1.1. UK policy

The growing extent of nontariff barriers (NTBs) present at the GB-European Union (EU) border in recent years has been shown to have increased prices (Bakker et al., Reference Bakker2022). Our dataset extends this analysis, providing a process that allows for real-time monitoring.

The EU remains Britain’s most important trade partner. Between 1999 and 2007, the EU accounted for over half of UK exports and imports. Although shares have fallen somewhat, the EU still accounts for 42% of exports and 48% of imports. The latter channel—£432 billion of imports—is one conduit through which trade shocks are transmitted. For example, the depreciation of sterling following the EU Referendum is estimated to have increased consumer prices by 2.9%, costing the average household £870 per year (Breinlich et al., Reference Breinlich2022). EU imports are also the main channel for inflationary pressure from NTBs.

NTBs will rise in 2024. The UK Government’s Border Target Operating Model, published in October 2023, outlines a phased introduction of full controls on imports from the EU to GB (Cabinet Office, 2024). This plan is divided into three key phases as follows:

1. a. 31 January 2024: Health certifications required for EU imports of medium-risk animal products, plants, plant products (e.g., meat, dairy, some fruits and vegetables) and high-risk nonanimal food and feed.

2. b. 30 April 2024: Initiation of risk-based checks (documentary, identity and physical) on imports of medium-risk goods from the EU mentioned above. High-risk plant inspections will shift from destination to Border Control Posts.

3. c. 31 October 2024: Enforcement of Safety and Security declarations for all imports from the EU and applicable territories.

The Cabinet Office guidance suggests that the inflationary impact will be minimal. The additional controls on plant and animal products are estimated to cost businesses £330 million annually, increasing food prices by 0.2% over 3 years (Cabinet Office, 2024). The analysis suggests that costs will depend closely on how businesses adapt their operations and supply chains to integrate the new control regimes. Internal modelling suggests that the impact on food prices ‘will not be significant’. Our paper provides a way to verify these cost estimates.

1.2. Literature

Tariffs have been the traditional trade barrier used by countries seeking to protect domestic producers and are still in widespread use today. The United States, for example, imposed new tariffs in 2018, which targeted approximately $283 billion of American imports, principally from China. Recent studies find high pass-through rates of tariffs to consumer prices. Flaaen et al. (Reference Flaaen, Hortaçsu and Tintelnot2020) use weekly microdata from retail stores to examine the impact: tariffs on washing machines were found to have a pass-through rate of between 1.08 and 2.25. Similarly, Fajgelbaum et al. (Reference Fajgelbaum2020) find evidence for complete pass-through of tariffs, along with a large decline in imports of targeted products and a significant loss in real incomes. If the impact of NTBs is similar to tariffs, there are reasons to be concerned about the impact on UK households.

For many countries, NTBs have eclipsed tariffs as the primary barrier to trade. The estimated indices for trade restrictiveness show that NTBs are almost as restrictive as tariffs (Looi Kee et al., Reference Looi Kee, Nicita and Olarreaga2009). This means low tariff rates can give a false picture of ‘free’ trade. In 2018, the EU had an average ad valorem equivalent NTB of 13.1%, contrasting an average tariff rate of just 1.8% (Bakker et al., Reference Bakker2022). These differences mean that while tariff rates may be falling, trade protection may be rising due to NTBs: Niu et al. (Reference Niu2018) estimate ad valorem equivalents of NTBs between 1997 and 2015, confirming this concern.

Recognising this, many trade agreements aim to offer a ‘deep’ eradication of barriers. Deep trade agreements go beyond tariffs and quotas and include provisions that reduce NTBs. This can include the mutual recognition of health certifications used for livestock, for example. Dhingra et al. (Reference Dhingra, Freeman and Huang2023) measure the welfare effects of such agreements. They find a substantial impact on trade and argue that agreements that reduce NTBs improve welfare.

