Decoded: ChatGPT's Ranking FormulaReciprocal Rank Fusion and .Best Domains

About this paper · How to read the evidence

This paper examines one specific claim about AI search: that ChatGPT merges the results of its hidden searches with a classic formula called Reciprocal Rank Fusion (RRF), and that this formula decides which pages are in the running to be cited. It explains the formula in detail, sets out how strong the evidence for the claim is, and shows where a .Best domain fits.

The subject needs care. RRF itself is thoroughly documented: it was published in 2009 and ships today in Elasticsearch, OpenSearch and Azure AI Search. Its use inside ChatGPT is not documented by OpenAI. It rests on one independent researcher's reading of ChatGPT's web client in July 2025, repeated and extended by others since. We present that claim at full strength and with its limits, because a strategy built on it should know both.

The title deliberately echoes the article that started the discussion, Metehan Yesilyurt's “Is RRF the Secret to Dominating AI Citations? I Decoded ChatGPT's Ranking Formula” (July 2025). This paper builds on his finding and on the 2026 research that followed; it does not claim the discovery.

The five evidence tags

Every claim carries one tag, so you always know how firm the ground is.

Disclosure: this paper is published by the registry operator of .Best. Third-party figures are reported as published by their sources, listed at the end; several were read through secondary coverage because the original pages were not reachable when this paper was prepared. Readers should consult the originals before quoting a figure. AI search changes quickly; every figure is dated.

01 · ChatGPT does not rank one list. It may fuse several.

When ChatGPT searches the web, it rewrites the buyer's question into one or more searches of its own. Each search returns a ranked list. Something has to turn those lists into the handful of sources the answer cites. The leading public hypothesis is Reciprocal Rank Fusion, a formula that rewards pages for appearing in many lists more than for topping any one of them.

Six findings

1. ChatGPT turns one question into several searches. OpenAI says ChatGPT “typically rewrites” a query into “one or more targeted queries”. Independent 2026 measurements put the average around two per prompt, and far more for complex or reasoning-mode questions. MEASURED · GENERAL
2. Several lists need a fusion rule, and RRF is the standard one. It scores each page by summing 1/(k + rank) across the lists it appears in, with k = 60 the default in Elasticsearch, OpenSearch and Azure AI Search. MEASURED · GENERAL
3. RRF in ChatGPT is reported, not confirmed. RRF settings were observed in ChatGPT's web client in July 2025; one practitioner reproduced the observation; OpenAI has never described its ranking in these terms, and k = 60 was assumed, not seen. REPORTED
4. Under RRF, breadth beats peak. At k = 60, a first place is worth only 14.8% more than a tenth place in the same list, while each additional list a page appears in adds a full new term. Ranking #20 in four searches scores three times a single #1. MECHANISM
5. The searches speak the language of “best”. “Best” is the word ChatGPT adds most often to its fan-out searches, in 24.3% of searches for advice questions, alongside “top”, “comparison” and “reviews”. MEASURED · GENERAL
6. A .Best name repeats that word in every list. On a .Best site, the domain, the brand mention and many titles carry “best”, so every “best” sub-query finds the same entity again. RRF rewards exactly that repetition. The name cannot make a weak page relevant; it helps a relevant one be the same, recognisable answer everywhere. MECHANISM

What we are not claiming. We are not claiming that OpenAI has confirmed RRF, that k = 60 is ChatGPT's setting, or that ChatGPT gives any weight to a top-level domain as such. No published study measures how often ChatGPT cites .Best sites compared with others. The argument of this paper is narrower and testable: if fusion works as reported, the pages that win are the pages that appear across many related searches, and a name that states the searches' own key word is a structural fit for that. Section 12 sets out how to test it.

The evidence at a glance

What to do

• Name. Pair your category with .best: hotels.best, crm.best, movers-lyon.best. The name should state what the site is the best answer to.
• Cover. Map the sub-queries a buyer's question fans out into, and build one strong, current page for each, so you appear in many lists rather than one.
• Measure. Rank-track the sub-queries, compute your own fused score with the formula in section 10, and judge progress against your baseline.

02 · One question, many searches

Classic search answered the words you typed. AI search answers what it thinks you meant, and to find out it runs searches of its own. That step, called query fan-out, is the reason a fusion formula is needed at all.

