GPT-5.6 Sol vs GPT-6.1 Sol: How ChatGPT’s New Model Searches and Cites
Abstract
GPT-6.1 Sol runs 45% fewer web searches than GPT-5.6 Sol for the same prompts, yet cites the market leader’s own website almost twice as often. We ran 40 identical commercial prompts about online payments through both models on October 12, 2026, and logged every fan-out query, cited URL and brand mention.
Fan-out queries per prompt fell from 8.6 to 4.7. Citations per answer held steady (4.9 vs 5.05), but came from fewer places: 58 distinct domains instead of 73. Stripe-owned pages rose from 23.5% to 45.0% of all citations. Answers got 34% shorter and named 40% fewer brands, and every one of the 40 answers ended by asking the user a clarifying question (up from 5 of 40).
Some common assumptions about newer models did not hold in this category: queries did not get longer, comparison and question phrasing never appeared, and listicles and editorial sites were not cited at all.
1. Introduction
Every new ChatGPT model changes how it searches the web, and with it, which brands get seen. When ChatGPT answers a question that needs current information, it rewrites the prompt into several web searches, called fan-out queries. It reads the results and cites a handful of pages. A brand is visible only if its pages are among those results and those citations.
The model decides what to search, so a model switch can quietly change the rules. A brand that was well cited under one model can lose visibility under the next without changing anything on its own site. This study measures that change for one category, comparing GPT-5.6 Sol with its successor, GPT-6.1 Sol.
Research questions
How does the number and length of fan-out queries change from GPT-5.6 Sol to GPT-6.1 Sol?
How does the wording of those queries change?
How many sources are cited, and what kinds of sites are they?
How much do the two models agree on which domains and URLs to cite?
How does brand visibility shift, for the market leader and its competitors?
2. Methodology
We sent the same 40 prompts to both models on the same day, through the same pipeline, and compared every fan-out query, citation and brand mention they produced.
Prompt set. Forty unbranded, commercial-intent prompts in the online payments category, eight in each of five sub-categories: Subscription Billing, Developer Payment APIs, Online Payment Processing, Global and Cross-Border Payments, and Marketplace and Platform Payments. The prompts mix “best X” questions (“best payment API for developers”) with how-to questions (“how to split payments between multiple parties”). None name a brand. Stripe is the tracked brand. The full list is in Appendix A.
Runs. All runs used XLR8 AI’s visibility tracking platform, with web search left to the model’s discretion and no location set.
Run | Model | Date | Prompts | Role in this study |
|---|---|---|---|---|
A | GPT-5.6 Sol | Oct 12, 2026 | 40 | Baseline |
B | GPT-6.1 Sol | Oct 12, 2026 | 40 | Test |
C | GPT-5.6 Sol | Aug 26, 2026 | 40 | Same-model drift check |
The headline comparison is A vs B. Running them a minute apart means the web they searched was effectively the same, so differences come from the model. Run C, an earlier run of GPT-5.6 Sol on the same prompts, shows how much one model varies on its own over seven weeks. That sets a bar the A vs B differences need to clear.
What we measured.
Fan-outs: number per prompt; words per query; share of queries that target one domain, use “official”, include a year, name a brand, use “vs” or “compare”, or are phrased as a question.
Citations: number per answer; distinct domains and URLs; source type.
Agreement: per-prompt overlap of cited domains and URLs between runs (Jaccard index: shared items divided by all items).
Answers: length in characters; brands named; whether the answer ends by asking the user a question.
Source types. Each cited domain was placed in one of five groups:
Stripe-owned (stripe.com, docs.stripe.com, support.stripe.com).
Competitor and vendor sites (PayPal, Adyen, Square, Paddle, Chargebee and other payments or billing vendors).
Government and standards bodies (IRS, CFPB, EU institutions, PCI Security Standards Council, FCA and similar).
Platform and developer ecosystems (Apple, Google, GitHub, Shopify, Wix).
