Overview
This case study was conducted during Phase 1 of Narr Theory research, following the initial Phase 0 investigation. The objective was to find further evidence on how the free version of ChatGPT recommends products, with particular attention to the role of its Internal Knowledge Model (IKM).
The central question was not simply which products does ChatGPT recommend? Rather, it was:
How does ChatGPT arrive at a product recommendation before, and alongside, its access to web-based evidence?
This question is particularly relevant to AI Product Recommendation (AIPR) because of the token-based nature of large language models.
At a fundamental level, a large language model generates tokens based on the context established by the preceding tokens. In a product recommendation query, the user’s decision narrative establishes that context. The model then generates the next sequence of words according to learned patterns and associations.
This means that the model’s existing semantic associations matter.
If ChatGPT has already established strong associations between particular brands, products, attributes, and decision narratives, those associations can influence what it generates next. The model does not begin every product recommendation from an entirely blank state. Its existing knowledge provides a set of weighted associations that can influence its reasoning and recommendations.
This case study therefore investigated whether the products recommended through ChatGPT’s Internal Knowledge Model would remain semantically consistent with products subsequently identified through web search.
Methodology
Decision Narrative
A deliberately simple decision narrative was used:
“I am looking for a laptop for professional work. I want something extremely fast, easy to use, and reliable. My budget is $2,000.”
The narrative establishes a clear premise while providing a limited number of constraints that could be investigated individually.
The four primary constraints were:
- Professional work — establishes the use case.
- Extremely fast — introduces a technical, experiential, and comparative requirement.
- Easy to use — primarily experiential.
- Reliable — primarily experiential.
- $2,000 budget — establishes a financial constraint and priority.
The intentionally generic language was important.
Terms such as fast, easy to use, and reliable are not completely objective. Even “fast,” despite having measurable technical dimensions such as processing speed and RAM, remains comparative and contextual. A laptop can only be considered fast relative to other laptops and relative to the user’s intended use.
This made the narrative useful for investigating how ChatGPT translates relatively ambiguous language into product recommendations.
Offline and Online Search
Two separate environments were used to distinguish between recommendations generated primarily from ChatGPT’s existing knowledge and recommendations generated with access to web search.
Offline Search
The first test used the ChatGPT desktop application with web search and connected search disabled.
The model was ChatGPT 5.6 Free.
This environment was used to investigate recommendations generated from the model’s existing knowledge rather than from a live search of the web.
Online Search
The second test used an incognito Chrome browser with a separate account and the ChatGPT 5.6 Free model, with online search available.
The purpose was not to determine which environment was “better,” but to compare the products and semantic patterns produced under each condition.
Epitome Tests
The full decision narrative was then broken down into individual constraint “verticals.”
This was called the Epitome Test.
Instead of asking ChatGPT to solve the entire decision narrative, the prompt would isolate one particular requirement. For example:
“I need a laptop for professional work. I would like to prioritize the speed of the laptop. I want a fast laptop under $2,000. What do you recommend?”
This allowed each constraint to be examined independently.
The purpose was to identify the products that ChatGPT appeared to associate most strongly with each particular attribute.
If a particular brand or product repeatedly appeared when a specific decision constraint was isolated, this provided a useful indication of its semantic association with that constraint.
The results could then be compared against the recommendations produced by the complete decision narrative.
Key Finding 1 — Product Recommendations
The first finding was the consistency of brand-level recommendations between the Internal Knowledge Model and web search, despite differences at the individual product-model level.
Full Decision Narrative: Offline Search
Using only the Internal Knowledge Model, ChatGPT recommended products including:
- Lenovo ThinkPad T14 / T14s
- MacBook Air / MacBook Pro with Apple Silicon
- Dell Latitude 7000 series
- HP EliteBook 800 series
- Microsoft Surface Laptop
The recommendations were therefore distributed across several established professional-laptop brands.
Full Decision Narrative: Online Search
When the same decision narrative was tested with online search, ChatGPT recommended:
- MacBook Pro 14-inch
- Lenovo ThinkPad X1 Carbon
- Dell XPS 14
The exact product models changed.
However, the brand associations remained remarkably consistent.
Lenovo, Dell, and Apple remained present across the two environments, even though the particular models recommended were different.
This creates an important hypothesis for further AIPR research:
The web-search recommendation may sometimes function as an extension or further articulation of associations that already exist within the Internal Knowledge Model.
In other words, the model may already have a strong association between a particular brand and a particular decision narrative before it searches the web for supporting evidence.
The web search can therefore change the specific product expression of the recommendation without necessarily changing the underlying brand association.
This distinction becomes particularly important when examining the descriptions ChatGPT provides for each recommended product.
The description attached to a recommendation is itself generated through tokens and therefore provides another layer of evidence about how the model associates that product with the decision narrative.
For example, if a product is recommended for being “fast,” the explanation of why it is fast provides additional clues about the semantic relationship between the product, the attribute, and the original decision narrative.
This suggests that future case studies should not examine only which products are recommended, but also how ChatGPT describes those products.
