How Often Does ChatGPT Update Its Information Sources: Results of Independent Experiment #10

A common situation: you ask ChatGPT about the best noise-cancelling headphones and get a set of results. Will you receive the same recommendations if you repeat the query a week later? Will the model change its output? And if so — by how much?

We decided to find out. Not just once, but systematically — every week for a month, from different accounts, in incognito mode, and with a US geolocation.

Methodology: How We Tested

10 queries — from informational to commercial:

1. What do you know about Headway Inc company?
2. What do you think about iPhone 16 Pro Max?
3. How to build AI applications for healthcare
4. Recommend a company that provides dedicated development teams
5. Where to buy iPhone 17 Pro Max?
6. MacBook Air vs MacBook Pro for students
7. Best noise-cancelling headphones under $200
8. Best UX/UI design agency for startups
9. Dedicated software development team providers
10. What is the best VPN for remote work?

4 account types: 2 logged-in accounts + 2 incognito sessions

Frequency: 1 check per week from each account = 4 checks per week over one month

What we tracked: sources cited in ChatGPT’s responses and how they changed week to week

Test Results

How often does ChatGPT update its sources for the same query

Parameter	Observations	Conclusion
Rate of change	From 25% (VPN, iPhone) to 70% (dedicated teams, headphones)	The model updates its sources in proportion to how much the market changes. For stable topics (Apple products, VPN) — almost always the same sites. For commercial/service queries — a fresh selection each time.
Core stability	30–70% — always authoritative sites (Apple, FDA, Tom’s Guide, TechRadar)	The model reliably maintains a “gold standard” of sources. Core conclusions can be trusted.
Account type / incognito impact	Minimal (5–10%) in most cases. Noticeable (15–30%) only for commercial purchases and services	A logged-in account skews slightly more official and premium. Incognito yields more niche options. The difference is not critical.
Overall trend	The more commercial/price-driven the query — the higher the variability. The more informational/evaluative — the more stable.	The model behaves predictably: for facts/opinions — very reliable; for purchases/services — a “fresh snapshot” each time.

Research Observations

The model does not serve random sources. Nor does it repeat the exact same ones. It balances between a stable core and a shifting context.

Three Speeds of Source Updates

Slow change (25–40% updates)

Queries where the industry has reached a consensus:

• “What is the best VPN for remote work?” — ~25–35%
• “What do you think about iPhone 16 Pro Max?” — ~30–40%

Why so few changes? Market leaders have been fixed for years. NordVPN, ExpressVPN, ProtonVPN — their positions are stable. Reviews update rarely because the product doesn’t change radically month to month.

Moderate change (35–55% updates)

Most queries fall here:

• MacBook comparison — ~35–45%
• UX/UI agencies — ~40–55%
• AI in healthcare — ~40–50%
• Dedicated teams — ~45–60%

Why moderate? Annual product cycles and fresh rankings, but authoritative sources (TechRadar, Tom’s Guide, FDA) remain unchanged.

Rapid change (50–70% updates)

Dynamic commercial markets:

• Noise-cancelling headphones — ~50–65%
• Dedicated team providers — ~55–70%
• “Where to buy iPhone 17 Pro Max?” — ~40–60% (with high reseller variation)

Why so many changes? Prices shift, new models appear, stock changes, discounts emerge — SEO competition pushes the model to seek out the most current sources.

Stable Core vs. Shifting Periphery

Regardless of query type, there is a baseline set of sources that appears again and again.

Stable sources (40–70% of responses)

• Official sites: Apple.com and other manufacturers
• Regulators: FDA, HHS (for healthcare queries)
• Authoritative review outlets: Tom’s Guide, TechRadar, CNET, MacRumors
• Communities: Reddit, occasionally Wikipedia

These sources dominate 60–70% of responses for Apple products, regulatory questions, and top-tier reviews.

Variable sources (30–60% of responses)

• Niche blogs covering specific headphone models
• Commercial aggregators (price comparison sites)
• Third-party resellers (Amazon, BestBuy, B&H Photo)
• Local startup and agency reviews

Their share grows in commercial queries: headphones, dedicated team services, purchase location searches.

Average split: ~40% stable core (authoritative sources) / ~60% variable context (current/dynamic sources)

Does Account Type Affect Sources?

We expected to see strong personalization. In reality, the difference was minimal but noticeable.

Logged-in accounts more frequently surfaced official sites (Apple, FDA), major retailers (Amazon, BestBuy), and premium media (Tom’s Guide, TechRadar).

Incognito mode produced more niche and “gray-area” sources: lesser-known resellers, academic arXiv papers, local blogs, and specialized forums.

The difference is 5–10% in most cases, rising to 15–30% for commercial queries about purchases and services. US geolocation affects all account types equally — focus on the American market, USD, HIPAA/FDA.

Study Conclusions

1. The model seeks balance The source core (authoritative reviews, official sites, regulators) remains stable (~40–60%), while the variable portion adapts for recency — new articles, pricing, stock availability.

2. Commercial queries = maximum variability If you’re optimizing content for queries like “where to buy” or “best [product] under $200,” expect stiff competition. ChatGPT actively refreshes its sources in search of current prices and availability.

3. Informational queries = maximum stability Queries like “what do you think about” or “how to build” are the most predictable. If your brand appears in a response once, it will likely stay there for a long time — provided it carries sufficient authority.

4. Personalization is not decisive The difference between a logged-in account and incognito is only 5–10%. Personalization exists, but it is far less pronounced than in traditional Google search.

What’s Next?

This research showed that ChatGPT’s output is neither static nor chaotic. It is a balancing system in which authority competes with recency.

Questions we are exploring next:

• How often does the model update sources for news queries? (Hypothesis: >80%)
• Can a brand “anchor” itself in the stable core through specific signals?
• How does user behavior (clicking, copying) affect subsequent responses?