How AI Search Works: Retrieval, Citations and Platform Differences

By Reputation.blog Editorial

Published: October 31, 2026
Last reviewed: October 31, 2026

Direct answer

AI search is not one system, one index, or one ranking algorithm. It is a family of products that combine a language model with a method of finding information outside the model. In a web-grounded answer, the system interprets the question, decides whether external information is needed, retrieves candidate sources, selects material, generates an answer, and—in many interfaces—shows links or citations readers can inspect.

Google’s generative AI features in Search are rooted in Google’s core Search ranking and quality systems. ChatGPT search can rewrite a question into targeted queries and sometimes sends them to search partners. Perplexity presents itself as a real-time answer engine with numbered citations. The Gemini workflow documented here is the Gemini API grounding tool, not a proven map of the consumer Gemini app. The Microsoft details come from Microsoft 365 Copilot and Microsoft 365 Copilot Chat, not every Copilot surface. Claude’s clearest public workflow is documented for its API and crawler system.

A page can be crawlable but not indexed, indexed but not retrieved, retrieved but not used, or used without a visible citation. No platform publishes a formula that converts publication into a citation. Eligibility is not selection, and selection is not a guarantee.

This guide explains the shared pipeline, the documented platform differences, and where public knowledge stops. For the technical companion, see How AI Systems Retrieve Information. For attribution, see Understanding AI Citations. The wider library is in the Knowledge Center.

What AI search is—and what it is not

Traditional web search primarily returns ranked documents and result modules; the user chooses what to open. AI search adds a generation step: a model receives information from selected sources and composes a direct response. Links therefore do more than act as destinations—they can become evidence for statements inside the answer.

A chatbot without web access is different. It may answer from training, conversation context, uploaded files, or connected private data. Fluent language is not evidence that a page was found, read, or cited.

AI search is also not a single shared “AI index.” Google documents grounding through its own Search system. OpenAI, Perplexity, and Anthropic each document separate crawlers for search, training, and user-directed actions. Microsoft 365 Copilot sends a generated query to Bing when web grounding is needed. These architectures are related, not interchangeable.

Nor is AI search simply “SEO with citations added.” Conventional search quality still matters where the product is built on a search index—Google says so explicitly—but answer retrieval adds filters. A system may issue several narrower searches, combine evidence from different pages, and visibly cite only the sources attached to the final statements.

The shared pipeline

Platforms do not publish the same internal sequence, and stages may occur in parallel, repeat, or be handled by separate services. This pipeline is an explanatory model grounded in public documentation—not a claim that every vendor uses identical software.

1. Interpret the request

The system interprets the prompt, using conversation context where the product supports it. A short request can contain several tasks: compare products, use current prices, prefer a location, and explain trade-offs. Retrieval depends on separating those requirements.

This is not human understanding. The model identifies likely intent, entities, constraints, dates, and missing information well enough to choose the next step. Ambiguity here can send retrieval toward the wrong entity or a broader question than the user intended.

2. Decide whether external information is needed

Not every prompt triggers web retrieval. Stable explanations may be answered from model knowledge. Current events, changed prices, new documentation, local conditions, and specialized facts are more likely to benefit from search.

The trigger may be automatic or manual. OpenAI says ChatGPT may search automatically when current information would help, and users can invoke Search. Anthropic’s API documentation describes Claude deciding whether web search would improve a response when the tool is enabled. Google’s Gemini API documentation says the model determines whether Google Search can improve the answer. The thresholds are proprietary.

3. Rewrite, decompose, or fan out the query

A conversational question is rarely a good database query. Systems may shorten it, extract entities and constraints, convert it into targeted terms, or split it into subquestions.

Google calls one version query fan-out: its model issues concurrent, related queries to gather broader evidence. Google’s example—“how to fix a lawn that’s full of weeds”—fans out to herbicides, chemical-free weed removal, and prevention.

OpenAI gives a similar ChatGPT example. A question about drugs targeting CCR8 might become “CCR8 immunotherapy drug development 2025,” followed by a more specific conference query after the first results are reviewed. That documents rewriting and possible iteration, not the number of queries used ordinarily or how result sets are weighted.

4. Retrieve candidate sources

Retrieval narrows a large collection to candidates. Depending on the platform, candidates may come from a prebuilt web index, a partner search service, licensed or partner content, a live page fetch, or a combination.

