Aug 13,2026 
Hoverboard analyzed 143,295 responses from anonymous ChatGPT sessions and found that some of the answers returned on different days were byte-for-byte identical, including the wording, structure and citations.
The behavior remains rare: 0.09% of the responses we analyzed in July matched an earlier response, compared with 0.44% so far in August. That does not prove ChatGPT served those answers from a cache. But separate research by RESONEO recently uncovered an experimental OpenAI configuration designed to match prompts against precomputed answers, giving us a plausible mechanism worth investigating.
For brands, the implication is potentially significant: AI visibility may not depend only on whether ChatGPT discovers and recommends you, but also on how long that answer persists.
That makes the question much more interesting: Are we beginning to see ChatGPT’s answer cache operating in the wild?
We Are Finding Byte-Identical ChatGPT Answers Across Different Days
Across the ChatGPT responses analyzed by Hoverboard, we identified cases where ChatGPT returned exactly the same response to the same prompt on different days.
Not approximately the same answer. Not the same companies presented differently. Not similar wording.
The responses were byte-identical.
That means thousands of characters of text appeared exactly the same across separate prompt checks performed on different days. Sometimes this occurred on back-to-back days; in other cases, identical responses were separated by longer intervals. In a few cases we found the same multi-thousand-character response appeared three or four times across several days.
Here’s one example from Hoverboard AI:
Identical ChatGPT response from August 6 and August 8
The example does not establish why the response was repeated. But identical text, structure and citations across separate sessions give us a concrete pattern to investigate.
How Often Are We Seeing Duplicate Answers?
Of the 143,295 ChatGPT responses we tracked, only a small percentage matched an earlier response byte for byte.
In July: 0.09% of responses were duplicates. In August so far: 0.44% of responses were duplicates. So, we’re talking a very small number at the moment. It’s important to put these numbers in perspective:
More than 99.5% of the responses we’re currently observing are not duplicates.
So we’re not suggesting that ChatGPT has suddenly started serving cached responses for a large percentage of questions. Quite the opposite. The behavior remains rare.
What Types of Prompts Trigger Duplicate Answers?
Although the duplicate sample remains small, nearly all of the cases identified so far involve single-brand factual or reputation prompts. But rare behavior becomes detectable when AI responses are measured at sufficient scale. The next question is whether these duplicate answers correlate with particular prompt categories, models or retrieval behavior.
What RESONEO Found Inside ChatGPT
Much of what we currently know about this particular ChatGPT experiment comes from research by Olivier de Segonzac and RESONEO.
On his LinkedIn post, Olivier talks about how he found that ChatGPT pages expose a Statsig payload containing feature gates, dynamic configurations and experimental layers that OpenAI can use to test functionality across different groups of users. RESONEO built a tool that monitors those experiments over time. Among the configurations they discovered was one that appears designed for semantic answer caching. According to their analysis, a configuration associated with a free-tier GPT-5.5 variant contained a system capable of matching short prompts against a prebuilt index called atc_cache_graph, using an OpenAI retriever named labrador.
The configuration they observed included:
- Prompts up to approximately 1,000 characters
- Approximate nearest-neighbor vector search
- A cosine similarity threshold of 0.93
- BM25 lexical weighting set to 0
- 10 retrieval candidates
- Top-1 reranking
- A 24-hour cache lifetime
- A roughly 500 millisecond lookup budget
- Additional safeguards involving numbers, URLs, emails and other literal differences
Their discovery gives us a potential technical mechanism for something we’re now independently observing in response data.
What Exactly Is Semantic Answer Caching?
Most people probably imagine ChatGPT working something like this:
Question → AI model → newly generated answer
When current information is required, another step might occur:
Question → search/retrieval → AI model → answer
Semantic caching introduces another potential path:
Question → find highly similar previous question → reuse precomputed answer
The important word is semantic.
The questions don’t necessarily have to contain exactly the same words.
Consider these two questions:
“Where can I find the highest quality peptides?”
and:
“What’s the best place to buy high-quality peptides?”
Those aren’t identical strings, but they essentially mean the same thing. Embedding models can represent the meaning of each question mathematically and measure how similar they are. If the similarity exceeds a sufficiently high threshold, a caching system could potentially conclude: We’ve effectively answered this question already.
Instead of repeating the entire process, it could retrieve the existing answer.
Why the 0.93 Threshold Is Interesting
One particularly interesting detail about the RESONEO findings is the 0.93 cosine similarity threshold. That’s a very high number (and it should be high). It suggests that OpenAI’s experimental system isn’t intended to treat vaguely related questions as interchangeable. The prompts need to be extremely similar.
