Did Milan restaurants’ Instagram posts become less alike during Italy’s ChatGPT ban? In the comparison with other cities, Milan restaurants’ posts used less similar wording during the ban, with an estimated relative drop of about 15%. This suggests that access to AI may be linked to businesses using similar wording.
ACADEMIC WORK REVIEWED
Generative AI and Content Homogenization: The Case of Digital Marketing
ResearchersChaoran Liu, Tong Wang, S. Alex Yang
SSRN · 26 July 2025 · First draft · Working paper · Source language: English
English review: 26 October 2026 · The source date appears in the citation above.
Editorial review draft
The source has been checked; the review still carries its original editorial status. DijitalPi did not conduct a customer experiment for this paper.
LET'S READ THE RESEARCH TOGETHER
We first explain the researchers' question, method and findings. We then discuss how to interpret the results, clearly separating DijitalPi's commentary from the source.
01 / WHAT DID THE RESEARCHERS WANT TO UNDERSTAND?
How is access to ChatGPT related to similarity in restaurant marketing copy and the likes it receives?
The researchers wanted to examine whether access to generative AI makes businesses’ marketing texts more alike. They also investigated the connection between access and consumer engagement, including the possible role of content similarity. Engagement here was measured through likes.
Explaining the context · DijitalPi commentary
DijitalPi commentary: Filling a content calendar and maintaining a distinctive business voice deserve separate assessment. Shared words and sentence structures can prompt a check on whether business-specific information is getting lost. This paper suggests considering similarity alongside production volume, without establishing how much differentiation would be commercially useful.
02 / HOW WAS THE RESEARCH CONDUCTED?
How did they test the question?
Chaoran Liu, Tong Wang and S. Alex Yang of London Business School used Italy’s ChatGPT ban, which took effect on 31 March 2023 and ended on 28 April 2023, as a natural experiment. A natural experiment uses an external change rather than an intervention randomly assigned by researchers. The data covered Instagram texts published in weeks 8-17 of 2023, from late February to late April. The affected group comprised 486 independent restaurants in Milan. Comparison groups included 429 in Paris and 432 in Lyon, France, and 544 in Madrid and 315 in Valencia, Spain. Public Google Maps and Instagram data were collected using Apify, with international chains excluded. Eligible businesses had posted before 1 January 2023 and had not become inactive after the ban.
Understanding the method · DijitalPi commentary
Businesses were not randomly assigned to groups. The researchers compared changes in Milan before and during the ban with changes in the control cities. The researchers used businesses in other cities to build a comparison that followed Milan’s pattern before the ban. They then estimated how far Milan differed during the ban. The methods are called Generalized Synthetic Control and Synthetic Difference-in-Differences. However, the researchers did not directly observe which businesses or posts used ChatGPT. The observed change concerned access to the tool.
The analysis covered post text only. Each business’s texts were compared with those of other businesses in the same city and week for vocabulary, word sequences, meaning and language style. The engagement measure was each business’s weekly average likes per post. The researchers used a calculation that reduced the influence of very large counts in the main analysis. This is called a logarithmic transformation. Likes were collected in a single snapshot at the end of 2024, so they represent accumulated engagement rather than the response at publication. Posts from businesses that hid their like counts were absent from the data.
03 / RESEARCH FINDINGS
Lower similarity and higher likes coincided during the ban
Estimates from the main method showed lower similarity in Milan during the ban relative to control cities: 0.003 for vocabulary, 0.003 for word sequences, 0.008 for meaning and 0.013 for language style. The authors reported relative declines against the corresponding means of approximately 15%, 12%, 2% and 3%. These are not percentage-point changes. Under the second method, vocabulary, word-sequence and meaning results were significant, but the style result was not. The style result also lost significance when week 17, which included Italy’s Liberation Day holiday, was excluded. In the main likes model, businesses’ weekly average likes per post were estimated to be approximately 3.5% higher during the ban. When extreme like counts were capped at the 5% level, the result remained significant under the main method but was not significant under the second method. An inverted-U association, with moderate similarity linked to higher likes and excessive similarity linked to lower likes, was significant only for word sequences and language style. Causality could not be established. Posting frequency and text length were also lower during the ban. The authors described the shortening of approximately one word per post as a small change.