Four terms should not be conflated:

A crawler discovers and collects web content.
An index organizes collected content for fast search.
Retrieval selects candidates for the current question.
A user-directed fetch visits a page because the immediate request calls for it.

Allowing a crawler makes content eligible only for that crawler’s stated purpose; it does not promise retrieval or citation. Blocking one agent does not necessarily block every route.

Classical retrieval can match words and phrases. Semantic retrieval can compare meanings represented as numerical vectors, or embeddings. The foundational RAG paper by Lewis and colleagues combined a pre-trained language model with a dense vector index of Wikipedia accessed by a neural retriever. Commercial systems may combine lexical, semantic, entity, freshness, and quality signals in unpublished ways. AI search should not be reduced to “a vector database attached to a chatbot.”

5. Select, rank, filter, and read

Candidate retrieval is not the final source set. A system may rank results, remove duplicates, apply safety or policy filters, fetch full pages, work from indexed extracts, or search again when the first results leave a subquestion unanswered.

This is the least transparent stage. OpenAI says ChatGPT ranks results using multiple factors intended to surface relevant, reliable information, without publishing the factors or weights. Perplexity does not publish a ranking formula in its crawler documentation. Google says its generative Search features rely on core Search ranking and quality systems, but not how each retrieved page contributes to an answer.

“Crawled,” “indexed,” “retrieved,” and “read” are not synonyms.

6. Ground the generation in retrieved material

Retrieval-augmented generation, or RAG, supplies external material so a response can be based on more than the model’s stored parameters. Google also calls this grounding in Search: its systems retrieve relevant, up-to-date pages, review information from them, and generate a response linked to supporting pages.

Grounding does not guarantee correctness. Sources can conflict, be outdated, omit context, or be misread. A sentence can combine claims from several sources. Readers should open citations and check whether the page supports the statement. OpenAI explicitly warns that ChatGPT search results and citations can be incomplete, outdated, or incorrect.

Our separate guide to AI retrieval examines RAG, embeddings, query expansion, and candidate selection in greater depth.

7. Attach citations and handle follow-ups

Citation display varies. Google shows prominent clickable links to supporting pages. ChatGPT may include citations and a Sources view. Perplexity uses numbered citations. Gemini API grounding returns inline citation annotations and search metadata. Microsoft 365 Copilot Chat can expose generated web queries in a linked citation section, within the limits described below. Claude’s API web-search responses include citations.

A citation shows attribution in that response; it does not prove the cited page was the only influence. Not every retrieved page becomes a visible citation. The distinction between a brand mention and a source citation is central to the AI citations pillar.

Follow-ups can restart the process with new context. Because pages, indexes, models, and ranking systems change, the same prompt can produce a different source set later.

The eligibility chain

For a business page to become a cited source, it generally must pass several distinct gates:

Discoverable: a crawler, sitemap, link, partner feed, or user request can find the URL.
Accessible: the server and robots directives allow the relevant agent to retrieve it.
Available: the platform can store or access a usable representation.
Eligible: product-specific requirements, such as Google snippet eligibility, are met.
Retrieved: the page enters the candidate set for a particular prompt.
Selected: ranking, filtering, and reading leave it among the sources used.
Cited: the interface attaches a visible link or citation.

Failure can occur at any gate, and the remedy differs. A robots.txt problem is not fixed by rewriting an introduction; a clear page that is never retrieved has a different problem from a retrieved page that is not cited. This chain is a diagnostic model, not a published scoring formula.

How the major platforms differ

The sections below follow the same order where evidence allows: documented behavior, source discovery, attribution, eligibility, and unknowns. Documentation for one surface should not be transferred silently to another.

Google AI Overviews and AI Mode in Google Search

Google’s official AI optimization guide says generative AI features in Google Search are rooted in core Search ranking and quality systems. AI Overviews and AI Mode are therefore features of Google Search, not a separate public index disconnected from ordinary Google indexing.

Google describes two techniques. RAG, or grounding, uses core Search systems to retrieve relevant, up-to-date pages from the Search index; the systems review information from those pages and generate a response with prominent clickable links. Query fan-out issues concurrent related queries so an answer can draw on broader information than the original wording would return.

