After you update your services on your website, does AI give customers the new information? Brand positioning describes what a business offers and who it serves. In this SSRN working paper, which has not undergone peer review, researchers found a 40-percentage-point decline in correct category descriptions during the first two days after fictional consulting brands changed positioning. DijitalPi commentary: Checking what AI tells customers could become part of your transition plan. The study measured neither sales, enquiries nor website traffic. Its specific models and controlled setting do not establish an updating timetable for Turkish answers or established businesses in Türkiye.
ACADEMIC WORK REVIEWED
Beyond Consumer Memory: How Large Language Models Update Brand Knowledge
English review: 14 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 a brand changes its services, how quickly and accurately does AI describe the change?
The researchers examined how changes to a company’s brand information appear in AI answers. They tested how updating varies with the model and the wording of the question. They also investigated recommendations during the transition.
Explaining the context · DijitalPi commentary
DijitalPi commentary: In a repositioning plan, checking whether the updated website is live is straightforward. Whether customers receive an updated description from AI requires a separate check. The marketing value of this study lies in distinguishing a brand’s presence in an answer from an accurate account of its current services.
02 / HOW WAS THE RESEARCH CONDUCTED?
How did they test the question?
Hamann, Stahl and Kannan conducted a controlled field experiment with 8 fictional consulting brands situated in a small town with little consulting competition. Websites established in May 2025 were submitted to Google and Bing for indexing one month before the experiment. On 11 October 2025, 3 brands moved from strategy to environmental consulting and 3 moved in the opposite direction; 1 brand in each category remained unchanged. Between 10 October and 15 November 2025, the researchers collected 1,392 observations through the consumer interfaces of ChatGPT-5, Gemini 2.5 Pro and DeepSeek Thinking (V3.2-Exp): 240 before repositioning and 1,152 afterwards. They used clean sessions without personalisation, keeping the device, user and location the same.
Understanding the method · DijitalPi commentary
The researchers changed website text. They kept the structure and design the same. The researchers compared answers before and after repositioning for changed brands against control brands whose services remained unchanged. This approach aims to distinguish the effect of repositioning from changes affecting all brands over time. Measurements took place every 6 hours during the first two days, daily for the following 12 days and weekly for a further 3 weeks. The main analysis used one answer per question and measurement wave.
Questions asked what the firm offered or whether its focus had changed recently. The researchers first recorded whether the answer identified the brand. For identified brands, they assessed whether the current category was correct and how closely the answer resembled the meaning of old or new website text; naming both the old and new categories counted as incorrect. A separate recommendation measure required models to choose a preferred provider for a target category from the 8 experimental brands alone. The authors’ classification of models by retrieval behaviour draws on documentation and observed responses; internal mechanisms were not directly verified.
03 / RESEARCH FINDINGS
AI can find a brand but describe its services incorrectly.
The researchers reported a statistically significant 40-percentage-point decline in correct category descriptions during the first two days, conditional on detecting the brand. The effect weakened in the medium term, days 3–14, and alignment recovered over days 15–35, although effect estimates for both later periods were statistically non-significant. Repositioning had no significant effect on brand detectability. Questions explicitly asking about a change yielded an estimated improvement of roughly 10 percentage points in category alignment; one relevant estimate met the 5% significance threshold, while the other did not. Similarity to old website text fell by roughly 6%, 13% and 18% over the short, medium and long term respectively; these are relative percentages, not percentage points. The estimated recommendation decline in the first two days was 48 percentage points relative to unchanged brands; however, it did not meet the 5% significance threshold. Medium- and long-term recommendation estimates were also negative but less precise.
Correct service descriptions in the first two days
40-percentage-point declineStatistically significant experimental estimate for accurately describing the current category when the brand was detected; not a change in sales or traffic.
Recommendations in the first two days
48-percentage-point declineEstimated change in selection for the target category from the experimental brands. Not significant at the 5% threshold and not a measure of lost customers.
DijitalPi commentary: Record which model produced an answer, when it was checked and which question was used. A correct answer to a question that mentions the change may conceal outdated information in responses to a general service question. Consider tracking whether updated pages have been re-indexed as well: the additional analysis found a positive association between re-indexing and correct category descriptions. This association does not establish that requesting re-indexing will correct answers within a particular time.
Differences between models depend on the versions observed in October–November 2025 and the authors’ inferred retrieval classifications. The experiment did not test established brands’ information in training data, press coverage or user content. Adding a blog or frequently asked questions page, and changing communication style, also produced no significant differences in exploratory analyses; with 2 brands per channel condition and 3 per style, these results should not be read as evidence that those approaches are ineffective.
05 / CONCLUSION AND OPEN QUESTIONS
What did we learn, and what do we still not know?
This SSRN working paper shows that updated website information may not enter AI answers immediately or completely. The date of 30 May 2026 belongs to the publication record; the manuscript is undated and there is no peer-reviewed journal version. Fictional brands, a single industry, a short period before the change and the small number of brands limit inference; formal assessment of comparable trends before the change was restricted. The study provides no validated updating timetable for Turkish or Türkiye. The authors registered their research plan before collecting data and disclosed four departures in the final implementation. No funding or competing-interest statement was found in the source; the authors are university academics, and no connection with the model providers examined is stated.
Our open question at DijitalPi: after changing your services, which general customer questions will you monitor for outdated descriptions, and how will you measure those answers separately from incoming enquiries?
We created these scenarios to make the topic concrete. They are not cases from the paper or measured client results.
EXAMPLE 01
B2B software: moving to a new use case
An illustrative software company repositions its general work tracking product as a service exclusively for procurement teams.
Does AI describe our new use case to customers?
Open the recommendation for example 01+
Check brand mentions separately from accurate descriptions of the current use case.
Save answers from before the change and repeat the same questions after updating the website. Record whether each answer describes the old use case, the new one or both. Record whether the search engine has stored the updated pages, alongside the date of each check. The updating times observed for consulting brands are not a delivery commitment for this software business.
EXAMPLE 02
Tourism: moving from city tours to nature trips
An illustrative tourism business stops offering city tours and focuses on nature trips.
Is a correct answer to a question about the change enough?
Open the recommendation for example 02+
Also ask a general service question; do not assume customers know a change has occurred.
In separate sessions with the same model, ask what the business offers and whether its focus has changed recently. Mark any references to old tours that are no longer available. Keep the questions and checking dates in a consistent record for comparison with later answers. The question-format finding does not establish the same result for tourism questions in Turkish.
EXAMPLE 03
Local services: moving from domestic to workplace cleaning
An illustrative cleaning business stops serving homes and concentrates exclusively on workplaces.
Does being recommended for the wrong service actually cost us customers?
Open the recommendation for example 03+
Track answer accuracy, recommendations and incoming enquiries separately.
During AI checks, record which service the business is recommended for. Separately ask incoming prospects which service they want and how they found the business. Interpret these records while retaining the statistical uncertainty around the paper’s estimated recommendation decline. Being absent from an answer is not itself a measure of lost customers or revenue.
TRY IT WITH YOUR TEAM
Create a record for the transition.
Keep one record of the updated pages, whether search engines have stored the changes, and answers to different questions. Use your own checks to assess whether answers consistently describe the new service, without treating the paper’s day counts as a target timetable. This is a proposed pilot that DijitalPi has not carried out.
You can start correcting information about your brand on your own site. SEOMACHINE supports content preparation using your brand information, source checks and editor approval. Updates to that information in external AI systems need to be monitored separately. Explore SEOMACHINE ↗
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