What the platforms say

OpenAI's help documentation describes the mechanism plainly: when ChatGPT search uses a search provider, it “typically rewrites your query into one or more targeted queries” and sends those to the providers, and it may use saved memories when rewriting. OpenAI also states that ChatGPT search draws on third-party search providers as well as content from partners. It does not describe how the results of several queries are combined. MEASURED · GENERAL

Google is explicit about its own version. Its Search Central documentation says AI Overviews and AI Mode “may use a ‘query fan-out' technique — issuing multiple related searches across subtopics and data sources — to develop a response.” MEASURED · GENERAL

What 2026 measurements show

Several independent teams have logged ChatGPT's fan-out queries at scale. Their figures differ because their bases differ (all prompts or only prompts that triggered a search; free or paid tiers; standard or reasoning models), so they should be read as a range, not one number.

Why fan-out changes the game. With one query, there is one ranking, and the top result wins. With several, a page can be first for one sub-query and absent from the rest, or solidly placed in all of them. Which of the two profiles wins depends entirely on how the lists are merged. That is the question RRF answers.

The fan-outs are not the prompt. A 2026 preprint that replayed 1,600 runs per model found no instant-model run that searched with the user's exact wording. Fan-outs add words the user never typed, such as a year, “best”, “reviews” or “official” (section 07). MEASURED · GENERAL

How many lists can a page collect?

The number of fan-out searches sets a ceiling on any page's fused score: with n searches, the most a page can earn is n first places, n/61 at k = 60. The ceiling explains why fan-out depth matters so much under fusion, and why the working threshold reported for ChatGPT (section 05) cannot be reached from a single search. MECHANISM

Deeper fan-outs widen the gap between a page that covers the cluster and a page that wins one query. As questions get harder and models search more, breadth becomes worth more. MECHANISM

03 · Reciprocal Rank Fusion, explained

Reciprocal Rank Fusion was introduced in 2009 by Gordon Cormack, Charles Clarke and Stefan Büttcher of the University of Waterloo, in a two-page SIGIR paper titled “Reciprocal Rank Fusion outperforms Condorcet and Individual Rank Learning Methods”. It solves a simple problem: how to combine several ranked lists into one, when the systems that produced them score documents on incomparable scales.

R is the set of result lists; r(d) is the position of document d in list r (1 = first); k is a constant, 60 by default. A list that does not contain d adds nothing.

Three properties that matter

1. It uses ranks, not scores. A keyword engine and a vector engine score relevance in units that cannot be compared. RRF ignores the scores and keeps only the order, so any retrieval system can be fused with any other without normalisation. Elastic's documentation sums it up: RRF “requires no tuning, and the different relevance indicators do not have to be related to each other”.
2. The constant k flattens the top. Without k, first place (1/1) would be worth ten times tenth place (1/10). With k = 60, first place is worth 1/61 and tenth place 1/70: a difference of 14.8%. The authors describe k as mitigating “the impact of high rankings by outlier systems”: one list's enthusiasm cannot outvote the others.
3. Every list adds a term. A document that appears in five lists collects five reciprocal scores; a document that appears in one list collects one. Agreement across lists, not dominance in a single list, is what accumulates.

Why 60

The value was not derived from theory. The authors fixed k = 60 during pilot experiments, fusing the results of 30 configurations of their search engine over four TREC collections, and reported that 60 was near-optimal but that the choice was not critical. It was then held constant through validation, where RRF beat every individual system and the Condorcet fusion method. Sixteen years later, 60 remains the default almost everywhere. MEASURED · GENERAL

Contribution of a single list relative to rank 1, computed from the RRF formula. At k = 60 the curve is nearly flat across the first page.

How search engines ship it

RRF is not an academic curiosity. It is the default fusion method of the three most widely used hybrid search platforms, which is one reason it is a plausible choice for any AI search pipeline that merges several queries or retrieval methods.

Weighted RRF generalises the formula: each list's term is multiplied by a weight w, so that score(d) = Σ w / (k + r(d)). Elastic's documentation gives exactly this form. A system that trusts one kind of retrieval more than another can say so without abandoning rank-based fusion.