To appear as a supporting link, a page must be indexed and eligible to be shown in Google Search with a snippet, meet Search’s technical requirements, and belong to a site included in Search generative AI features in Search Console. Those are eligibility conditions, not a promise of retrieval or citation.

Google is unusually direct about tactics it says are unnecessary for Google Search:

No llms.txt file or other special AI file, markup, or Markdown version is required.
Content does not need to be artificially divided into tiny “chunks.”
Content does not need to be rewritten solely for AI.
Inauthentic mentions are not a shortcut.
No special schema is required for generative AI features, although structured data can support rich-result eligibility and broader SEO.

Google’s long-term advice is to create valuable, non-commodity, people-first content, which it says is likely to influence generative-search presence more than its other suggestions. Site owners can use the Search Console Generative AI performance report.

Google does not publish the exact trigger for every AI Overview, the weighting of retrieved pages, or a formula for turning an indexed page into a supporting link. Those details are unknown.

ChatGPT search

OpenAI documents ChatGPT as a conversational product that can search the web when current information would improve the answer. Search can begin automatically or be invoked manually. Responses using web search may include citations, and the Sources view can show cited sources and other relevant links.

OpenAI’s Help Center says ChatGPT search sometimes partners with other search providers. When it does, it typically rewrites the question into one or more targeted queries and sends them to the providers; after reviewing initial results, it may issue more specific queries. OpenAI links readers to Microsoft’s and Shopify’s privacy policies for how search providers may process queries.

OpenAI does not publish a complete provider list, each provider’s share, or the weight assigned to partner results. Claims that ChatGPT is “just Bing” or “just Google” go beyond the evidence.

OpenAI’s crawler documentation separates three roles:

OAI-SearchBot surfaces websites in ChatGPT search. Sites that opt out are not shown in search answers, though they may still appear as navigational links.
GPTBot collects content that may be used for training. Disallowing it does not itself remove a site from search.
ChatGPT-User supports user actions and may visit a page because a person asked a question. It is not an automatic crawler and does not determine Search eligibility.

Each setting is independent: a site can allow OAI-SearchBot while disallowing GPTBot. OpenAI recommends allowing OAI-SearchBot and permitting its published IP ranges for search eligibility. Robots changes can take about 24 hours to affect search systems.

OpenAI says results are ranked using multiple factors intended to surface relevant, reliable information, and placement is not guaranteed. The factors and weights are not public. Crawler access is a threshold, not a ranking formula.

Perplexity

Perplexity describes itself as an AI-powered search engine that searches the web when a question is asked and returns a conversational answer backed by verifiable sources. Its Help Center says content is sourced in real time and answers use numbered citations linking to original sources: interpret the question, search, summarize, and cite.

Its crawler documentation distinguishes:

PerplexityBot, designed to surface and link websites in Perplexity search results. It is not used to crawl content for foundation models.
Perplexity-User, which supports user actions and may visit a page to help answer a question and include a link. It is not used for ordinary crawling or training collection and, because the fetch is user-requested, generally ignores robots.txt.

Perplexity says its robots gather and index information automatically or in response to user requests, and crawler-preference changes can take up to 24 hours to appear.

The documentation does not disclose ranking or source-selection factors. Perplexity can accurately be described as search-led and citation-forward; claims about domain-authority thresholds, citation-count formulas, or fixed publication preferences are not documented.

Gemini and Grounding with Google Search

This section has a strict scope. Google publishes detailed documentation for Grounding with Google Search in the Gemini API. It proves how that developer tool works when enabled; it does not prove that the consumer Gemini app triggers searches, ranks results, or displays citations identically.

In the API, the google_search tool connects a Gemini model to real-time web content so it can answer beyond its knowledge cutoff and cite verifiable sources. The documented sequence is:

The application sends the prompt.
The model decides whether Google Search can improve the answer.
If needed, it generates and executes one or more search queries.
It processes and synthesizes the results.
The API returns a grounded response with inline citation annotations and search-call/search-result information.

Search is therefore a model-invoked tool, not necessarily a step performed for every API request. Consumer Gemini triggering, ranking, and citation internals remain undisclosed in the sources reviewed.

Microsoft Copilot

The Microsoft details here come from Microsoft Learn documentation for web search in Microsoft 365 Copilot and Microsoft 365 Copilot Chat. They should not be generalized automatically to every consumer Copilot surface or application integration.