Relative decline in vocabulary similarity
Approximately 15%The estimate from the main method expresses Milan’s decline in vocabulary similarity during the ban relative to the mean similarity score. It is not a percentage-point change.
Average likes per post
Approximately 3.5%The main logarithmic model’s estimate of higher weekly average likes per post during the ban. It does not measure sales, reservations or your business’s expected return.
DijitalPi commentary: Monitor how the copy changes. Do not treat AI use itself as a measure of success or failure. Keeping drafts and comparing human edits that add business-specific information may help reveal where similarity arises. Record posting frequency too: the study suggests that content volume and response per post need not move in the same direction. This editing approach was not an intervention tested in the paper.
Lower similarity occurring alongside higher likes does not establish that one caused the other. Other Instagram activities, including paid promotion, cannot be ruled out, and older posts have had longer to accumulate likes. Access to ChatGPT through a VPN also cannot be excluded. The authors suggest this could make the estimated effect smaller. The inverted-U association provides neither a universal similarity target nor a reason to remove every shared expression.
05 / CONCLUSION AND OPEN QUESTIONS
What did we learn, and what do we still not know?
The study did not measure sales or reservations. The study points to lower similarity, particularly in vocabulary and sentence structure, in Milan restaurants’ Instagram texts during the 2023 access ban. Individual ChatGPT use was not observed, and style findings and some engagement results were sensitive to the estimation approach. The findings do not directly extend to Türkiye, Turkish-language content, other sectors, images or videos. The reviewed text is a first-draft working paper dated 26 July 2025, with no information establishing that it underwent peer review. All three authors are affiliated with London Business School. No funding or competing-interest declaration was found in the source. Data came from public pages rather than a platform or company partnership.
Our open question at DijitalPi: how would similarity, likes per post and customer enquiries change if your Turkish social media copy were edited to include business-specific information? No implementation testing this question has yet taken place; the illustrative examples are neither research cases nor client cases.
We created these scenarios to make the topic concrete. They are not cases from the paper or measured client results.
EXAMPLE 01
E-commerce: describing different products the same way
An illustrative home textiles retailer uses the same comfort and style claims in Instagram copy for different products.
Does our copy explain what distinguishes our products?
Open the recommendation for example 01+
Try replacing generic adjectives with verified product features.
Keep the AI draft and edited version to compare repeated words and sentence structures. Check whether the copy includes product-specific fabric, sizing or usage information. Record likes after the same elapsed time for each post, alongside its visual and any paid promotion. This study does not establish that the proposed editing will increase likes or orders.
EXAMPLE 02
Local services: interchangeable maintenance promises
An illustrative bicycle repair business publishes general maintenance messages resembling those of nearby businesses.
What does this text tell customers that is specific to us?
Open the recommendation for example 02+
Make the actual service scope and booking conditions clear.
Review service posts published in the same area and period, marking shared expressions. Add only verified service information to your draft and record the sentences you change. Track likes per post separately from booking enquiries. Findings about restaurant copy do not establish the same effect for bicycle repair businesses.
EXAMPLE 03
B2B: technical offers become generic promotion
An illustrative industrial equipment supplier prepares Instagram copy describing different machines through similar efficiency claims.
Can readers distinguish which machine suits which task?
Open the recommendation for example 03+
Ask a product specialist to review the drafts and clarify differences in use.
Add verified operating conditions and product differences alongside repeated sentences. Before publication, record which concrete information distinguishes each text. Assess likes separately from quotation requests and note paid distribution. The research did not test B2B purchasing decisions or this editing approach.
TRY IT WITH YOUR TEAM
Review repetition in your copy
Plan a comparison of AI drafts and human-edited copy, recording repeated expressions, likes per post and posting frequency separately. Account for visuals, paid promotion and the time available for likes to accumulate. This is a proposal that DijitalPi has not yet implemented.
SMMMACHINE creates social media copy and visuals in your brand’s voice and checks content quality. This production process does not replace the comparison proposed in this article. Explore SMMMACHINE ↗
The academic work belongs to the researchers named above. This page contains DijitalPi's explanatory review and original business examples; it is not a full translation of the paper.
Cookie preferences