Third-party editorial and user-generated content (reviews, media, forums, community sites).
Each URL was also tagged by page type from its path: docs or help, pricing, blog or guide, and other.
3. Finding 1: Fewer searches, but not longer ones
GPT-6.1 Sol ran 4.7 searches per prompt against GPT-5.6 Sol’s 8.6, a 45% drop, and searched less on 36 of 40 prompts. The searches did not get longer to make up for it.
Figure 1. Mean fan-out queries per prompt, by sub-category. Source: XLR8 AI visibility platform, Oct 12, 2026.
Metric | GPT-5.6 Sol | GPT-6.1 Sol | Change |
|---|---|---|---|
Fan-out queries, total | 345 | 189 | −45% |
Fan-outs per prompt (median) | 8 | 4 | −50% |
Most fan-outs on one prompt | 21 | 10 | −52% |
Prompts with no search at all | 0 | 1 | +1 |
Words per fan-out (mean) | 6.8 | 7.1 | +0.3 words |
Fan-outs with 11+ words | 4.6% | 4.2% | flat |
The cut was broad: every sub-category lost between 37% and 51% of its searches. Global and Cross-Border Payments, the most search-heavy group, fell from 10.5 to 5.6 per prompt.
The one prompt with no search was “how to split payments between multiple parties”. Instead of searching, GPT-6.1 Sol replied with a clarifying question: did the user mean splitting a bill with friends, paying several sellers, or paying with two cards? It shows the new model choosing to ask rather than search when a prompt is ambiguous (see Finding 5).
What this means. With fewer searches per answer, each one carries more weight. A page that ranked for the fifth or eighth search under GPT-5.6 Sol may now never be fetched at all.
4. Finding 2: Still searching domain by domain, but dropping “official”
Nine in ten GPT-6.1 Sol searches still targeted one company’s website, as under GPT-5.6 Sol. What changed is how those searches are written: the site: operator gave way to a plain-text form, and the word “official” nearly disappeared.
Share of fan-out queries that… | GPT-5.6 Sol | GPT-6.1 Sol |
|---|---|---|
Target one domain (either form) | 91.6% | 91.5% |
…using site:domain.com | 91.6% | 1.1% |
…using site.domain.com | 0% | 90.5% |
Contain “official” | 68.1% | 2.6% |
Name a brand | 67.0% | 71.4% |
Name Stripe | 28.4% | 39.2% |
Ask about price, fees or rates | 26.7% | 27.0% |
Include a year (2024–2027) | 4.9% | 3.2% |
Contain “best” or “top” | 0.6% | 2.1% |
Contain “vs”, “versus” or “compare” | 0% | 0% |
Are phrased as a question | 0% | 0% |
Side by side, for the prompt “best payment API for developers”:
GPT-5.6 Sol: site:stripe.com/docs payments API pricing developers Stripe official
GPT-6.1 Sol: site.stripe.com pricing payments API United States 2.9 30
Both go to Stripe’s site for pricing. The new query drops “official” and adds the concrete figures it expects to find on the page (the 2.9% + 30¢ card rate).
Three observations:
The domain-first strategy survived. Counted the old way, site: queries collapsed from 91.6% to 1.1%. Counting the new site. form, domain targeting did not fall at all. We report queries as the platform logged them; whether the search backend treats site. as an operator or as plain text cannot be told from this data, but the intent is the same.
“Official” disappeared. GPT-5.6 Sol attached it to most of its site: queries. GPT-6.1 Sol uses it in fewer than 3 in 100.
No comparison or question phrasing appeared. Neither model wrote a single “vs”, “compare” or question-style query. GPT-6.1 Sol builds its own comparisons by visiting each vendor’s site directly, rather than searching for someone else’s comparison.
Stripe’s share of fan-outs rose from 28.4% to 39.2%. With fewer searches to spend, GPT-6.1 Sol spent a larger part of them on the category leader.