In that context, web search is optional. When public web information would improve or ground a response, Copilot can fetch information from the Bing search service. It parses the prompt, identifies terms for which web information would help, and generates a search query from those terms. Microsoft says the generated query usually consists of a few words informed by the prompt and differs from the full prompt.

Microsoft documents privacy limits: the generated query does not include the entire prompt unless it is very short, entire Microsoft 365 or uploaded files, entire web pages or PDFs summarized in Edge for Copilot Chat, or identifying information based on the user’s Microsoft Entra ID.

After Bing returns information, Copilot uses it to compose the response. In Microsoft 365 Copilot Chat, web-search query citations can show the exact generated queries in the linked citation section. Microsoft says query citations are available in Copilot Chat, not in the Copilot pane inside Microsoft 365 apps such as Word or PowerPoint, and remain available in the thread for 24 hours.

The page does not publish a source-selection or ranking formula. It establishes the Bing connection and query-generation process for this Microsoft 365 surface only.

Claude

Anthropic’s public material supports a bounded account of Claude web search, especially on the API. When web search is enabled, Claude determines whether a request would benefit from current or specialized information. If so, it generates a targeted query, retrieves and analyzes results, and can run multiple progressive searches, using earlier results to inform later queries. Web-sourced API responses include citations for verification.

Anthropic separates its web agents:

ClaudeBot collects content that could contribute to training. Restricting it signals that future material should be excluded from training datasets.
Claude-SearchBot navigates the web to improve search-result quality. Disabling it prevents indexing for search optimization and may reduce visibility or accuracy in user search results.
Claude-User supports Claude users and may access a site when an individual asks a question. Disabling it prevents retrieval for that user query and may reduce visibility for user-directed search.

Training access, search indexing, and live user-directed retrieval are separate decisions. Anthropic does not publish a complete ranking formula for Claude search or confirm the full consumer-search backend in the sources reviewed.

A compact comparison

Platform or surface	Documented retrieval route	Attribution	Public boundary
Google AI Overviews / AI Mode	Google Search index and core ranking/quality systems; RAG and query fan-out	Prominent clickable supporting links	Exact triggers and page weighting unpublished
ChatGPT search	Automatic or manual search; rewritten queries may go to partners	Inline citations and Sources view	Provider mix, shares, and ranking weights unpublished
Perplexity	Real-time web search; PerplexityBot and user-directed fetches	Numbered citations	Ranking and selection factors unpublished
Gemini API grounding	Model may invoke Google Search and issue multiple queries	Inline annotations plus search metadata	Does not prove identical consumer-app behavior
Microsoft 365 Copilot / Copilot Chat	Generated query sent to Bing when web grounding is needed	Citations; generated-query view only in the documented Chat surface	Should not be generalized to every Copilot product
Claude API / search agents	Model decides whether to search; targeted and progressive searches	Citations in web-sourced responses	Ranking formula and full consumer backend unpublished

The table maps documentation, not quality. A missing detail means the vendor has not published it in the sources reviewed, not that the mechanism does not exist.

Practical examples

Example 1: One question becomes several searches

A homeowner asks how to repair a lawn full of weeds. Google documents a broader fan-out: herbicides, chemical-free removal, and prevention can all become related searches. The final answer may cite a page that never targeted the homeowner’s exact wording but answers one necessary subquestion well.

The lesson is not to create thin pages for every wording. A genuinely useful page should expose relevant sections clearly enough for retrieval and reading systems to match the page—or a section—to the subquestion it answers.

Example 2: A question whose answer changed recently

Suppose a user asks what changed in a software product’s pricing this month. Training knowledge may describe an older plan. A web-grounded system needs current retrieval, and source dates become part of the evidence.

Platforms may proceed differently. ChatGPT may rewrite the product and date constraints and search again if results are incomplete. Perplexity describes real-time sourcing. Google’s Search features can retrieve up-to-date pages from its index. Claude’s API can run progressive searches. Similar-looking answers can therefore rest on different evidence because query transformations, indexes, partners, and ranking systems differ.

Publishers should keep the authoritative pricing page current, show the effective date, and avoid leaving the only current explanation in a social post.

Example 3: A business is mentioned but not cited

A user asks about a company. The answer names the business but links only to a directory, news article, or marketplace profile. That is a brand mention, not a citation to the company’s website.

