Send a broadsheet, circular or catalogue page to the API. Our vision model crops out every deal, however crowded the page, and each crop is then extracted on its own. You get back every deal as JSON: products, pack sizes, prices, savings and deal type, checked against retail rules.
Sample retailer page. "2 for $5" is the store's offer to its shoppers, read off the page. Our price is $0.04 per deal extracted.
Retailers publish new broadsheets every week, across regions, banners and channels. Brands, analysts, retail media teams and price-comparison sites need the deals in them as data, and most of it is still keyed in by hand.
Every deal on every page is read and keyed in by a person. Add a retailer, a region or an edition and the cost and the delay grow with it.
"2 for $7", a 6-pack priced each, a three-product bundle, a members-only price. One small misreading gives you the wrong price, unit or deal type.
One promotion runs in several regional editions, each with different artwork. Unless they're matched, it gets captured and counted more than once.
A broadsheet is usually valid for a week or two. If capture takes days, most of that window has passed before anyone can use the data.
Before a deal can be read, it has to be cut out of the page. On a broadsheet that's the hard part: deals are packed edge to edge, price flashes overlap the next product, and every retailer lays its pages out differently. We couldn't find anyone who did this reliably, so we trained our own model to do it.
Built through a long labelling and training effort on real retailer broadsheets. It draws a tight box around each deal, taking in the product, its price and its deal wording without pulling in the deal next door.
The extraction engine only ever sees one deal at a time, so a price can't be borrowed from a neighbouring deal and products from different tiles can't be merged. Every field is checked against retail rules for deal type, pack size and pricing before it's returned.
dealType: Combo3 productsprice: 240savingValue: 123Two stages do the work. A vision model trained on retail pages crops out each deal, then an extraction engine reads every crop separately and checks it against retail rules.
Send a broadsheet, circular or catalogue page image to the REST API. Multi-page catalogues can be submitted as one advert.
The vision model cuts each deal out of the page and returns its pixel coordinates, even when tiles touch and price flashes overlap.
Each crop is read on its own, so nothing is borrowed from the deal next door. Deal type, pack size and pricing are checked against retail rules before anything is returned.
Poll for status or fetch results in bulk. Each deal has a match key that stays the same across regional editions, and a link to its cropped image.
Crops from retailer adverts, and what the API returned for each. These are South African pages, so prices are in rand.
One crop comes back as four records that share a dealImageKey, each with its own pack and price.
4x1x250gBulkR183.998x1x125gBulkR199.9910x1x50gBulkR126.9915x1x20gBulkR96.99
Three packs of six, not three cartons: quantity 3, pack size 6. Mixed flavours are flagged as assorted.
3x6x200mlBulkR90assorted: true
Each product gets its own row with its own brand and size, all sharing the combo price.
1x1x2.5kgComboR99.991x1x2.5kgComboR99.991x1x2kgComboR99.99Every deal comes back in the same shape, ready for analytics, price monitoring and compliance work. These two records are real output from a SPAR broadsheet page.