A detailed analysis of the 2021 TCA is provided by Freeman et al. (Reference Freeman2022), who examine UK-EU trade relative to UK trade with the rest of the world using a difference-in-differences event study. They reason that uncertainty and anticipation effects could mean trade impacts may predate new rules being imposed, but they find no evidence of such ‘forward looking’ impacts between 2016 and 2021. Following the TCA, however, they find a sudden and persistent fall in relative UK imports from the EU, estimated at 25%. In contrast, they find a smaller and temporary decline in relative UK exports to the EU, but a large and sustained drop in the number of firms exporting. Low-value trade relationships may have been made unviable, the authors argue.

The quantification of NTB pass-through to consumer prices set out by Bakker et al. (Reference Bakker2022, 2023) is the closest paper to ours. By matching ONS price microdata with UN data on bilateral trade flows, they are able to estimate a measure of EU trade exposure for individual food items. The authors find that since the EU referendum, UK food products that are more heavily imported from the EU have seen larger price increases. Using a difference-in-differences event study, they find that price increases are driven by food products with high NTBs, with an implied pass-through rate of 50–88%. The cost of this, estimated between 2019 and 2023, is almost £7 billion across UK households. These costs are likely to amplify preexisting inequality in the United Kingdom, with households in the bottom income decile suffering cost of living increases that are 52% more than those in the top decile.

2.1. Protected names

Many food and drink product names provide information on their origin. A geographical indication (GI) is a ‘distinctive sign used to identify a product whose quality, reputation, or other such characteristics relate to its geographical origin’ (European Commission, 2023). Within this umbrella categorisation, there are three systems holding varying degrees of protection.Footnote ² We utilise the strongest categorisation: Protected Designation of Origin (PDO). This covers food and wine products, specifying that each stage of production, processing and preparation must occur in a specific region.

PDO products have strong links to their place of origin. For instance, “Champagne” from France, “Prosciutto di Parma” from Italy and “Stilton Cheese” from the United Kingdom are all emblematic PDO products that consumers associate with authenticity and region-specific character. For most products sold in supermarkets, retailers can substitute goods sourced from one country with goods sold in another. For example, a retailer can switch suppliers for vegetables in response to a change in NTBs. In contrast, as PDO products’ character is defined by their origin, their geographic substitutability is low, and their exposure to NTBs is potentially high.

To match products from our dataset to their PDO counterparts, we use the European Commission’s PDO register. By searching product titles for combinations of protected terms, as defined in the European Commission’s eAmbrosia database, we identify the origins of over 1,100 products from our wider dataset. The matched origin countries in this subset are legally defined and protected.

Our PDO products are mainly from regions with a tradition of GIs, notably the United Kingdom, France, Italy and Spain. Table 1 describes the price data collected for PDO products; Figure 1 provides a map.

Table 1. Summary of PDO products

Notes: Compares the number of PDO items registered in the eAmbrosia Geographical Indicators dataset (as of 20 February 2024) with the count of matching products in our dataset. Top eight countries by matched product count are shown.

Source: Authors’ calculations, European Commission (eAmbrosia).

Figure 1. Composition of product origin. Share of products assigned to each country from main prices dataset.

Notes: Excludes products originating from the UK (roughly 50%). Share calculation uses country-of-origin data identified in this paper, covering 67,000 products sold across seven UK supermarkets.

Source: Authors’ calculations.

2.2. Large language model text extraction

Many traded products do not have protected names. Our second method aims to assign provenance for a further 67,000 items by parsing item descriptions. Listings on supermarket websites often include an indication of a product’s origin. For example, the page of a juice drink might list the source of the fruit and the location where the packaging took place. By parsing these descriptions using a novel large language model (LLM)-driven methodology, we can gain a better understanding of the geographical distribution of product origins.

The challenge we face is the huge number of products on offer, and the complexity and variability in the language that retailers use in their descriptions. To solve this problem, we use an LLM to interpret and extract the relevant country or countries of origin from the text. LLMs are artificial intelligence (AI) tools that have been trained on large amounts of text, enabling them to parse and generate new text. In practical terms, the LLM acts as a research assistant (RA) tasked with summarising details from text, such as identifying the nuanced mentions of a product’s origin within its description.