# Reciprocal Rank Fusion, as documented by Elastic (pseudocode)
score = 0.0
for q in queries:
    if d in result(q):
        score += 1.0 / ( k + rank( result(q), d ) )
return score

What RRF is not

RRF decides the order of candidates; it does not decide relevance on its own. Pages first have to be retrieved by at least one list, and in modern pipelines the fused list is often re-scored by a reranking model before anything is shown. A 2026 industry paper from Dell Technologies found that fusion's gains in recall can be “largely neutralized after re-ranking and truncation”. Fusion opens the door; it does not guarantee the seat. Section 06 returns to this. MEASURED · GENERAL

04 · The arithmetic of a citation

The consequences of the formula are counter-intuitive enough to be worth working through. All figures in this section are computed directly from RRF with k = 60 and equal weights. They describe the formula, not any measured ChatGPT behaviour. MECHANISM

Five pages, five sub-queries

Imagine a buyer's question that ChatGPT fans out into five searches. Five pages compete. Each has a different profile across the five result lists.

Profile A is the classic SEO trophy: first place for one query. Under RRF it finishes last. Profile E would never be noticed in a rank tracker that only watches the top 10, yet it collects five small scores that add up to three times A's single large one.

The rule of thumb

The strategic consequence. Under RRF, the unit of competition is not a keyword. It is a cluster of related searches. A site that is solidly present across a cluster outranks a site that wins one query and is missing from the rest. Every additional list you appear in adds a full new term to your score; every place you climb within a list adds only a fraction of one.

How sensitive is this to k?

The flatness comes from k. If a system used a small constant, first places would matter more. The 2026 SemEval shared task offers an example: one strong team tuned k down to 10 for its hybrid retrieval. The table shows how the #1-versus-#10 premium moves with k.

Even with k = 10, breadth still wins: four mid-page appearances beat one first place. The direction of the advice does not depend on the exact constant, only its strength does. That matters, because ChatGPT's constant has not been observed.

05 · Decoded: what is claimed about ChatGPT

Everything specific about RRF inside ChatGPT traces back to one observation, made in July 2025 and discussed widely since. This section sets out exactly what was observed, by whom, and what remains inference.

The July 2025 observation

On 22 July 2025, SEO researcher Metehan Yesilyurt published a thread and an article reporting that, while inspecting ChatGPT's web client in Chrome's developer tools, he had found settings indicating that ChatGPT uses Reciprocal Rank Fusion to combine search results. Secondary write-ups of his finding list configuration values such as rrf_alpha: 1 and rrf_input_threshold: 0. REPORTED

Two points in his own account matter for anyone building on it. First, the constant k was not visible; he used 60 because no specific setting appeared in the code, and 60 is the standard default. Second, he wrote that he was not claiming to have found the exact formula, which could change. REPORTED

The follow-up: a top-n playbook

In September 2025 the same author published “The RRF Top-n Playbook”, which he labelled highly experimental. Using a browser extension that exposes ChatGPT's search path, he reported that ChatGPT retrieved between 38 and 65 results, and proposed a working threshold: a fused score of about 0.020. The arithmetic of that threshold translates into a practical rule. REPORTED

Under this reading, no page is cited on the strength of one search alone. The threshold itself is an experimental estimate.

How strong is the evidence?

OpenAI's own public statement on ranking is brief: its publisher FAQ says ChatGPT ranks results using “multiple factors intended to help users find relevant, reliable information”, and that placement is not guaranteed. Nothing in it confirms or rules out RRF. A practitioner who studied ChatGPT's network traffic in 2026 reported seeing no ranking scores in the data sent to the browser, which is consistent with fusion happening server-side, out of view. MEASURED · GENERAL

Our reading. The RRF claim is credible but unconfirmed. It is credible because fan-out demonstrably produces several lists, because RRF is the industry-standard way to merge them, and because the client-side settings point that way. It is unconfirmed because the evidence is one observation, the key constant is assumed, and OpenAI can change its pipeline at any time. The useful question is therefore not “is it exactly RRF?” but “what kind of page wins under any rank-fusion rule?”. The answer is robust across fusion methods: pages that appear across many related searches. That is the ground this paper builds on.

Why the client would carry fusion settings

A plausible explanation, offered as inference, is that ChatGPT's search tool is configured per request: the client or the orchestration layer passes parameters for how many results to fetch and how to merge them, and the merging itself runs on OpenAI's servers. That would explain why parameter names are visible while scores are not. MECHANISM

06 · From retrieval to citation

Fusion is one stage in a funnel. Pages are retrieved, fused, often reranked, and only then do a few become citations. The 2026 data shows how steep that funnel is.

