitability scores, policy decisions and audience segments, all computed before or after the auction.

On the buy side, creative review is the heaviest user. Search image assets in Google Ads sit under a long-standing ban on text overlays, and in July 2026 Google wrote an exemption for AI disclosure labels into that policy, a rule that presupposes systems able to detect lettering placed on a picture. Meta enforced a stricter version for years. Images with more than 20% text were rejected or had delivery restricted, and the checking tool, according to Instapage, divided each image into a grid of 25 rectangles. Facebook dropped the rule in September 2020, according to Search Engine Journal.

Publishers carry a mirror obligation. AdSense for Search rules effective November 1, 2025 require verbatim transcription of image text, video text and visible street signs from upstream creatives that send traffic to Related Search for Content pages.

Verification and contextual vendors apply the technique to the content surrounding an ad. Integral Ad Science listed text extraction from images among the methods behind its TikTok measurement, next to speech-to-text and frame-by-frame scoring. DoubleVerify's streaming product with NBCUniversal was presented as using computer vision and OCR to identify what appears in a programme. Infrastructure providers now offer it at the stream level: Amazon Web Services' Contextual Metadata runs vision models alongside encoding to flag objects, brands and on-screen text in live video, mapping results to the IAB content taxonomy. The common thread, as the PPC Land reference on sentiment notes, is that classification starts by converting content into text.

Commerce data depends on reading printed till slips. Fetch, whose receipt data Media.net routed into open web targeting and measurement on April 23, 2026, processes more than 13 million receipts a day. Perion reached the same data through a LiveRamp integration.

Visual search completes the list. A Lens query contains no typed text at all, so any words must be derived from the image. Google added Shopping ads to Lens on October 3, 2024, putting monthly visual searches near 20 billion.

Origin and evolution

The idea predates computers. In 1914 Emanuel Goldberg built a machine that read characters and converted them into telegraph code, while Edmund Fournier d'Albe developed the Optophone, a handheld scanner producing tones for letters. Ray Kurzweil founded Kurzweil Computer Products in 1974 to pursue omni-font recognition, unveiled a reading machine for blind people on January 13, 1976, and began selling software in 1978, with LexisNexis among the first customers.

Tesseract, the most widely used open engine, was developed at Hewlett-Packard in Bristol and Greeley, Colorado, between 1985 and 1994. HP open-sourced it in 2005, Google developed it from 2006 until August 2017, and version 5.0.0 arrived on November 30, 2021, with support for more than 100 languages. The CAPTCHA era added a twist: reCAPTCHA, spun out in 2007, routed human answers into book digitisation, using people to read words that machines had failed on.

Large multimodal models changed the architecture again. Speaking on a Google podcast in July 2025, a Gemini team member said separate OCR models had been bundled into Gemini, removing the old step of extracting text before feeding it to a model. Microsoft had already folded OCR into its MarkItDown conversion library by December 2024. Mistral launched an OCR API in 2025 priced at 1,000 pages per dollar, and DeepSeek published DeepSeek-OCR in October 2025, a model trained on nearly 100 languages that reached general availability on Google's Vertex AI on October 23, 2025.

Why it matters for marketers

Standards made it harder to avoid. On October 18, 2025, the Media Rating Council restricted the phrase brand safety to vendors analysing images, video and audio at content level, with a grace period that, according to PPC Land's explainer on the council, ended on April 18, 2026. Words burned into a thumbnail or overlaid on a video are invisible to keyword-only tools, and OCR is among the cheapest ways to expose them.

Cost is the other lever. Frame-level classification is priced by how much it costs to look at each frame, which is why Google's move to cut Gemini video analysis tokens by up to 88% matters for contextual vendors. Enforcement is scaling as well: Google's 2025 Ads Safety Report credited Gemini-based review as its main response to generated scam creative.

Limitations and disputes

Accuracy varies widely. A 2009 study of historic newspaper digitisation, cited by Wikipedia, found character accuracy between 81% and 99% across commercial software. A 1% character error rate can become a word error rate of 5% or worse. Advertising creative, with stylised type, gradients and motion, is harder than a clean scan.

Published benchmarks are frequently self-reported. Mistral's launch figures, which placed its model ahead of Google Document AI and Azure, came from the company's own tests. AWS, for its part, has published no accuracy or error rates for its live-video models. Enforcement tools have drawn criticism too: Instapage argued that Meta's grid responded to where text sat more than to how much there was.

