Running the same query in ChatGPT's free Think mode, paid Thinking mode, or an instant answer may not produce the same web research. Resoneo replayed its July prompts in August and found that modes did more than spend different amounts of time reasoning. They could draw from different source pools, retrieve different numbers of URLs, and select different sets of domains.
The finding means a single screenshot cannot answer whether a brand is visible in ChatGPT. A company may be absent from a free instant answer and become a source in paid thinking. Automated monitoring through the API may also fail to predict the consumer product reliably. These figures are not official technical disclosures from OpenAI. They are independent observations based on network traffic and should not be treated as permanent quotas or ranking rules.
What changed in the August replay?
Resoneo reran the same test prompts it had used in July across free and paid accounts and Instant, Think, and Thinking configurations. Free Think collected an average of 35.3 URLs and 16.3 distinct domains per conversation. July's free instant path had averaged 15.1 URLs and 9.8 domains. Source volume had nearly reached paid Thinking at medium effort.
Similar volume did not mean the same source route. Under the researchers' classification, 74.7% of free Think URLs came from “labrador”, believed to be OpenAI's in-house index, 3.1% from scraped classic Google results, and 22.2% from a news-provider path. Paid Thinking at medium effort was almost the reverse: 75.3% of URLs were classified as scraped Google results and 24.7% as the in-house route.
“Labrador” is a name inferred by Resoneo from network markers. OpenAI does not document an index by that name or confirm the shares above. The “scraped Google” classification is likewise a technical interpretation based on request and response patterns. The finding should be understood as repeatable but externally observed behaviour, not a verified product architecture.
Does more thinking always mean more sources?
The August data also breaks a simple rule that a higher mode always retrieves more material. In paid Thinking at medium effort, average fan-outs fell from 3.56 to 1.90, URLs from 47.2 to 33.9, and distinct domains from 21.9 to 15.5. At advanced effort, URLs dropped from 67 to 40.5 and domains from 29.5 to 14.7.
Meanwhile, 96.7% of free Instant answers produced a reasoning recap in August, compared with 0.6% in July. The background flow can therefore change within weeks even under a familiar user label. A model name or subscription tier alone does not explain which web route is being used.
The study also observed large differences in site-operator use. Across the test prompts, site: queries represented 3.5% on free Instant, 10.8% on free Think, 41.9% on medium Thinking, and 58.1% on high Thinking. High effort rose from 40.8% in late July to 58.1% in mid-August. The researchers interpret this as a more navigational pattern: fewer queries aimed more directly at expected brands and official domains.
Why may API monitoring be a poor proxy for the product?
Resoneo compared the same prompts across the ChatGPT product and API models. Jaccard similarity between the sets of brands that appeared together was only 0.23 to 0.27. Values near zero indicate limited overlap. In this sample, the brands present in API output were a weak predictor of those shown in the consumer product.
That does not make API monitoring useless. The API offers a controlled and repeatable laboratory. The mistake is labelling its output as a measurement of “our actual share in ChatGPT”. Unless plan, mode, location, account state, and date in the consumer product are also sampled, the result describes one model endpoint.
Our earlier Fark Studio analysis examined how smaller sites could appear in ChatGPT's in-house search index and how page openings might be used in citation context. The August update introduces a different problem. Even when content is technically accessible, the source pool selected by a mode may alter whether it appears. Accessibility and selection are not the same metric.
What does OpenAI officially say?
OpenAI's current ChatGPT Search help page says search is available to free and paid users, prompts may be rewritten into one or more targeted queries, and third-party search providers can be used in some cases. The official page explains that ranking uses multiple factors intended to provide reliable, relevant information and that there is no guarantee of top placement.
The official, actionable requirement for site owners is OAI-SearchBot access. If robots.txt, a CDN, or a firewall blocks the bot, a site may not be included in ChatGPT Search. OpenAI does not disclose mode-level URL counts, an index called “labrador”, Google source shares, or the ranking factors. That boundary is important when separating news from researcher inference.
Who is affected in Türkiye?
SEO and content teams, reputation managers, publishers, e-commerce businesses, and local operators are first in line. Turkish-language queries may use a different source pool from English prompts. Location can influence local results. Free-user behaviour is not necessarily represented by an enterprise team's paid account.
There is a more direct risk for teams buying or selling AI-visibility reports. A tool that uses only an API and fails to disclose that method can make a claim that is too broad for consumer ChatGPT. A manual audit based on one plan and one mode has the opposite blind spot. Without language, city, date, and session state, comparisons between brands in Türkiye are not reliable.
What is confirmed and what remains uncertain?
The confirmed finding is that Resoneo replayed the same test prompts in July and August and reported differences in URLs, domains, and source classes across modes in its sample. OpenAI officially confirms that ChatGPT Search can rewrite queries, use third-party search partners, and requires OAI-SearchBot access for inclusion.
The uncertainties are larger. The sample cannot represent every Turkish prompt, account, or user cohort. Pipeline labels inferred from network markers are not official. Model routing, experiment cohorts, and source partnerships can change after publication. The observed percentages are not product commitments or ranking weights.
What should teams do now? A mode-aware visibility audit
1. Divide the prompt library by intent. Keep brand, category, comparison, local discovery, and problem-solving prompts in separate groups.
2. Record the test matrix. Store date, time, language, country or city, plan, mode, session state, and device with every result.
3. Sample free and paid paths. Compare at least Instant, free Think, and paid Thinking where available during the same period.
4. Label the API as a separate benchmark. Do not merge API output into the same chart as product results. Calculate overlap separately.
5. Separate source route from answer presence. Record the URL, source domain, citation position, and answer context, not only whether the brand was mentioned.
6. Verify technical access. Check robots.txt, CDN, WAF, and logs for OAI-SearchBot. State clearly that access does not guarantee top ranking.
7. Avoid decisions from one observation. Repeat prompts on different days and keep visible timestamps because the product flow can change quickly.
8. Connect to business outcomes. Interpret citations and mentions alongside referral traffic, qualified visits, demand, and conversion signals.
This process keeps SEO, content production, and digital marketing teams from overloading one “AI visibility” score. A measurement vendor should state exactly which surface it tested.
Where should brands act, and where should they wait?
Act now if OAI-SearchBot is blocked, source pages are stale, or a visibility report relies only on the API. Fix access, add method labels, and report modes separately. For business-critical prompts, establish a recurring sample from the actual product.
Wait before treating a 74.7% in-house share or a 75.3% Google-path share as permanent. Do not rebuild content formats or budgets around those ratios. Do not present undocumented pipeline names as official OpenAI features. First test whether the direction repeats within your own prompt, language, and market sample.
Fark Studio perspective
In AI-search visibility, inaccurate measurement can be as damaging as invisibility. A result from one mode is not a photograph of the entire ChatGPT experience. A large API panel can be valuable because it is controlled, but it is not a mirror of the consumer product.
Fark Studio's durable approach is to measure technical access separately from answer selection and attach a method label to every observation. Plan a mode-aware AI-search visibility audit with Fark Studio to combine your prompt library, product and API samples, source relationships, and business outcomes in one measurement architecture.
Sources
Resoneo, What ChatGPT pulls, what it shows, what it cites, July 2026, updated August 2026.
Search Engine Land, Inside ChatGPT's retrieval stack: The index, cache, and pages it actually reads, August 17, 2026.
OpenAI Help Center, ChatGPT Search, updated August 14, 2026.