Most retrieved pages are never cited

AirOps analysed 15,000 prompts that ChatGPT expanded into 43,233 queries and 548,534 retrieved pages. Only 15% of the retrieved pages were cited. A companion analysis reported that a page's position in ChatGPT's own result list was the strongest predictor of citation, falling from 58% at the first position to 14% at the tenth. MEASURED · GENERAL

Your Google rank is not your ChatGPT rank

Several studies have compared ChatGPT's citations with classic search results. The overlap is partial and unstable:

• A preprint posted in October 2026 replayed 80 questions, 20 times each, on two ChatGPT instant models. Google results plus OpenAI's own search API recovered only 28.3% and 26.2% of the cited URLs, but about half of the cited hostnames. MEASURED · GENERAL
• Ahrefs found that about 12% of URLs cited by ChatGPT, Gemini and Copilot also rank in Google's top 10 for the original prompt. MEASURED · GENERAL
• AirOps found that pages ranked first on Google for a fan-out query were cited 43.2% of the time, about 3.5 times more than pages beyond the top 20, and that 32.9% of cited pages appeared only in the results of a fan-out query, not of the original prompt. MEASURED · GENERAL

Read together, these figures fit a fusion model well. Citations correlate with ranking (position helps), but the relevant rankings are those of the fan-out queries, not of the prompt, and the same site can be cited through a different page than the one that ranks. Hostname overlap being twice URL overlap is exactly what you would expect if sites, more than individual pages, accumulate presence across lists. MECHANISM

Rerankers: the second gate

In most modern AI search systems, the fused list is passed to a reranking model that reads passages and re-scores them for the question. Peec AI's 2026 explainer puts the order of operations bluntly: if a page was never retrieved, rewriting a paragraph for a reranker cannot fix it. Fusion and reranking reward different things. Fusion rewards presence across lists; the reranker rewards a passage that answers the question cleanly. A page needs both. MEASURED · GENERAL

What the funnel means for strategy

1. Visibility is earned per cluster. A buyer's question produces a small family of searches. The pages that appear in most of that family form the candidate pool. Strategy should be planned at that level.
2. Sites, not just pages, accumulate. When different pages of one site appear in different lists, a system that groups by site or brand sees the same source again and again. A consistent, recognisable name helps that grouping be obvious. MECHANISM
3. Position still matters, less than expected. Moving from tenth to first in a list is worth having, but it is worth less than appearing in one more list.
4. The last mile is the passage. Once in the pool, the page that states the answer plainly is the page that gets quoted.

A note on volatility. Two ChatGPT runs of the same question often cite different pages. A 2026 preprint comparing engines found exact-URL overlap below 1% across engines for the same prompt. Single checks are anecdotes; only repeated measurement over many runs says anything. Section 10 builds that into the method.

07 · “Best” across the fan-out

Under fusion, what matters is which lists a page appears in. Which lists exist depends on which words ChatGPT adds when it rewrites a question. The most frequent of those words is “best”.

The words ChatGPT adds

Peec AI analysed about 5 million fan-out queries collected between 1 and 21 April 2026. The word ChatGPT most often added to its hidden searches was “best”: it appeared in 24.3% of fan-outs generated for advice-style prompts, where the user had not typed it. Other frequent additions were “top”, “comparison”, “reviews”, “tools”, “software” and “features”. MEASURED · GENERAL

A vocabulary that moves

The vocabulary is not fixed. Peec AI data discussed by Lily Ray indicates that after a mid-2026 model update, “best”, “top”, “vs” and “reviews” became less frequent while site: searches and “official” surged, and the listicle share of citations fell from 15.77% to 7.80%. Ahrefs found “official” and the current year rising through 2026. MEASURED · GENERAL

Two signals, one direction. The 2026 shift is sometimes read as “best is fading”. A closer reading is that ChatGPT is looking for two things at once: the best options for a need, and the official source for each option. The first produces “best” searches; the second produces brand and site: searches. A site whose own name is the category plus “best” is addressed by both kinds of search at once. That is the subject of the next section. MECHANISM

What a “best” fan-out looks like

Why this matters under fusion

If one in four advice sub-queries contains “best”, then for a typical cluster of three to five searches, there is a good chance that at least one list is a “best” list, and often two. Under RRF, every list a page appears in adds a term. Pages that naturally match the “best” phrasing gain a list; pages that do not, lose one. MECHANISM

08 · Where a .Best name meets RRF

A domain name cannot change a ranking formula. It can change which lists a site appears in, how consistently it is recognised across them, and whether people click when it is cited. Under fusion, those are the levers that count.