Possible explanations include an access or eligibility problem, a third-party page answering the subquestion more directly, the official page being retrieved but not selected for the visible link, or the model taking the name from retrieved material while citing another source for the claim. The observation alone does not identify the cause.

Diagnosis requires checking access, platform eligibility, source clarity, consistency with third-party records, and the exact prompt and platform tested. The AI visibility pillar covers that work without treating one answer as a permanent ranking.

What AI search is not

Training-data knowledge is not live retrieval

Training gives a model patterns and facts learned during development. Retrieval supplies material for the present request. The two can be combined, and an answer may not reveal the boundary. A citation is evidence of attribution, but its absence does not prove every sentence came only from training.

Indexed is not cited

Indexing means a system has a usable record of a page. Citation means that, for a particular prompt at a particular time, the page survived retrieval and selection and was visibly attached to the answer. Most indexed pages will not be cited for most questions; that is not a penalty by itself.

A mention is not a citation

A mention means an entity appeared in generated text. A citation gives the reader a source link. They measure different things and should not be merged. A company can gain mentions while its domain receives no citations, or be cited for an industry claim without being recommended.

A citation is not a recommendation

A source may be cited because it supports a definition, reports a disputed claim, supplies a date, or is the page being criticized. Context matters; counting links without reading the surrounding sentence produces a misleading metric.

One answer is not a stable ranking

AI answers can change when a prompt is rephrased, an index refreshes, a page changes, a model changes, or the system issues different searches. A screenshot is an observation, not a durable position. Measurement requires repeated prompts, dates, platforms, and saved responses.

Business implications

The useful response to AI search is disciplined source management, not a layer of tricks on a weak information footprint.

Keep access gates intentional. Review robots.txt and CDN rules for the search agents relevant to each platform. OpenAI, Perplexity, and Anthropic document separate search and training controls; a blanket block copied from an AI-policy template can remove search eligibility while trying to prevent training.

Maintain ordinary index quality. Google makes basic Search eligibility foundational to its generative features. Crawlability, indexation, snippet eligibility, technical quality, and useful content remain practical concerns, even though every platform has its own retrieval system.

Publish pages that can serve as evidence. A strong source page answers a defined question directly, identifies the responsible organization, shows dates where currency matters, distinguishes fact from opinion, and links important claims to primary evidence. Thin pages built only to repeat a query phrase give a retrieval system little reason to prefer them.

Make important facts consistent. Company and product names, locations, specifications, prices, leadership details, and dates should agree across the official site and material third-party profiles. Inconsistency creates reconciliation work and makes another source easier to cite.

Use structured data for clarity, not as a citation spell. Google says no special schema is required for generative Search features. Accurate structured data can support conventional features and machine understanding, but it must match the visible page and must not invent ratings, offers, credentials, or relationships.

Measure by platform and prompt. Record the prompt, platform, date, answer, mentioned entities, cited URLs, and citation context. Separate branded from unbranded prompts and mentions from citations; one screenshot is not a durable improvement.

None of these actions guarantees a citation, ranking, recommendation, or visibility increase. They improve the conditions under which a relevant page can be discovered, understood, retrieved, and trusted as a source—a distinction controlled by third-party systems, not by the publisher.

Common misconceptions

“There is one AI search algorithm.”

No. The platforms share concepts—retrieval, ranking, generation, and attribution—but use different products, indexes, partners, triggers, and displays.

“If a page ranks in Google, it will be cited by every AI assistant.”

Google ranking is directly relevant to Google’s Search features because they are rooted in Google Search. It does not determine ChatGPT, Perplexity, Claude, Gemini app, or Copilot source selection.

“Blocking the training crawler removes a site from AI search.”

Not necessarily. OpenAI, Anthropic, and Perplexity document separate agents or functions for training, search, and user-directed access. The exact user agent matters.

“Every AI answer is searched live.”

Some answers use no web retrieval. Some products decide automatically whether to search; some API tools must be enabled; some user-directed fetches happen only for a specific request. “Live” also does not guarantee that every source was fetched at the moment the answer was generated rather than retrieved from a current index.

“A special AI file is now basic SEO.”

Google explicitly says llms.txt is not required for its generative Search features. A file may have other uses in a particular workflow, but it should not be sold as a universal eligibility requirement.

“Chunking content guarantees retrieval.”