The proposed LLM is faster than a human RA, and better than a naïve trawl through text using standard search functions. The LLM recognises diverse linguistic expressions that refer to locations identifying, for example, that ‘wine from the Rhône Valley’ gives an indication of origin. The LLM behaviour here varies between extrapolating for the country (so returning ‘France’) or allocating the named origin (so returning ‘Rhône Valley’). Responses of the latter type are limited in number and so can be manually added to a lookup table and replaced with the actual country. A more traditional method—parsing text for a list of target words, including ‘France’—would miss this. Our approach thus increases information capture.

Unlike Champagne or Stilton, many grocery products mention multiple countries in relation to their ingredients. In such cases, we attribute the product to all listed countries, based on an assumption of equivalence for each mentioned origin. For instance, a product described as ‘packed in France using meat from France, Germany and Spain’ is attributed an origin of (France, Germany and Spain). This pragmatic approach, of course, does not identify the true cost contribution from each location.

We take steps to maximise the accuracy and relevance of the LLM responses we get back from our requests. For each request to the LLM, we provide the product description (text extracted from its individual webpage) along with an array of previous messages. The first message takes the role ‘system’, which shapes the behaviour of the LLM for the subsequent conversation. The other messages are sample examples of previous user-assistant interactions. Including examples can improve the model’s reliability in understanding and executing the task, especially when dealing with nuanced or complex product descriptions. The examples act as blueprints for the model’s outputs. By seeing how past queries were structured and answered, the LLM can more accurately mimic this format, improving the consistency of output. Essentially, this serves to dynamically tune the responses to our task, while still leveraging the LLM’s preexisting knowledge. The example illustrates a typical interaction (Figure 2).

Figure 2. Example instructions to LLM. Simplified array of system and example messages. These are passed to the LLM with each request.

Notes: Full message available in Figure 5 in Annex. We use OpenAI’s gpt-3.5-turbo-1106 chat completions model.

The LLM responses (which are formatted as JSON data) are then processed, cleaned and merged with our prices databases. This application of LLMs augments our price dataset with origin data for 62% of the daily prices, significantly enhancing our ability to test the role of geography in our food price analysis. We identify a country of origin for over 67,000 products. Of these, less than half (49.4%) originate from the United Kingdom. The distribution of origin among imported products is shown in Figure 3. Italy is the top import source (11%), followed by France (10%) and Germany (8%).

Figure 3. Composition of product origin. Total number of products assigned to each country from main prices dataset.

Notes: Includes all products with an identified non-UK origin from our dataset of UK supermarket prices, irrespective of product category.

Source: Authors’ calculations.

Table 2 shows the price characteristics for the top 15 countries by item count. Beyond the United Kingdom itself, Italy (3,944), France (3,694) and Germany (3,008) are at the top of the ranking, with South Africa (690), India (664) and Australia (629) sitting at the bottom.

Table 2. Summary of prices dataset by country

Notes: N refers to the number of products assigned to each country (equivalent to Figure 3). Obs refers to the quantity of price quotes for those N products. This includes all products collected in the main prices dataset and so may include nonfood items sold by UK supermarkets. These nonfood items are removed prior to further analysis by matching each product with a food/drink COICOP division.

Source: Authors’ calculations.

Figure 4 shows the distribution of imports between EU and non-EU origins for products in our dataset, by store. The stores are anonymised and given numbers 1–6. All stores import more EU products than non-EU products, although the extent to which this is true varies. For example, store 1 sees 62% of its imports originating from the EU, while for store 5 this is 79%. There is also substantial inter-store variation in the share of UK-sourced products (not shown in the chart), ranging from 46% for store 1–62% for store 5. Store 1 is the only store with more than 50% of products showing non-UK origin.