Epitome Test Results
Speed
Offline Search
When the decision narrative was narrowed to prioritize speed, the Internal Knowledge Model recommended:
- Lenovo ThinkPad P1
- Dell Precision 5000 / 7000 series
- Asus ProArt P16
- MacBook Pro 14-inch
- Lenovo ThinkPad T14 / T14s
Online Search
The online-search version recommended:
- Lenovo ThinkPad T16 Gen 4
- Lenovo ThinkPad T14 Gen 6
- Lenovo ThinkPad E16 Gen 3 AMD
- Asus ExpertBook P5
- Dell Pro 14 Plus
The most notable pattern was Lenovo ThinkPad’s repeated appearance.
Three of the online-search recommendations were Lenovo ThinkPad models.
The brands themselves, however, remained broadly consistent with the Internal Knowledge Model.
Asus appeared in the online results as well. This may represent a new association introduced through web search, although this conclusion requires further testing. It is possible that Asus would have appeared in the original Internal Knowledge Model recommendation if a larger number of recommendations had been requested.
This is therefore an area for future experimentation.
Ease of Use
Offline Search
The Internal Knowledge Model recommended:
- MacBook Air 15-inch M-series
- ThinkPad T14s
- MacBook Pro 14-inch
- Dell Latitude 7450 / 7000 series
- Asus ProArt P16
Online Search
The online search recommended:
- Lenovo ThinkPad E14 Gen 7
- Lenovo ThinkPad T14
- HP EliteBook
Again, the exact models changed, but several of the underlying brands remained within the broader recommendation ecosystem.
Reliability
Offline Search
The Internal Knowledge Model recommended:
- Lenovo ThinkPad T14 / T14s
- Dell Latitude 7000 series
- HP EliteBook 800 series
- Lenovo ThinkPad P14s
- MacBook Pro 14-inch
Online Search
The online search produced:
- Lenovo ThinkPad
- Dell Pro / Latitude
- HP EliteBook
- Lenovo ThinkPad E / L series
- Consumer Dell / HP / Lenovo
- Gaming laptops
The “reliable” epitome produced more generic product categories and broader brand references than some of the other tests.
One possible explanation is that the underlying knowledge associated with reliability may encompass a larger and more diverse set of products. However, this remains an observation rather than a confirmed explanation and requires additional testing.
Key Finding 2 — Internal Knowledge Model Advantage
The results provide a potential strategic advantage for brands that already possess strong semantic representation within the Internal Knowledge Model.
The reason is fundamental to the nature of large language models.
LLMs use tokens to establish semantic relationships between words, phrases, concepts, products, brands, and contexts. Consequently, having a brand or product strongly represented within those learned associations may increase the likelihood that it is generated when a corresponding decision narrative is introduced.
This is analogous, in some respects, to human recall.
If someone asks:
“What car should I buy?”
A person may immediately respond with:
“Toyota.”
Or Honda. Or Lexus.
The person does not necessarily begin by calculating every vehicle currently available in the market. They recall a brand that already possesses meaning within their knowledge structure. Specific products are then variations or manifestations of what that brand represents.
The same general phenomenon may occur within AI Product Recommendation.
A brand can become associated with particular semantic concepts:
Toyota → reliable
ThinkPad → professional
MacBook → easy to use
These are simplified examples, not conclusions of this case study. But they illustrate the underlying research question.
If a brand has a strong semantic association with the words and concepts contained within a user’s decision narrative, that brand may have an advantage when the model begins generating its recommendation.
This creates an important implication for mass-market brands:
Brands need to understand how they are represented within the Internal Knowledge Model.
Toward Narrative Intelligence
This raises a practical question:
How could a brand determine how it is semantically associated with a particular decision narrative?
One possible solution is a Narrative Intelligence Tool capable of examining the semantic relationships between a brand, its products, and the decision narratives that consumers use.
At first glance, this might suggest analyzing the entire web.
However, attempting to process and continuously monitor the entire web would be inefficient. The volume of information is enormous, and a third-party system would face significant challenges in processing, storing, and interpreting the complete dataset.
A more efficient approach may be to identify the Sources of Authority that matter most within a particular product category and decision narrative.
Rather than asking:
“What does the entire internet say about this product?”
the system could ask:
“Which sources of authority shape the semantic association between this product and the decision narrative?”
It could then track how brands and products are associated with specific attributes, use cases, problems, comparisons, and other decision variables.
This provides a potential direction for future AIPR research and for the development of a Narrative Intelligence system.
Key Finding 3 — ChatGPT Already Knew the Answer
One of the most significant observations from the case study was the possibility that ChatGPT already knew the answer before performing web search.
The Internal Knowledge Model already contained associations between:
- laptops and professional work
- laptops and speed
- laptops and ease of use
- laptops and reliability
- particular brands and those attributes
The model therefore appeared to have an existing conceptual structure that could produce a recommendation before retrieving live evidence.
Web search could then provide additional evidence for that recommendation.
This creates an important distinction in how AI Product Recommendation may operate:
Recommendation → Search → Evidence
rather than necessarily:
Search → Evidence → Recommendation
If the first sequence is occurring, then the model’s pre-existing semantic associations become strategically important.