Five mechanisms

1. The word in every “best” list. A page on hotels.best titled “Best boutique hotels in Paris” states “best” in its address, its title and its brand. Keyword retrieval for a “best” fan-out finds an exact lexical match; semantic retrieval finds a page whose identity is the query's intent. Each “best” sub-query is an additional list the site can appear in. MECHANISM
2. One entity across all lists. When different pages of a site rank for different sub-queries, a name that states the category makes them read as one source: hotels.best for “best hotels”, hotels.best for “hotel reviews Paris”, hotels.best for the brand search. Hostname-level overlap in 2026 data (about half of cited hostnames recovered, against about a quarter of URLs) suggests sites, not just pages, are what persists across searches. MECHANISM
3. The brand search is a category search. When ChatGPT adds “official” or runs a site: search to verify an option, the brand name it searches for is, on a .Best site, also the category and the word “best”. The verification list and the recommendation list point to the same place. MECHANISM
4. Readable addresses get cited. Ahrefs' study of 1.4 million prompts found natural-language URL slugs cited at an 89.78% rate against 81.11% for other slugs. A .Best domain is the most natural-language part of a URL: crm.best reads as words. MEASURED · GENERAL
5. The click after the citation. Being cited is not the end: people still choose which source to open. In the only controlled test of .Best, identical Google ads earned 20% more clicks and paid 8.1% less per click on .Best than on .com, on a small sample. Clicks feed back into how classic search engines judge a page, and classic rankings feed the fan-out lists. MEASURED · .BEST

A worked comparison

Take a question that fans out into four searches, and two sites with equally good content. The only difference is that one site's name and titles carry the word “best”, which earns it two additional lists. The ranks below are illustrative.

The generic site scores 1/68 + 1/75 = 0.0280. The .Best site scores 1/72 + 1/69 + 1/75 + 1/82 = 0.0539: about 1.9 times higher, without ranking first anywhere, because it is present in two more lists. MECHANISM

What the name does not do

The .Best advantage, in RRF terms. RRF adds one term for every list a source appears in. A .Best name is built to be found by the most common phrasing ChatGPT adds, to be recognised as the same source in every list, and to be chosen when it is shown. It does not replace good pages. It makes good pages count more times.

09 · The RRF content playbook

If fusion rewards presence across a cluster of related searches, content strategy changes shape: from one page per keyword to one well-covered cluster per buyer question.

1 · Map the fan-out, not the keyword

For each buyer question you want to win, list the searches ChatGPT is likely to run. Use the patterns the 2026 data shows: the category plus “best” or “top”; a “comparison” or “vs” form; a “reviews” form; a form with the current year; and the brand or “official” verification search for each option. Five to eight searches per question is a practical size.

2 · Give every search a page that deserves to rank

3 · Win the middle of many lists

Do not pour every resource into one first place. Under fusion, moving a page from #25 to #12 in three lists does more than moving one page from #4 to #1. Spread effort across the cluster.

4 · Make the answer quotable

Fusion gets a page into the pool; the reranker picks the passage. Put a one-paragraph answer at the top of each page, with the entity, the claim and the figure together, and keep comparisons in real tables.

5 · Keep it fresh and dated

The current year appears in a large share of 2026 fan-outs. Pages that show a visible update date and current information match those searches; pages that do not are filtered out of them.

6 · Be one entity everywhere

Use the same name, the same description and the same facts on your site, your profiles and the pages that mention you. On a .Best site, the name already states the category, so every mention reinforces the cluster.

7 · Let the crawlers in

OpenAI states that sites that opt out of its search crawler, OAI-SearchBot, will not be shown in ChatGPT search answers. Check robots.txt, render content in HTML, and avoid hiding answers behind scripts.