Figure 4. Distribution of import origins by store. Food & drink (including alcoholic) items.

Notes: Share of products by region of origin. Stores anonymised to IDs 1–6. Share of unallocated products ranges from 20 to 50% across the stores. Share of UK products ranges from 46 to 62% across the stores.

Source: Authors’ calculations.

3.1. Frequency and size of price changes

The United Kingdom has continued to experience food inflation during our period of study. Between August 2023 and February 2024, CPI inflation for food was 0.67%. Microdata tend to show how prices ebb and flow (there are price cuts, even in inflationary periods) and that the net proportion of prices rising is a simple and powerful predictor of inflation (Davies, Reference Davies2021). We find a consistently positive net balance—that is, more rises than cuts—during our period of study, in line with the aggregate data.

Imported foods changed price more frequently than those produced in the United Kingdom. Across our period, 0.41% of the daily price quote pairs for UK products saw price rises (Table 3). Products originating from the EU were more likely to increase in price (0.55% frequency), as well as non-EU products (0.5% frequency). These internationally sourced products are also more likely to experience price cuts. Across each geography, price rises were more common than price cuts. Net rises were highest for non-EU products, at 0.12%. Net rises for EU products was 0.11%, while for UK products the figure was 0.09%.

Table 3. Frequency of price changes by origin of product

Notes: ‘Obs’ is a count of price quote pairs available for products from each region. These are consecutive daily price observations for all products originating from that area. ‘Total’ includes observations from all products in our dataset and so includes products whose origin was not identified by the LLM. Tables 3–5 consider the price change dynamics across every available price quote pair—that is, consecutive price observations for any specific product. ‘Frequency’ gives the proportion of consecutive price observations that represented a price change in any direction, while ‘Up’ and ‘Down’ split these by price rises and price cuts.

Source: Authors’ calculations.

Goods from the EU saw slightly more substantial price moves (mean and median price changes) than those from the United Kingdom, with both the EU and United Kingdom showing larger moves than non-EU imported foodsFootnote ³. EU products exhibited the largest mean price changes, with a mean rise of 27.8% against a mean fall of −22.4% (Table 4). The median price rise and fall is grouped around 20%, which may be driven by sales (Anderson et al., Reference Anderson2017).

Table 4. Size of price changes by origin

Notes: These values correspond to the observations detailed in Table 3. The mean and median values are each conditional on a price change occurring.

Source: Authors’ calculations.

We are also able to investigate price dynamics for individual countries. The 10 countries with the highest quantity available price quote pairs are shown in Table 5. Food from France showed the highest frequency of price changes at 1.47% (0.79% were up, 0.68% were down). The United States and Italy, similarly, see frequent price changes. Domestic (i.e., UK origin) foods changed prices much less, as did foods from Belgium and China. The size of price changes also varies by country (Table 10, Annex).

Table 5. Frequency of price changes by country

Notes: ‘Obs’ is a count of price quote pairs. These are consecutive daily price observations for all products originating from that area.

Source: Authors’ calculations.

3.2. Weighted price change

The importance of the price changes identified in our data rests on the food products weight in UK consumer expenditure. We capture this using a modified Laspeyres Price Index. Since we know the retailer selling each product, we can include retailer-specific ratings based on each supermarket’s market share ( $ {w}_s $ ). We also include product-specific CPI weights ( $ {w}_i $ ) by matching our products to COICOP divisions.Footnote ⁴ Baskets of products that are available on two dates $ \left(0,t\right) $ are constructed and the weighted price ratios are calculated. Weighting by CPI weights, store market shares and the number of prices from a given store $ s $ and type $ i $ , $ {N}_{i,s} $ , the price index at time $ t $ is equal to:

$$ {p}^{0,t}=\sum \limits_{S=1}^S\sum \limits_{i=1}^{N_s}\sum \limits_{j=1}^{N_{s,j}}\left(\frac{w_i{w}_s}{N_{s,j}}\times \frac{p_j^t}{p_j^0}\right) $$

As a benchmark, the official UK CPI index shows food prices to have increased by 0.7% between August 2023 and February 2024.Footnote ⁵ Our index suggests that imported foods contributed almost twice as much (2% and 1.9%, respectively, for non-EU and EU origins) to domestically sourced (1.3%) goods of rises in UK supermarket prices (Table 6).