{
"coordinates": {
"x": 432,
"y": 974,
"width": 405,
"height": 387
},
"products": [
{
"brand": "Spring Meadow",
"subBrand": "UHT",
"name": "Full Cream Milk",
"packSizeNotation": "1x6x1L",
"quantity": 1,
"packSize": 6,
"unitOfSize": 1,
"unitOfMeasure": "L",
"assorted": true
},
{
"brand": "Spring Meadow",
"subBrand": "UHT",
"name": "Low Fat Milk",
"packSizeNotation": "1x6x1L",
"quantity": 1,
"packSize": 6,
"unitOfSize": 1,
"unitOfMeasure": "L",
"assorted": true
}
],
"dealType": "Straight",
"description": "Spring Meadow UHT Full Cream or Low Fat Milk 6 x 1 litre",
"savingText": "Save R10",
"savingValue": 10,
"price": 89.99,
"loyalty": false,
"asSeenOnTv": false,
"discountPercentage": 0,
"contentHash": "eb035f19022a963e",
"dealImageKey": "9cd752bf-92aa-4d53-9801-30415db3a17e"
}{
"coordinates": {
"x": 78,
"y": 519,
"width": 885,
"height": 449
},
"products": [
{
"brand": "Allsome",
"subBrand": "",
"name": "Long Grain Parboiled Rice",
"packSizeNotation": "1x1x10kg",
"quantity": 1,
"packSize": 1,
"unitOfSize": 10,
"unitOfMeasure": "kg",
"assorted": false
},
{
"brand": "SPAR",
"subBrand": "",
"name": "Super Maize Meal",
"packSizeNotation": "1x1x10kg",
"quantity": 1,
"packSize": 1,
"unitOfSize": 10,
"unitOfMeasure": "kg",
"assorted": false
},
{
"brand": "SPAR",
"subBrand": "",
"name": "Sunflower Oil",
"packSizeNotation": "1x1x2L",
"quantity": 1,
"packSize": 1,
"unitOfSize": 2,
"unitOfMeasure": "L",
"assorted": false
}
],
"dealType": "Combo",
"description": "Combo deal for Allsome Long Grain Parboiled Rice, SPAR Super Maize Meal, and SPAR Sunflower Oil",
"savingText": "Save R123",
"savingValue": 123,
"price": 240,
"loyalty": false,
"asSeenOnTv": false,
"discountPercentage": 0,
"contentHash": "61ebe15ea801660d",
"dealImageKey": "68e7a9b8-e6f4-4b7e-a46f-932ab978a284"
}coordinatesdealImageKeycontentHashbrand · subBrandnameassortedpackSizeNotationquantity · packSizeunitOfSize · unitOfMeasuredealTypepricesavingText · savingValuediscountPercentageloyaltyasSeenOnTvdescriptionThe same wording always gets the same classification, whichever retailer, edition or market it comes from.
A single sellable item at a price. "Each" and "A or B" offers are Straight, and so is a multipack when the pack is physical.
Organic Whole Milk 6 x 1 L, $8.99 per pack
1x6x1LThe shopper has to buy more than one item: "2 for", "any 3", "buy 6". Ambiguous quantities default to Bulk.
Baked Beans 14 oz, 2 for $3
2x1x14ozA bundle of different products sold as one deal. Each product gets its own row, sharing one price.
Rice 5 kg + Pasta Sauce + Olive Oil, all for $15
3 products · 1 priceMix-and-match promotions where the lowest-priced item in the basket is free.
Buy any 3 haircare, cheapest free
MultibuyAlmost identical wording can mean one multipack or several separate items. Get it wrong and every unit price and comparison built on it is wrong too.
"Yogurt 6-pack 150 g, $4"
1x6x150g"6 yogurts 150 g for $4"
6x1x150g"Any 2 cola 6 x 330 ml packs for $9"
2x6x330mlAlongside the deals, each page returns the advert's key number, validity dates, promotional copy and legal text.
{
"keyNumber": "",
"termsAndConditions": "Maximum 10 deals per customer on all SAVE EXTRA products. Specials available at SPAR and SUPERSPAR country stores in KwaZulu-Natal from 7 October up to and including 18 October 2026, while stocks last. We reserve the right to limit quantities. TV products available nationally. Prices include VAT. E & OE.",
"advertText": "Super Savings. Unleash your inner chef with nutrifresh. Save up to 80% on nutrifresh Kitchen Essentials.",
"advertStartDate": "2026-10-07",
"advertEndDate": "2026-10-18",
"mediaUrls": [],
"contactNumbers": []
}Usually tiny print in a corner, and easy to miss. If the page doesn't carry one, the field is left empty rather than guessed.
Start and end dates in ISO format, read from wording like "prices valid through" or "offers end".
A competition or partner offer with its own end date isn't mistaken for the broadsheet's validity period.
Captured word for word and kept separate from the promotional copy, for compliance review.
Web addresses and phone numbers printed on the advert.