# robots.txt: allow AI search crawlers on a .Best site
User-agent: OAI-SearchBot
Allow: /

User-agent: Googlebot
Allow: /

Sitemap: https://hotels.best/sitemap.xml

Common mistakes under fusion

• Chasing one trophy keyword. A first place in one list is worth 1/61 and no more.
• Splitting the entity. Different names on different pages and profiles make one source look like several weak ones.
• Writing for the prompt instead of the fan-out. The searches add words the user never typed; pages that only match the prompt miss most lists.
• Undated content. Fan-outs carrying the year skip pages that look stale.
• Burying the answer. In the pool, the clearest passage wins the citation.

The .Best version of the playbook. Choose a name that states the category and the word the searches already use: hotels.best, crm.best, ebike.best. Build the cluster of pages behind it. Every page then carries “best” in its address, every mention of the brand repeats the category, and every list the site appears in shows the same, recognisable source.

10 · Measure your RRF footprint

Nobody outside OpenAI can see ChatGPT's fused scores. You can, however, compute a proxy from data you control, and track whether your cluster strategy is working. The method below needs a rank tracker and a spreadsheet.

Step by step

1. Pick ten buyer questions that matter to your business, phrased the way people ask ChatGPT.
2. Write the likely fan-out for each: five to eight searches, using the patterns in section 09. Where you can, capture real fan-outs with a fan-out logging tool or a browser extension that shows ChatGPT's searches, and replace your guesses.
3. Track rankings for every search in a classic search engine, top 100.
4. Compute your fused score per question with k = 60, and repeat with k = 10 and k = 100 to check the conclusion does not depend on the constant.
5. Compare with competitors on the same searches.
6. Check real citations: ask each question in ChatGPT several times a week, log which sources are cited, and see whether fused scores and citations move together.

# Fused score per question, k = 60 (Python)
def rrf(ranks, k=60, weights=None):
    """ranks: list of positions (1 = first), None if absent."""
    weights = weights or [1.0] * len(ranks)
    return sum(w / (k + r) for r, w in zip(ranks, weights) if r)

rrf([None, 8, 15, None])   # generic site → 0.02804
rrf([12, 9, 15, 22])       # hotels.best  → 0.05391

Five numbers to track

How to read it. The fused score is a model of the candidate pool, not of the answer. If it rises and citations do not, the bottleneck is the second gate: make passages clearer. If citations rise with it, the cluster strategy is working. Judge only trends over several weeks; single runs are noise.

11 · Use cases: three .Best sites, worked through

The cases below apply the method to three category names. Fan-outs follow the patterns reported in 2026 data; ranks and scores are illustrative targets, not measurements.

hotels.best · “Where should I stay in Paris for a romantic weekend?”

software.best · “Which CRM should a ten-person startup use?”

restaurants.best · “Where should I eat in Paris tonight?”

None of the three sites needs a first place. Each clears the reported 0.020 threshold by appearing in several lists in the middle of the first two pages, with a name that matches the “best” phrasing of most of them. The customer journey that follows is the one GEO is built on: question, fan-out, retrieval, fusion, citation, visit, booking. MECHANISM

12 · Limits and open questions

A white paper about an unconfirmed mechanism should say clearly what would prove it wrong. These are the limits we see, and the study that would settle the .Best question.

What could be wrong

1. ChatGPT may not use RRF. The evidence is one client-side observation and one hedged replication. OpenAI may use another fusion method, or none. The breadth-over-peak logic holds for most rank-fusion methods, but not for a system that, say, takes only each sub-query's top results. OPEN QUESTION
2. The constant is unknown. A small k would restore much of the value of first places. Section 04 shows breadth still wins at k = 10, by a smaller margin.
3. Rerankers may dominate. If a strong reranker re-scores a wide pool, fusion matters mainly for entry into the pool. Industry research in 2026 found fusion gains largely neutralised after reranking in some settings.
4. The pipeline changes. 2026 data shows ChatGPT building its own index, changing fan-out counts and shifting vocabulary within months. Any formula decoded today can change tomorrow.
5. Fan-out data is vendor data. Most 2026 measurements come from companies that sell AI visibility tools. Their samples are large but not independent, and their bases differ.
6. No study isolates the domain name. We found no 2026 study that tests whether a descriptive or keyword domain changes AI citation rates. The .Best fusion advantage is a mechanism until measured.

The .Best fusion study

The registry intends to publish such a study and invites .Best registrants to contribute anonymised data. Until then, the honest position is the one this paper takes: the formula is public, its use in ChatGPT is reported, and the .Best advantage under it is a reasoned hypothesis with one measured click effect behind it.