Table 6. Size of price changes by origin—whole period

Notes: Tables 6 and 7 present the price changes and a price-change metric across the entire dataset. ‘Product’ refers to the number of unique products. This is calculated using a single period, with prices $ {p}_0 $ and $ {p}_1 $ for each product found in mid-August 2023 and mid-February 2024, respectively. These estimates are not directly comparable to an inflation rate. Mean and median changes are conditional on there being a price change.

Source: Authors’ calculations.

Turning to country of origin, the weighted price change of products purchased in the United Kingdom varies substantially. According to our dataset, products purchased in the United Kingdom originating from South Africa (4.6%), Spain (4.5%) and Thailand (4%) have increased the most over the period of study (see Table 7). In contrast, the weighted price change for products originating from France (1.4%), Ireland (1.6%) and Italy (1.7%) is the lowest in our dataset.

Table 7. Size of price changes by country—whole period

Notes: Equivalent conditions to Table 6. Top 10 countries with highest number of products matched from prices dataset.

Source: Authors’ calculations.

3.3. Inflation risks by product type

Our data allow us to identify pricing ‘risk’ by product type. Table 8 displays the top five products by import destination. Focusing on the EU, the table displays the top 10 items by the share of items imported from the chosen destination. For example, in our dataset, 77% of ‘olive oil’ is bought from EU countries. Baby food (70%) and wine (55%) are the next most concentrated item types in terms of UK imports from the EU.

Table 8. Top product type by origin

Notes: Shares and product counts relate to the sample of goods in the main prices dataset, including products with both an identified origin and a matched COICOP category, giving a pool of 28k items. Share ranges used for combined COICOP categories.

Source: Authors’ calculations.

As outlined in Section 1, on 31 January 2024, the UK introduced new border controls concerning ‘medium-risk’ and ‘high-risk’ food products. This means health certifications are now required for EU imports of medium-risk animal products, plants and plant products (e.g., meat, dairy, some fruits and vegetables). According to our dataset, meat products see an exposure of between 14 and 35%. Milk has a relatively small exposure, at 8–10%, while milk derivatives (e.g., cheese, yogurt etc.) have a higher exposure, between 25 and 46%.

The next policy changes are due on 30 April 2024 and 31 October 2024. In April, risk-based checks (documentary, identity and physical) on imports of the same medium-risk goodsFootnote ⁶ from the EU will be initiated. From October, safety and security certificates will be required. Although many fruit and vegetables are classified as medium risk, a temporary policy of ‘easement’ ensures these are being treated as low risk until October. Table 9 provides our best assessment of the exposure to EU imports at these three states.

Table 9. New EU food regulations

Notes: Data reflect products tracked from July 2023 to January 2024, corresponding to ‘food and drink’ COICOP groups. Out of 51k identified products, 28k were matched with a country/region of origin across six supermarkets. ‘Product count’ pertains only to these matched items; actual numbers affected by checks are likely higher. ‘Other dairy’ combines cheese, yogurt and similar products. ‘Fresh vegetables’ category excludes potatoes.

Source: Authors’ calculations.

Initial changes in January and April 2024 primarily affect meat and dairy products, with exposure ranging from 3% for eggs to 46% for cheese (other dairy products). October 2024 is expected to see the end of the easement policy for fruit and vegetables, marking a significant increase in exposure to 35% and 31%, respectively, for these categories.

To assess the potential effect of future NTBs, having a clear understanding of item origin is useful. Our new dataset and country-matching analysis provide an indication of potential exposure to item- and category-specific NTBs.