Google says artificial chunking is not required for Google Search’s generative features. Clear headings and coherent sections can help human and machine readers, but fragmenting an article into unnatural pieces is not a documented universal rule.

“The cited page must be the page the model ‘believes’ most.”

Citation is an interface and attribution decision following retrieval and generation. Public documentation does not expose a simple belief score, and a citation may support only part of a sentence.

“AI citations are permanent once earned.”

Indexes, pages, models, and generated queries change. A citation is attached to a response, not awarded as a permanent badge to a domain.

Frequently asked questions

Is AI search replacing traditional search?

AI search changes the interface—from a list of links toward a generated answer with sources—but major AI features still depend on retrieval systems and web indexes. Google’s generative Search features are explicitly rooted in core Search, so the two should be studied together.

Does every AI answer use the web?

No. A model may answer from training knowledge, conversation context, uploaded material, or connected data. Web search may be automatic, manual, developer-enabled, or unavailable. A citation shows web material was attributed; it is not a complete log of every influence.

What is query fan-out?

Query fan-out expands one question into multiple related searches. Google uses the term for concurrent related queries in its generative Search features. Other systems document rewriting or progressive searching, but those implementations should not be treated as identical.

What does a website need to be eligible for AI citations?

The platform must be able to discover and access the page through an applicable index, crawler, partner, or fetch route. Google also requires indexation and snippet eligibility for its Search features. OpenAI identifies OAI-SearchBot access as the ChatGPT search-eligibility route. Eligibility does not ensure retrieval, selection, or citation.

Should a site block GPTBot or ClaudeBot?

It depends on whether the objection is to training, search, or both. GPTBot and ClaudeBot are associated with training-related collection, while OpenAI and Anthropic document separate search agents. Blocking a training agent is not the same as blocking search; decide by function.

Does structured data guarantee an AI citation?

No. Google says structured data is not required for its generative AI features and that no special AI schema exists for that purpose. Accurate structured data can help conventional features and clarity, but it guarantees no citation anywhere.

Why did an AI system cite a competitor or directory instead of the official site?

One answer rarely identifies the cause. Possibilities include access or eligibility problems, a clearer third-party answer to the subquestion, fresher information, stronger corroboration, or a different retrieval path. Diagnosis requires page checks and repeated prompt tests.

Can a business make an AI platform cite it?

A business can improve discoverability, clarity, consistency, and source quality, but no one can honestly promise control over a platform’s retrieval and citation decisions. Guaranteed AI citations or recommendations are outcomes controlled by third-party proprietary systems.

What remains unknown

The public record leaves consequential questions unanswered:

The exact triggers for every Google AI Overview or ChatGPT search invocation.
The complete source-ranking weights on the major consumer platforms.
The share of ChatGPT answers attributable to each search partner or source route.
Perplexity’s ranking and source-selection formula.
Claude’s search-ranking formula and full consumer-search backend.
Whether consumer Gemini follows the Gemini API grounding workflow for triggering, ranking, and display.
How consumer Copilot surfaces differ from documented Microsoft 365 Copilot and Copilot Chat behavior.
How much each retrieved page contributes to a generated sentence when sources are combined.

These unknowns are not an excuse for inaction. They are a reason to separate documented mechanics from observation, test results from causation, and eligibility work from guaranteed outcomes.

References

Google Search Central. https://developers.google.com/search/docs/fundamentals/ai-optimization-guide?hl=en
OpenAI Help Center. https://help.openai.com/en/articles/9237897-searching-the-web-with-chatgpt
OpenAI Developers. https://developers.openai.com/api/docs/bots
Perplexity Help Center. https://www.perplexity.ai/help-center/en/articles/10352895-how-does-perplexity-work
Perplexity Documentation. https://docs.perplexity.ai/docs/resources/perplexity-crawlers
Google AI for Developers. https://ai.google.dev/gemini-api/docs/google-search
Microsoft Learn. https://learn.microsoft.com/en-us/copilot/microsoft-365/manage-public-web-access
Anthropic. https://claude.com/resources/articles/web-search-api
Anthropic Help Center. https://support.claude.com/en/articles/8896518-does-anthropic-crawl-data-from-the-web-and-how-can-site-owners-block-the-crawler
Lewis et al., NeurIPS 2020. https://arxiv.org/abs/2005.11401v2

Reputation.blog Editorial