Do customer ratings fall around new AI model releases?
Are new AI model releases linked to lower customer ratings on Trustpilot? In the 7 days after model releases and price cuts, the share of 1-star reviews rose by 0.34 percentage points among unverified reviews relative to verified reviews, which are linked to a purchase or service experience. A business’s rating changes can differ depending on whether reviews are verified.
English review: 23 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?
When AI becomes cheaper or new models arrive, how do unverified customer reviews change relative to verified reviews?
The researchers investigate whether changes in AI pricing and capabilities are associated with changes in online review behaviour. Because they cannot directly observe which reviews were written with AI, they compare activity around model releases and price reductions.
Explaining the context · DijitalPi commentary
DijitalPi commentary: Monitoring review totals alone can miss changes in the distribution of ratings. This study provides a reason to include verification status and timing in reputation monitoring. The useful business question is where the change occurs and whether it matches customer experience records, without immediately labelling individual reviews as fake.
02 / HOW WAS THE RESEARCH CONDUCTED?
How did they test the question?
Valeria Lerman, Oren Rigbi and Yaniv Dover examined data covering 13,818,281 Trustpilot reviews and 136,750 companies between 1 January 2023 and 31 December 2024. No country is specified and no country breakdown is provided. The main analysis covers 2,797,199 reviews and 72,298 companies within the 7 days before and 7 days after OpenAI price reductions and new model releases, excluding event day. The design is observational: changes in unverified reviews are compared with changes in verified reviews linked to a purchase or experience.
Understanding the method · DijitalPi commentary
The researchers do more than compare ratings before and after each event: they estimate how the change in unverified reviews differs from the change in verified reviews. Verified reviews serve as the comparison group. The analysis accounts for persistent differences between companies and conditions shared within calendar weeks. It assumes that, without the event, the groups would have followed similar trends. Pre-event rating trends support that assumption.
Outcomes are measured by company, day and verification status: average ratings, shares of star categories, review counts, review length and textual similarity. The data came from third-party provider Bright Data, rather than a collaboration with Trustpilot, and are not publicly accessible. They contain reviews remaining after the platform’s fake-review filtering. The study does not label individual reviews as AI-generated or measure changes in customer purchases, sales or revenue.
03 / RESEARCH FINDINGS
Ratings shifted towards negative reviews, with no significant change in review counts
The researchers found a significant decline in the average rating of unverified reviews relative to verified reviews during the 7 days after events. The coefficient in Table 3 is −0.0171 rating points. In the same comparison, the share of 1-star reviews increased by 0.34 percentage points and the share of 5-star reviews decreased by 0.46 percentage points, both with p < 0.01. This reports statistical significance, not the probability that reviews were written with AI. The change in verified reviews’ average rating was not significant. Effects on review counts, length, textual similarity and the shares of 2-, 3- and 4-star reviews were also not significant. Separate analyses located the pattern mainly around new model releases. Results for price reductions alone and events combining price reductions with model releases were generally small and insignificant. In the daily analysis, the average rating decline was most pronounced approximately 3-5 days after events before recovering.
Share of 1-star reviews
+0.34 percentage pointsThe change in unverified reviews relative to verified reviews in the 7-day post-event window. This is neither a percentage increase nor a review count.
Share of 5-star reviews
−0.46 percentage pointsThe change in the share of high ratings over the same period and comparison. It does not measure lost sales or revenue.
DijitalPi commentary: Start by comparing the spread of ratings, whether reviews are verified and when the model was released. The absence of a significant review-count result suggests that monitoring triggered solely by rising volume might not capture the pattern reported here. A new model release can be recorded as a date to investigate, without treating it as an established cause of negative reviews.
The authors interpret the negative shift as a pattern potentially consistent with AI being used to target competitors rather than promote a business itself. However, neither the reviewers’ identities nor their intentions are known. More effective removal of positive fake reviews could also explain the observed distribution. Verified reviews may be an imperfect comparison group, so these findings cannot establish responsibility for an individual review, business or competitor.
05 / CONCLUSION AND OPEN QUESTIONS
What did we learn, and what do we still not know?
The study does not identify which individual reviews were written with AI or measure lost sales. This preprint provides indirect evidence of an association between new model releases and a short-lived negative shift in unverified Trustpilot ratings. It does not estimate the overall prevalence of AI-generated reviews. The authors describe the observed findings as a lower bound on the relevant activity. The analysis covers one platform, OpenAI events only and the 2023-2024 period. Longer-term persistence is unknown. It offers no results specific to Türkiye or Turkish-language reviews, so the pattern cannot be transferred directly to local businesses. The paper has not undergone peer review. All authors are university researchers, with no company-affiliated author. No funding or competing-interest statement was found in the source.
Our open question at DijitalPi: when your rating distribution changes, do you have the records to examine whether it coincides with verification status or a genuine service problem?
We created these scenarios to make the topic concrete. They are not cases from the paper or measured client results.
EXAMPLE 01
E-commerce: the share of low ratings changes
An illustrative online shop notices that low ratings seem more prominent after a new AI model is released.
Are more reviews arriving, or is the distribution of ratings changing?
Open the recommendation for example 01+
Examine review counts and rating shares separately by verification status.
Group existing review records by date, star rating and verification status. Comparing the 7 days before and after the model release, excluding release day, would adapt the paper’s window for a local investigation. Add records of delivery problems during the same period. This simple comparison does not reproduce the paper’s analysis or detect AI use.
EXAMPLE 02
Local services: ratings fall among a small review base
An illustrative cleaning business sees negative feedback becoming prominent among its limited customer reviews.
Does this reflect a service problem or a review pattern that needs investigation?
Open the recommendation for example 02+
Check the experiences described alongside existing service records.
Compare the dates and subjects of complaints with the relevant service records. Record verification status where the platform provides it, and support genuine customers in sharing verified reviews of their experiences. Remember that the paper groups companies by review volume, rather than turnover or employee numbers. You cannot assume that this business will experience the same pattern or that every verified review is unproblematic.
EXAMPLE 03
B2B software: similar wording raises suspicions
An illustrative B2B software company is considering attributing negative reviews with similar wording to AI.
Does similar wording show that reviews were generated with AI?
Open the recommendation for example 03+
Do not use textual similarity as your sole decision criterion.
List review ratings, dates and verification status together. Examine product updates and support records to investigate whether customers share a genuine complaint. The paper found no significant effects on textual similarity or review length. Similar wording therefore cannot establish that a review was AI-generated or came from a competitor.
TRY IT WITH YOUR TEAM
Separate review monitoring by verification status
Prepare a view of existing records that tracks rating distributions, review counts and verification status together. Plan to compare changes around model releases with service records, without treating timing as grounds for an accusation. This is a proposal that DijitalPi has not yet implemented.
Valeria Lerman; Oren Rigbi; Yaniv Dover. The Hebrew University of Jerusalem, Business School (Valeria Lerman, Oren Rigbi, Yaniv Dover); Department of Cognitive and Brain Sciences (Yaniv Dover); The Federmann Center for the Study of Rationality (Yaniv Dover).
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