Take the question from above “Where can I find the highest quality peptides” and now compare that to “How do I ensure I’m using the highest quality peptides?” and it’s a different question. Those questions are obviously related, but they ask for different information and may not clear a threshold intended for near-equivalent prompts. The exact similarity score would depend on the embedding model and implementation, but this wouldn’t trigger the same response.
That distinction matters because semantic caching creates an obvious potential problem from an answer engine optimization perspective: if you’re not in the cached response, are you effectively locked out of that answer until the cache expires or a sufficiently different question triggers a fresh response?
In other words, even if something changes that would normally cause ChatGPT to mention your brand (a new article gets published, your reputation improves, or a source ChatGPT relies on changes), none of that may matter if the system decides it has already answered that question and simply serves the existing response again. This matters even more when the model used is not using search or query fan-outs (fyi, the free version of ChatGPT does a lot less of both of these things).
We do not yet know whether this behavior affects how quickly brand visibility changes, but it is now measurable, and worth monitoring.
Could ChatGPT Skip Search Entirely?
Imagine someone asks “What can you tell me about Company X?” ChatGPT determines that web retrieval would improve the response. It creates a series of fan-outs (Company X reviews, site:companyx.com, Company X reputation, etc.), then searches the web, collects sources and generates a detailed answer.
Now imagine another user asks an extremely similar question just 2 hours later. If ChatGPT has a valid cached answer available, there’s potentially no reason to repeat the search. The architecture could theoretically look something like:
User asks question
↓
Acceptable cached answer available?
If yes:
Return cached answer
If no:
↓
Does this question require retrieval?
If yes:
Search → retrieve sources → generate answer
If no:
Generate from model knowledge
That would mean the cache isn’t simply reducing model inference costs. It could potentially prevent the retrieval process from happening at all. It’s just a theory about how such a system could be implemented. The exposed configuration does not tell us the complete production routing architecture. But our duplicate response data raises an interesting question about whether we’re beginning to observe some form of answer reuse in practice.
The Identical Citations Are Particularly Interesting
We also need to talk about the identical citations. You could see it in the screenshot above, but some duplicate responses aren’t only identical in their written answer. Their citations are identical as well.
That’s important because citations are produced as part of the retrieval and answer-generation process. If ChatGPT independently performed a new search and generated a new response on two separate days, it certainly could retrieve the same sources and construct the same answer.
But producing thousands of characters of byte-identical text, with the same structure, wording and citations, makes the result notable. Again, we’re not saying this is proof of caching. There could be other mechanisms responsible for deterministic or reused responses.
But it’s exactly the kind of behavior we’d expect to investigate when looking for evidence of answer reuse.
Is your brand’s ChatGPT visibility changing—or simply persisting? Hoverboard tracks repeated responses, brand mentions, citations and query fan-outs over time, then turns those patterns into concrete actions for improving AI visibility. Click here to book a demo.
Why Would OpenAI Want to Do This?
OpenAI is projected to spend $600 billion on computing power through 2030, according to Reuters. With over one billion users, you can imagine how many people ask ChatGPT the same questions every day:
- How do I reset my iPhone?
- What are the best credit cards for travel?
- What is intermittent fasting?
- What are the best running shoes?
- Is Company X reputable?
With one billion users, you’re talking enormous compute power needed here, so lots of cost involved. If ChatGPT can determine with extremely high confidence that it already has a satisfactory answer, generating a completely new response every time may be unnecessary.
Even if semantic caching were only appropriate for a small percentage of requests, the savings at ChatGPT’s scale could potentially be substantial. And our early data is interesting in that context. We’re currently observing duplicates in considerably less than 1% of responses. A system wouldn’t need to cache the majority of ChatGPT requests to produce meaningful infrastructure savings.
Also keep in mind that ChatGPT uses different models with different capabilities for free and paid users. ChatGPT routes requests across models and product configurations with different computational costs. If answer caching is partly intended to reduce inference or retrieval costs, high-volume free-tier traffic would be a logical environment on which to test it. We do not yet have comparative duplicate-response data from paid accounts. If caching is partly intended to reduce inference and retrieval costs, free-tier traffic would be a logical place to test and use it.
For brands seeking to maximize consumer visibility, ChatGPT’s free tier deserves the most focus and attention. OpenAI reported more than 50 million consumer subscribers in February 2026 and more than one billion active users in July. Those figures are not directly comparable measures, but they indicate that paid consumer subscribers represent only a small fraction of ChatGPT’s overall audience.
Caching Could Create a New Metric: Persistent Visibility
Consider a commercially valuable query:
“What are the best personal injury law firms in Los Angeles?”
Imagine ChatGPT generates an answer recommending Firm A, Firm B, Firm C and Firm D.
If someone asks a similar question the following day and ChatGPT generates a fresh response, the recommendations may change. Firm A might disappear, Firm E might appear, or the order might shift. That is the probabilistic behavior marketers generally expect from large language models.