The web may not always be determining the recommendation from scratch. It may instead be corroborating or refining an association that already exists.
This hypothesis requires considerably more research, but the consistency between the offline and online brand recommendations provides a useful starting point.
Key Finding 4 — The Epistemic Semantic Chain
The preceding finding leads to a broader concept: the Epistemic Semantic Chain.
For an intelligent system to recommend a product, it must first establish what the words in the user’s decision narrative mean.
Consider the word:
Reliable
“Reliable” does not have a single universal meaning.
In the context of laptops, reliability could refer to:
- infrequent hardware failure
- durability
- ease of repair
- resistance to physical damage
- long-term performance
- battery consistency
- warranty or service support
Likewise, “easy to use” could involve:
- operating-system familiarity
- keyboard quality
- screen visibility
- setup simplicity
- startup speed
- software integration
- repairability
The decision narrative therefore begins as a string of words.
Those words must be interpreted.
They then acquire meaning within a domain.
The domain establishes additional relationships between concepts.
Within that domain, brands and products acquire meaning.
Finally, those brands and products can be evaluated against the user’s specific decision narrative.
This can be represented conceptually as an epistemic semantic chain:
Word → Meaning → Context → Domain → Brand/Product → Decision Narrative
The stronger and more interconnected these semantic relationships become, the more effectively an AI system can potentially interpret the relationship between a product and a user’s decision.
This is why simply “being mentioned on the internet” may not be sufficient.
A product needs to be associated with the right concepts, in the right contexts, through the right sources of authority.
Further research into the Epistemic Semantic Chain will therefore be necessary.
Key Finding 5 — The ChatGPT Free Model Problem
The final finding was considerably less theoretical:
The ChatGPT Free model produced a number of practical problems during product recommendation.
Some of these problems were already expected from previous experience, but the case study provided additional examples.
Incorrect Citations
ChatGPT sometimes cited the wrong brand website.
For example, it could discuss a Dell product while providing a Lenovo citation.
This creates a particularly serious problem for product recommendation because the citation is supposed to provide evidence for the recommendation.
If the source does not actually support the product being discussed, the recommendation becomes difficult to verify.
Weak Reasoning
At times, ChatGPT’s reasoning did not appear coherent.
When questioned about its own reasoning, it could move in circles rather than providing a clear explanation of how it arrived at its conclusion.
This is consistent with the broader problem of attempting to interpret a token-generation system as though it were necessarily performing a linear human reasoning process.
The generated explanation can be lengthy without necessarily providing proportional informational value.
Excessive Output
ChatGPT also tended to produce too much text relative to the value returned.
This creates a cognitive-load problem.
For product search, consumers generally want to reduce the amount of work required to make a decision. If an AI system generates a large amount of text without producing a correspondingly strong increase in decision value, the technology can become less efficient than the search behavior it is intended to replace.
Implication for AI Product Search
These weaknesses could have significant consequences for AI Product Recommendation adoption.
Consumers may be willing to experiment with AI for product discovery because it offers a conversational alternative to traditional search.
However, product purchasing introduces a higher requirement for accuracy, evidence, and efficiency.
Google already possesses an established consumer behavior around product search. If consumers repeatedly encounter inaccurate citations, questionable recommendations, circular reasoning, or excessive text in ChatGPT, they may return to Google for purchase-related queries.
This does not necessarily imply that ChatGPT will disappear from AI Product Recommendation.
Rather, it suggests a potential competitive vulnerability during the adoption period.
ChatGPT has demonstrated significant utility for asking questions, writing, reading, and general information discovery. But product recommendation introduces a different standard: the recommendation needs to be useful enough, accurate enough, and efficient enough to justify changing an established purchasing behavior.
If it fails to meet that standard consistently, consumers may simply return to the search engine they already trust.
The broader implication is therefore not that AI Product Recommendation will fail.
It is that accuracy and semantic reliability may become decisive factors in determining which AI product-search systems consumers ultimately adopt.
Conclusion
This case study provides preliminary evidence that ChatGPT’s product recommendations may be influenced by semantic associations already present within its Internal Knowledge Model.
The strongest observation was the consistency between offline and online recommendations at the brand level, despite differences in the specific product models recommended.
This supports a hypothesis worth investigating further:
Web-based product recommendations may sometimes represent a further token articulation of semantic associations that already exist within the model’s Internal Knowledge Model.
If this is true, then the implications for brands are significant.
A brand’s visibility in AI Product Recommendation may depend not only on whether it can be retrieved from the web, but on what that brand means to the model before retrieval occurs.
This shifts the strategic question from:
“How do I get my product onto the web?”
toward:
“How is my product semantically understood within the decision narratives that matter?”
That question leads directly toward the development of the Epistemic Semantic Chain, Source of Authority analysis, and eventually a potential Narrative Intelligence system capable of monitoring the semantic relationships that influence AI Product Recommendation.
The laptop case study therefore represents an early step in investigating not simply what ChatGPT recommends, but why certain products appear to be available for recommendation in the first place.
Leave a Reply