Industries / Finance and business services / Advertising and marketing / Market research
Tracking sales
Sales tracking counts what people actually buy, from shop tills, shop audits and households that record their purchases.
5
Milestones achieved, from 1933 to 2023.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1933
Counting sales in shops
The A.C. Nielsen Company of Chicago launched an index of drug store sales in 1933 and food and department store indexes in 1934. Nielsen's own history credits Arthur C. Nielsen with the idea of "market share," a brand's portion of a category's sales, in 1935, when its auditors checked shop shelves to work out sales patterns.
Credit
Arthur C. Nielsen, A.C. Nielsen Company
1974
The first scanned barcode
On 26 June 1974 cashier Sharon Buchanan scanned a 10-pack of Wrigley's Juicy Fruit gum, priced at 67 cents, at a Marsh Supermarket in Troy, Ohio, near NCR's scanner development site; Clyde Dawson, Marsh's head of research and development, was the buyer. Checkout scans now feed sales tracking: NIQ says it receives point-of-sale data from more than 8,900 retailers in 88 countries.
Credit
Marsh Supermarkets and NCR; the barcode standard is now run by GS1
2001
Walmart stops sharing sales
In May 2001 Wal-Mart Stores said it would stop sharing its sales information with ACNielsen and five other data-tracking firms, to keep it out of the hands of competitors. In July 2011 it agreed to share its US sales data, including Sam's Club, with Nielsen "for the first time in 10 years."
LessonSales tracking rests on retailers' consent, and one large refusal left a gap in the data for a decade.
Credit
Walmart; Nielsen
2009
Home scanners checked
Households in Nielsen's Homescan panel record their shopping with a scanner at home. Comparing those records with a large supermarket chain's own checkout data, researchers found no matching store transaction for about 20% of recorded trips, and on matched trips about 20% of items went unrecorded, though quantities matched 94% of the time.
LessonHousehold panels miss snacks eaten on the way home and items from big shopping trips, and the misses vary by type of household, so the gaps are not random.
Credit
Liran Einav, Ephraim Leibtag and Aviv Nevo, for the US Department of Agriculture's Economic Research Service (2008 study, summarized 2009)
2023
The big trackers merge
NielsenIQ (NIQ), which grew from A.C. Nielsen's business, completed its combination with GfK on 10 July 2023; IRI and NPD, merged in August 2022, took the name Circana in March 2023; and in January 2025 Kantar folded Worldpanel, its consumer-panel business in 49 countries outside North America, into Numerator. NIQ's annual report lists consumer panels of more than 5.5 million people in more than 50 countries, and more than 12,000 field auditors in 80 countries for small and local shops.
Credit
NIQ and GfK; IRI and NPD; Kantar
Watch for
Sales through AI shopping agents
NIQ set up a Commerce Lab in April 2026 to build measurement for buying done through AI assistants. In September 2026 it announced work with Similarweb, a web-traffic data firm, to connect AI-driven product discovery to "measured sales outcomes," with a first version due in the fourth quarter of 2026, "beginning with a focused set of categories and markets."
Achieved whenthe product ships and its sales figures are checked by someone outside NIQ, such as a retailer comparing them with its own records.
Claimed by
NIQ, April and September 2026
Open
01
Decode receipt lines into products
A public test set of photographed till receipts from several chains and countries, with each line labeled by brand, product, size and category, and a scored comparison of AI vision models at turning the shortened text on each line into the product bought.
The briefCredit · unclaimed02
Test AI product coding
A public test of tools that assign products to GS1 Global Product Classification "bricks," the standard category codes used in product data, scored on whether the right code comes first or in the top three, against labels set by two people.
The briefCredit · unclaimed
Open 01 · the brief
Decode receipt lines into products
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test set of photographed till receipts from several chains and countries, with each line labeled by brand, product, size and category, and a scored comparison of AI vision models at turning the shortened text on each line into the product bought.
Why it is still open
Public receipt datasets label other things. The ICDAR 2019 SROIE competition set covers reading receipt text and extracting "key texts" such as the company, date, address and total, and CORD, a set of Indonesian receipts, labels lines for parsing; neither links a line to a product. Trackers collect receipts commercially, as NIQ's annual report shows, but no public test of decoding them was found.
Built with what exists
Receipts donated by volunteers from their own shopping, with card numbers and loyalty IDs blacked out; Open Food Facts, an open database of food products and their barcodes; GS1's Global Product Classification (GPC) for categories; AI vision models such as Claude, Gemini and Qwen-VL; Label Studio for labeling.
First step
Collect 100 of the team's own receipts from five chains, label every line, and run two models on them.
Done looks like
At least 2,000 receipts and 20,000 lines from at least three countries, with each model's accuracy for brand, category and exact product, broken down by chain, and the dataset released.
Open 02 · the brief
Test AI product coding
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test of tools that assign products to GS1 Global Product Classification "bricks," the standard category codes used in product data, scored on whether the right code comes first or in the top three, against labels set by two people.
Why it is still open
GS1 Belgium & Luxembourg offers a GPC Predictor that suggests the three best-matching codes from a description or photo, but states that it "is a support tool and may contain errors," and its page gives no accuracy figure. Searches found no independent accuracy test.
Built with what exists
GS1's published GPC schema; product photos and descriptions from Open Food Facts; the GPC Predictor and general AI models; Label Studio.
First step
Download the current GPC schema and label 200 grocery products by hand.
Done looks like
At least 2,000 products from at least 30 GPC segments, labeled by two people with their agreement reported, and top-1 and top-3 accuracy for each tool.