OCR also cuts both ways. The same capability that reads creative has long been used to defeat text-based CAPTCHA challenges. Commercial ownership raises further questions: Publicis agreed in May 2026 to acquire LiveRamp for $2.5 billion, the layer through which some receipt-derived signals now flow.

Not the same as

Nielsen OCR. Online Campaign Ratings, announced by Nielsen in 2010, was tested in the fourth quarter of that year with Procter & Gamble, Verizon Wireless, Starcom MediaVest and Facebook, combining Nielsen panels with demographic data from online providers. It was renamed Digital Ad Ratings on April 1, 2015, the product that Google later integrated into Ad Manager for OTT inventory.

OCR at HHS. In US health privacy, the letters stand for the Office for Civil Rights, the enforcer of HIPAA that warned roughly 130 health organisations about tracking technologies in July 2023.

Automatic content recognition (ACR). ACR fingerprints audio or video and matches it against a reference library to identify what a television shows. It recognises content, not characters.

ICR and OMR. Intelligent character recognition applies machine learning to handwriting; optical mark recognition, according to IBM, detects ticked boxes and filled bubbles.

Recent developments

On September 14, 2026, Snipp Interactive opened a ChatGPT plugin and a Claude connector that expose receipt OCR parsing and SKU-level validation inside AI conversations. AWS's Contextual Metadata, covered this month, ties on-screen text detection to live ad decisioning, and Google's July 2026 policy change means text detection on Search image assets must now tell a disclosure label from a promotional overlay. Extraction is moving from standalone utility into general models and live pipelines, while independent accuracy data remains scarce.

Timeline

  • 1914: Emanuel Goldberg builds a character-reading machine; Edmund Fournier d'Albe develops the Optophone
  • December 29, 1931: Goldberg granted US Patent 1,838,389 for his "Statistical Machine"
  • 1974: Ray Kurzweil founds Kurzweil Computer Products
  • January 13, 1976: Kurzweil reading machine unveiled with the National Federation of the Blind
  • 1978: Kurzweil begins selling commercial OCR software
  • 1985 to 1994: Tesseract developed at Hewlett-Packard
  • June 1993: Unicode 1.1 adds OCR characters
  • 2005: HP open-sources Tesseract
  • 2006: Google takes over Tesseract development
  • 2007: reCAPTCHA spun out, feeding answers into digitisation
  • 2010: Nielsen announces Online Campaign Ratings, tested in Q4 2010
  • April 1, 2015: Nielsen renames Online Campaign Ratings to Digital Ad Ratings
  • August 2017: Google's lead development of Tesseract ends
  • September 2020: Facebook drops its 20% image text rule
  • November 30, 2021: Tesseract 5.0.0 released
  • October 3, 2024: Google adds Shopping ads to Lens
  • 2025: Mistral launches an OCR API at 1,000 pages per dollar
  • May 14, 2025: Mistral OCR 25.05 released on Vertex AI
  • October 18, 2025: MRC restricts brand safety claims to content-level analysis
  • October 23, 2025: DeepSeek-OCR reaches general availability on Vertex AI
  • November 1, 2025: AdSense for Search verbatim creative text rules take effect
  • April 18, 2026: MRC grace period ends
  • April 23, 2026: Media.net announces Fetch receipt data partnership
  • July 9, 2026: Google changelog on AI labelling exempts disclosure labels from the image text overlay ban
  • September 14, 2026: Snipp brings receipt OCR parsing into ChatGPT and Claude

Summary

Who. Ad platforms run OCR in creative review; verification and contextual vendors such as Integral Ad Science and DoubleVerify apply it to content; cloud providers including Google, AWS, Microsoft, Mistral and DeepSeek supply engines; receipt platforms such as Fetch and Snipp turn it into purchase data.

What. Optical character recognition converts text held in images, scans and video frames into machine-readable characters, returned with positions and confidence scores.

When. Character-reading machines date to 1914, commercial software to 1978 and the open Tesseract engine to 2005. Multimodal models absorbed the task from 2024 onwards, while the MRC's October 2025 policy raised demand for content-level analysis.

Where. Outside the bid stream: in pre-serve creative review, pre-bid contextual classification, post-bid verification, receipt-based measurement and visual search.

Why. Most automated advertising systems reason about text, while much of what they must judge arrives as pixels. OCR bridges that gap, with accuracy that remains uneven and largely self-reported.