5. Finding 3: Same number of citations, far more of them from Stripe
GPT-6.1 Sol cited about as many sources per answer as GPT-5.6 Sol (5.05 vs 4.9), but nearly half of them were Stripe’s own pages, up from under a quarter.
Figure 2. Share of cited URLs by source type. Source: XLR8 AI visibility platform, Oct 12, 2026.
The citations came from a smaller pool: 58 distinct domains instead of 73, while the number of distinct URLs stayed at 166. GPT-6.1 Sol read more pages on fewer sites, and the site it read most was Stripe’s. Competitor and vendor sites lost 13 points of share, and government and standards bodies lost 5.5.
The kind of page cited barely moved. Docs and help pages were about 39% of citations under both models, pricing pages 17–18%, and blog or guide pages 7–9%. The shift is in whose pages get cited, not which kind.
Neither model cited a listicle, review site, forum or news outlet in any meaningful number: two third-party citations under GPT-5.6 Sol (1.0%, both to an unofficial Stripe help site) and none under GPT-6.1 Sol. In this category, for these prompts, ChatGPT goes straight to the vendors.
6. Finding 4: The two models agree on less than half their sources
For a given prompt, the two models shared 45% of their cited domains and only 16% of their exact URLs. Across the whole run, GPT-6.1 Sol dropped 30 domains that GPT-5.6 Sol had cited and added 15 new ones; 43 domains appeared in both.
Comparison | Domain overlap per prompt | URL overlap per prompt |
|---|---|---|
GPT-5.6 Sol vs GPT-6.1 Sol (same day) | 0.45 | 0.16 |
GPT-5.6 Sol, Aug vs Oct (same model, 7 weeks apart) | 0.53 | 0.16 |
Overlap is the Jaccard index: shared items divided by all items cited for that prompt by either run. The drift row matters. The same model, asked the same questions seven weeks apart, kept only 16% of its exact URLs, so URL-level churn is normal whether or not the model changes. At the domain level, the switch to GPT-6.1 Sol caused more churn than seven weeks of drift (0.45 vs 0.53).
Biggest gains and losses in citations
Domain | GPT-5.6 Sol | GPT-6.1 Sol | Change |
|---|---|---|---|
stripe.com | 19 | 40 | +21 |
docs.stripe.com | 27 | 45 | +18 |
support.stripe.com | 0 | 6 | +6 |
github.com | 1 | 4 | +3 |
irs.gov | 3 | 5 | +2 |
payoneer.com | 0 | 2 | +2 |
docs.adyen.com | 17 | 9 | −8 |
developer.paypal.com | 8 | 3 | −5 |
pcisecuritystandards.org | 6 | 1 | −5 |
paypal.com | 11 | 9 | −2 |
eur-lex.europa.eu | 3 | 1 | −2 |
The gains go almost entirely to Stripe’s own pages. The losses fall on competitors’ developer docs (Adyen, PayPal) and on standards and regulatory sources. The GitHub gain is Stripe’s too: all four of its citations point to Stripe’s own repositories (the React Native SDK, the Stripe CLI and Stripe’s AI best-practices guide).
7. Finding 5: Shorter answers, fewer brands, and a question at the end of every one
GPT-6.1 Sol’s answers were 34% shorter and named 40% fewer brands. Stripe kept its place while competitors lost mentions, and every answer now ends by asking the user for more detail.
Answer metric | GPT-5.6 Sol | GPT-6.1 Sol | Change |
|---|---|---|---|
Average answer length (characters) | 4,571 | 3,002 | −34% |
Brands named per answer | 7.2 | 4.3 | −40% |
Answers that mention Stripe | 33 of 40 | 34 of 40 | +1 |
Answers whose first citation is a Stripe page | 4 of 40 | 18 of 40 | +14 |
Answers that end with a question to the user | 5 of 40 | 40 of 40 | +35 |
Figure 3. Number of answers (of 40) mentioning each brand. Source: XLR8 AI visibility platform, Oct 12, 2026.