A general-purpose model can describe a broadsheet. It can't reliably crop a crowded page into separate deals, or tell "2 for $7" from a 6-pack priced each. We looked for a service that could, found none, and built Deal Extractor to do exactly that.
| Capability | OCR | General-purpose AI | Deal Extractor |
|---|---|---|---|
| Crops each deal cleanly on a crowded page | No | Unreliable | Purpose-trained model |
| Coordinates and a crop for each deal | No | No | Yes |
| Reads each deal from its own crop | No | Reads the whole page at once | Yes |
| Deal type and pack-size rules | No | Inconsistent | Checked on every deal |
| Output format | Raw text | Varies between runs | Fixed JSON schema |
| Matches a deal across regional editions | No | No | Yes |
| Advert validity dates | Raw text | Often picks the wrong date | Ignores side-promotion dates |
Submit page images, poll for status and collect JSON over HTTPS. We host and run the service, so there's nothing to deploy, no models to run and no prompts to tune.
No models, GPUs or ML team on your side. Your developers work with JSON over HTTPS.
Pages go through a background queue with automatic retries. Every result is validated, and failures alert us, not you.
Review extracted deals, track usage and cost, and export reports to Excel.
/Advertisement/SubmitAdvertisementImageSubmit a broadsheet page to Deal Extractor/Advertisement/SubmitAdvertisementMetaImagesSubmit one or more pages for advert metadata/Advertisement/GetImageStatusFetch status and extracted deals for an image/Advertisement/GetImageStatusBulkFetch results for many images in one call/Advertisement/ResubmitAdvertisementImageRe-run extraction on an imagePages with nothing usable, and deals that can't be read, aren't billed. There are no setup fees, platform fees or minimums.
One price
$0.04
per deal extracted
Set your monthly volume, and what manual capture costs you per deal today.
Deal Extractor
$400
per month at $0.04 a deal
Manual capture
$7,500
About $7,100 a month less than capturing by hand, 95% lower.
The manual figure assumes a few minutes of a data-entry operator's time per deal. Change it to your own number.
The detector was trained to recognise deals, not one retailer's design, so it works on grocery, pharmacy, DIY and general-merchandise broadsheets from any market. Prices come back as numbers in the currency printed on the page.
If you need specific retailers or languages covered, send us a few sample pages and we'll show you the output before you commit to anything.
Reading a whole page at once is where most mistakes come from: a price gets attached to the wrong product, two neighbouring deals are merged, or a small deal is missed. Cropping first means the extraction step only ever sees one deal.
Cropping a crowded broadsheet accurately is the hard part. We couldn't find a service that did it reliably, so we trained our own vision model on real retailer pages. It's the core of the product, and it keeps improving as we retrain on new layouts.
$0.04 for each deal successfully extracted. A page with eight deals costs $0.32. Pages with nothing usable, and deals that can't be read, are free. There are no setup fees, platform fees or minimums.
Each deal is extracted from its own crop, then checked against retail rules for deal type, pack size and pricing. Discount percentage, loyalty pricing and "as seen on TV" are only set when they're printed on the page.
Every record links to the image it was read from, so anyone on your team can check a deal against the artwork. We track quality in production and retrain as retailers change their layouts, and those improvements reach the API without any change on your side. The best test is your own pages: send us some and we'll show you the results.
Submit page images to the REST API, poll for status or fetch results in bulk, and receive JSON in a fixed schema. There's nothing to host, and most teams are integrated within days. The admin portal lets non-developers review deals, track usage and export reports to Excel.
Retailers run the same promotion in many regional editions, each with different artwork and layout, so comparing pixels doesn't work. We match on what the deal says instead. Each deal carries a content hash that is the same for the same offer wherever it appears, so you can remove duplicates across regions without extra work.
Happen Software, an engineering company working from Ireland and South Africa since 2016. Deal Extractor already processes real retailer broadsheets, through the same API and portal you would integrate with.
Tell us which broadsheets you work with and roughly how many deals a month. One of our engineers will run a sample of your pages, show you the results, and help with the integration if you go ahead.
Send us a few pages. We'll run them and send back the extracted deals, with no commitment.