Semantic caching introduces another possibility. If the second question is sufficiently similar and ChatGPT considers the previous answer reusable, the original recommendations could persist.
Firm A would no longer be appearing in only one generated response. It could be included in an answer that is shown repeatedly. That is good news for Firm A…and bad news for every relevant competitor left out of the reusable answer.
This introduces what could become an important AEO metric: persistent visibility.
Marketers currently ask:
Does ChatGPT mention my brand when someone asks this question?
They may also need to ask:
Once my brand becomes part of an answer, how long does that answer persist?
If every response is generated independently, a brand effectively competes for inclusion each time. But if some answers are reused, one successful inclusion could influence multiple subsequent responses.
The opposite matters just as much. A company could improve its website, publish stronger content and earn new authoritative mentions without immediately affecting the answer users receive. The underlying information environment may change before a reusable AI answer changes.
Caching could therefore create three distinct forms of AI visibility:
Model visibility
The model generates the brand from information already represented in its learned knowledge.
Retrieval visibility
The system searches or retrieves current information that contributes to the brand appearing in a newly generated response.
Persistent visibility
A previously generated answer containing the brand is reused without necessarily performing fresh retrieval for that interaction.
To the consumer, the outcome looks the same: the brand appears in ChatGPT. But the mechanism behind that appearance (and the actions required to influence it) could be completely different.
This Also Challenges the “AI Is Too Random to Track” Argument
There’s an interesting irony here. One criticism we frequently hear about AI visibility tracking is:
LLMs are probabilistic. You can ask the same question repeatedly and receive different answers. Therefore AI visibility tracking isn’t meaningful.
Probabilistic systems are measurable through repeated observations. That’s exactly why Hoverboard tracks prompts consistently over time rather than treating a single ChatGPT response as definitive.
But answer caching could push parts of the AI ecosystem in the opposite direction. It could make some responses less probabilistic. Imagine 1,000 users asking variations of:
What is the best CRM software for small businesses?
With independent generation, brand visibility exists as a probability distribution. One company might appear 73% of the time. Another might appear 41%. Another might appear 17%.
But if groups of semantically equivalent questions resolve to reusable answers, some portions of that distribution could become significantly more stable. The future AI ecosystem may therefore contain both:
Probabilistic visibility
and:
Persistent visibility.
Both can be measured.
They just need to be measured differently.
Can We Detect AI Answer Caching From the Outside?
OpenAI doesn’t need to tell us whether a particular response came from a semantic cache for us to investigate the behavior statistically. At sufficient scale, we can look for signals.
For example:
- Byte-identical responses across different days
- Identical citations
- Identical formatting and ordering
- Clusters of semantically similar prompts receiving identical answers
- Sudden changes in duplicate-response frequency
- Responses that remain unchanged despite changes in available web information
- Duplicate behavior correlated with particular prompt categories
- Duplicate behavior correlated with particular ChatGPT models or account tiers
None of these individually proves caching. But together, across hundreds of thousands or millions of observations, they could show us patterns that aren’t visible from manually testing ChatGPT a handful of times.
That’s something we intend to continue studying.
AI Visibility Tracking Cannot Simply Become Traditional Rank Tracking
Traditional search gave marketers a relatively simple mental model:
Query → ranked search results
AI answer engines are fundamentally different.
There can be:
- Search routing
- Retrieval systems
- Rerankers
- Model knowledge
- Tool calls
- Model selection
- Experiments
- And potentially semantic answer caches
The response the user sees is the final product of that entire system.
For marketers, this means AI visibility isn’t determined by one ranking algorithm. It’s determined by a pipeline. Understanding that pipeline is going to become increasingly important as consumers move more product discovery, professional research and purchasing decisions into AI assistants.
Methodology: Hoverboard tracked 3,495 unique prompts once per day from July 1 through August 10, 2026, producing 143,295 ChatGPT responses across the 41-day observation period. Every prompt was submitted to the free version of ChatGPT in a non-logged-in session. This was intended to reduce the influence of account-level personalization, saved memories and previous conversation history. Each prompt was monitored repeatedly so that responses to the same prompt could be compared across different days. Responses were compared only with earlier responses generated for the same prompt. We classified a response as a duplicate when its complete answer text matched an earlier response byte for byte. Similar responses with changes in wording, formatting, ordering or citations were not counted as duplicates. When the same answer appeared multiple times, the first occurrence served as the baseline and each subsequent byte-identical occurrence was counted as a duplicate response.
We’re continuing to measure duplicate-answer frequency across models, prompt categories and account tiers. If you want to monitor the answers affecting your own brand, start tracking them in Hoverboard AI today.