13 · A 30-day plan and conclusion

Conclusion: count more times

AI search does not hand out one first place. It runs several searches, merges what they return, and cites the few sources that hold up across them. If the merging works as reported, by Reciprocal Rank Fusion, the arithmetic is stark: a first place is worth barely more than a tenth, and every additional list a source appears in adds a whole new share of the score.

That changes what a name is for. A .Best domain cannot make a page relevant, and nothing suggests ChatGPT weighs the extension. What it can do is state, in the address itself, the word ChatGPT adds to its searches more than any other, and present the same recognisable source in every list where the site's pages appear. Under a formula that rewards being found again and again, that is a structural advantage worth testing, and worth building on.

When AI searches for the best, be in every list it builds. Map the fan-out, cover it with pages that deserve to rank, keep one name that says what you are the best answer to, and measure it.

About .Best

Sources

Read directly: the Elastic and OpenSearch documentation. Read through search coverage and secondary reports, because the original pages were not reachable from our network when this paper was prepared: all other sources. Figures are reported as published; verify each against its original source before quoting it.

• Cormack, Clarke & Büttcher. Reciprocal Rank Fusion outperforms Condorcet and Individual Rank Learning Methods. SIGIR 2009, pp. 758–759.
• Elastic. Reciprocal rank fusion (Elasticsearch reference).
• OpenSearch. Score ranker processor (introduced in 2.19), OpenSearch documentation.
• Microsoft. Relevance scoring in hybrid search using Reciprocal Rank Fusion (Azure AI Search).
• OpenAI. Introducing ChatGPT search, 31 October 2024.
• OpenAI Help Center. Searching the web with ChatGPT.
• OpenAI Help Center. Publishers and developers FAQ.
• Google Search Central. AI features and your website.
• Yesilyurt, M. Is RRF the Secret to Dominating AI Citations? I Decoded ChatGPT's Ranking Formula, 22 July 2025.
• Yesilyurt, M. How ChatGPT ranks citations (X thread), 22 July 2025.
• Yesilyurt, M. The RRF Top-n Playbook: How to Get Cited by ChatGPT's Web Mode (k≈60), September 2025.
• Peec AI (Rudzki, T.). Patterns we see in ChatGPT query fanouts, May 2026.
• Peec AI. Words ChatGPT adds to fanouts.
• Peec AI. ChatGPT fan-outs have doubled in length in 4 months.
• Peec AI. Rerankers for GEO/AEO: how AI search chooses passages and sources.
• Peec AI. ChatGPT built its own search index, September 2026.
• Ray, L. ChatGPT is changing which page types it cites in AI responses (new Peec AI data). X article, 2026.
• Ahrefs (Linehan, L.). How ChatGPT does query fan-out (15 million queries studied), October 2026.
• Ahrefs (Linehan, L.). Only 12% of AI cited URLs rank in Google's top 10 for the original prompt.
• Ahrefs (Linehan, L. & Guan, X.). Why ChatGPT cites one page over another (study of 1.4M prompts).
• AirOps. The influence of retrieval, fan-out, and Google SERPs on ChatGPT citations, March 2026.
• Understanding ChatGPT Citations: Evidence from Production Web Search Across Model Generations. arXiv 2610.15382, October 2026 (preprint).
• Scoring With the Engine: Retrieval Exposure, Cross-Engine Divergence, and the Limits of Engine-Agnostic GEO Scores. arXiv 2609.22655, 2026 (preprint).
• Medrano, Verma & Chhabra. Scaling Retrieval Augmented Generation with RAG Fusion: Lessons from an Industry Deployment. arXiv 2603.02153, March 2026.
• Meng & Ilvovsky. Sifei at SemEval-2026 Task 8: Hybrid Retrieval and Query Rewriting for Multi-Turn RAG. arXiv 2606.28352, June 2026.
• Mohanadasan, S. How ChatGPT actually picks sources (I read the network traffic, not the outputs), 2026.
• Discovered Labs. How ChatGPT uses Reciprocal Rank Fusion for AI citations.
• GlobeRunner. .BEST vs .COM Case Study (controlled Google Ads test, four Dallas sectors, small sample).

All example sites, ranks and scores marked “illustrative” are fictional. The playbook and measurement method in sections 09 to 13 describe the registry's recommended practice, not measured results.