The shortlist got shorter, and the leader stayed on it. For “how to accept international payments online”, GPT-5.6 Sol named nine providers and platforms; GPT-6.1 Sol named three (Stripe, PayPal, Wise). For a challenger brand, being the fourth or fifth name in an answer used to mean being mentioned. Now there may be no fourth or fifth slot.
Every answer ends with a question. GPT-6.1 Sol closed all 40 answers with a clarifying question tailored to the prompt, for example:
“Are you building SaaS, e-commerce, or a marketplace, and where is your business registered?” (best payment API for developers)
“About how much do you process monthly, what’s your average sale, and are payments mostly in person, online, or invoices?” (cheapest payment processing rates for small business)
“Are you B2B or B2C, what’s your approximate monthly revenue, and do you charge flat-rate, per-seat, or usage-based?” (best subscription billing software for SaaS)
The answer to the first prompt is now a draft. The next turn, shaped by the user’s reply about country, volume or business model, is where the final recommendation is made, and it gets its own searches.
8. Discussion: What this means for GEO
GPT-6.1 Sol behaves like a buyer with less time: it visits fewer sites, trusts the category leader’s own pages more, and asks the user to narrow the question before committing. In this category, that favors the incumbent and squeezes everyone else.
Implications for brands
Your own site is the citation surface. In this category, 82% of GPT-6.1 Sol’s citations went to vendor-owned pages (Stripe’s plus competitors’). Docs, pricing pages and help-center articles are what get read and cited, so they need to state facts plainly: rates, supported countries, features, limits.
Write the numbers the model is searching for. GPT-6.1 Sol’s queries carry concrete figures (“2.9 30”, “0.7% pay as you go”, “5% 50 cents”). Pages that state current numbers in plain text match those queries; pages that hide them behind “contact sales” do not.
Fewer slots favor the leader. With 4.3 brands per answer instead of 7.2, a challenger needs to own a specific use case (merchant of record, interchange-plus pricing, contractor payouts) where it is clearly the best answer, rather than hoping to appear on a long list.
Plan for the second turn. Every answer now ends with a question about business model, country or volume. The follow-up is where the final recommendation happens. Content that answers “best X for a B2B SaaS in the EU at $50k a month” is positioned for that turn.
Track continuously, at the URL level. Even with no model change, only 16% of cited URLs survived seven weeks. One-off audits go stale fast.
9. Limitations
These results describe one category, 40 prompts and one run per model. They show direction, not universal rates.
One category, one brand. Payments is documentation-heavy, with a clear market leader. Categories driven by reviews (consumer products, travel, local services) may behave very differently.
Single run per model. LLM output varies between runs. GPT-5.6 Sol itself ran 25% fewer searches and 23% fewer citations in October than in August. The GPT-6.1 Sol changes are larger than that drift on most measures, but not all: URL-level overlap (16%) is no different from drift.
Pipeline, not the ChatGPT app. Runs were made through XLR8’s tracking platform. Logged fan-outs and answers may differ from what a signed-in user sees in ChatGPT, which can add memory, personalization and interface features.
Fan-outs as logged. We report fan-out strings exactly as recorded and cannot see whether the search backend treats site.domain.com as an operator.
Heuristic classification. Source types, page types and “ends with a question” were classified by rules on domains, URL paths and answer text. Spot checks agreed, but some edge cases may be misfiled.
No location set. Results may differ for users in specific countries.
10. Conclusion
GPT-6.1 Sol searches less and concentrates what it cites. In online payments, that concentration went to the category leader’s own pages: Stripe’s share of citations nearly doubled, to 45%, while competitors lost mentions and shortlist slots.
The model still builds answers by going vendor by vendor, but with half the searches, a third less text and a question at the end of every reply. The safest conclusion for any brand is not a new checklist. It is to measure its own category on each new model, as soon as that model ships.
A natural next step is to re-run this prompt set on future ChatGPT releases and extend it to review-driven categories, to test whether the leader effect holds where third-party content matters more.
