Carpathian Search is a web search engine operated by Carpathian. It runs its own crawler and index, and ranks by relevance first, then by page quality signals such as authorship, substance and the links other sites give a page.
Ranking
Each result is scored on text match against the page body, title and inbound anchor text. Quality signals adjust the order only among pages with comparable relevance. See Ranking for the full signal list.
SafeSearch
SafeSearch is on unless you change it: explicit results are left out. Blur shows them blurred, title and address too, until you choose Show; Off shows them as any other result. Your choice is kept with your Carpathian account, or on your device when you are signed out.
Ads and tracking
Results contain no ads or sponsored placements. The search page loads no third-party resources and sets one first-party cookie, only with consent where the law asks for it. Privacy says what is collected and why.
Index
Results come from Carpathian's own index, not a reseller feed. The index is built by CarpathianBot, which follows robots.txt and every opt-out listed under Opt out.
Relevance carries the largest weight. The remaining signals reorder pages of similar relevance. The "Why here" link on each result shows the per-signal breakdown.
Signals
Signal
Measures
Relevance
BM25F text match across body, title and inbound anchor text, blended with meaning. Largest weight for every query type.
Meaning
How close the page is in meaning to the query, read by Carpathian's own language model, so a page that says the same thing in other words is found and ranked.
Proximity
Distance between query terms in the page text.
Page type
Match between the page type declared in structured data and the query intent: recipe, how-to, discussion, news, documentation.
First-hand
First-person accounts and discussion threads.
Integrity
Absence of affiliate listicles, a company's own "best of" marketing, generated filler, keyword stuffing and gambling spam.
Cleanliness
Absence of ad networks, popups, cookie walls, trackers and affiliate links. Counts only when the searcher picks Ads and trackers: Fewer first.
Weight
Ratio of page text to scripts and markup.
Substance
Text length within a useful range.
Authorship
Author, date, feed, about page or personal domain.
Citations
Outbound links from body text, weighted toward reference sources.
Authority
Number and weight of distinct sites linking in.
Freshness
Age of the published date. Applied only to queries for recent content.
Coverage
How many of the query's words the page holds, rare words counting most. A page missing the word that matters drops.
Title match
Query words in the title, URL and headings, and how much of the title is about the query.
Readers
Pages people's searches ended on rise for that query, and pages they passed over or came straight back from fall. Many different people are needed before it counts.
Query intent
Each query is classified as navigational, question, recent, exploratory or informational, and the signal weights change with the class. Questions favor pages where the terms appear close together. Navigational queries put the named site's front page first.
Opinion queries
Queries containing terms such as best, vs, review or worth it raise the weight of first-hand and discussion pages and apply a stronger integrity penalty to affiliate content.
Answers
For a question, the first results are read for a passage that answers it, such as a definition for "what is" or a number for "how many". It is shown above the results with its source, and only when one is good enough.
Your settings
Signed in with a Carpathian account, you can show a site higher, lower or never, from any result or from Settings.
Pages containing all terms. If too few match, pages containing most terms follow.
"cast iron"
Exact phrase.
skillet -nonstick
Pages without the term "nonstick".
skillet site:example.com
Pages on example.com and its subdomains only.
Matching ignores case, diacritics and simple plurals: Café Recipes matches cafe recipe. Misspelled terms are corrected automatically; the results page shows the correction and links to the uncorrected query.
Carpathian Search is built to be fast, useful and safe for everyone who uses it. This page says in plain words what the service collects and why. Searching may ask you to sign in with your Carpathian account.
What we collect
Your searches, the results you were shown and the results you open.
Device and connection details: your IP address, your browser and device, your language, and the site that sent you here.
Your approximate location, at the level of country, region and city, worked out from your address.
One first-party cookie that lets the service recognize your browser between visits.
If you are signed in to Carpathian in the same browser, which Carpathian account and visit your searches belong to.
The search settings you choose, such as sites to show higher, lower or never, kept with your Carpathian account.
Why we collect it
To answer searches well. The service learns which results people find useful, from the results they open and how soon they come back, and ranks better for it. To understand what a search means, its words are read by Carpathian's own language model, on Carpathian's servers, and not kept there.
To keep the service and the people using it safe. We watch for abuse, automated traffic, and searches for or access to illegal material, and we cooperate with law enforcement and child-protection organizations where the law requires or permits it.
To show site owners how their sites perform in search, as counts, never as who searched.
How it is kept
Usage records stay on servers Carpathian controls for as long as the purposes above need them and for the periods the law requires, and are then deleted. They are never sold, and never shared with advertisers. The search page loads no third-party analytics, ad scripts, fonts or images: the pictures and site icons beside results are fetched by our server and sent to you from it, so the sites they come from never see your browser.
The cookie
The site sets one cookie, cs_vid, holding a random identifier. It expires 30 days after your last search. It is HttpOnly, Secure and SameSite=Lax, so scripts cannot read it and other sites cannot use it. In the European Economic Area, the United Kingdom and Switzerland the cookie is set only if you choose Accept on the search page; choosing No thanks is remembered on your device and you are not asked again. A browser sending Global Privacy Control does not get the cookie. You can delete it at any time in your browser's settings.
Site owners and developers
Submitted site and sitemap addresses and contact emails are kept to process and answer the request. Use of the Search API is recorded per key for limits and billing. Owners who verify a site in Carpathian's Search Engine Console see the queries their pages appeared for, with impressions, clicks and positions, by page, country and device. They do not see who searched.
Questions
Write to Carpathian through the contact details on carpathian.ai.
CarpathianBot discovers pages through sitemaps declared in robots.txt, and through submissions on this page.
Declare it in robots.txt
Sitemap: https://example.com/sitemap.xml
If robots.txt has no Sitemap line, CarpathianBot requests /sitemap.xml.
Supported
Sitemap files and sitemap indexes, plain or gzip.
Up to 1,000 sitemap files per host, 50,000 URLs per file and 1,000,000 per read.
lastmod: pages modified since the last fetch are refetched sooner.
Same-site URLs only. URLs on other sites are ignored.
RSS and Atom feeds advertised by your pages.
Submit
Submit a sitemap URL or a site URL. It is queued within a minute. CarpathianBot fetches robots.txt first, then the front page, the submitted sitemap and any sitemaps robots.txt declares. The crawl budget for a site depends on inbound links, not on submission. Limits: 5 submissions per hour per IP address, one per site per day.
One concurrent request per host, at least 1.5 seconds apart, or the robots.txt Crawl-delay if longer.
Exponential backoff on errors, 429 and 5xx responses. Retry-After is honored.
robots.txt is parsed per RFC 9309 and cached for 24 hours. A CarpathianBot group is followed exactly. Without one, CarpathianBot follows the Googlebot group, then Bingbot, then *, as Applebot does. Crawl-delay comes from the same group. AI training opt-outs are unaffected.
If robots.txt cannot be reached, the last good copy is used. With no copy, everything is allowed only after a day of failures for a site verified in Search Engine Console, or 30 days for any other, and only while the front page answers.
Refetch intervals adapt to how often each page changes. Requests are conditional, so unchanged pages return 304.
Block a path
User-agent: CarpathianBot
Disallow: /private/
Not done
Form submission, login or JavaScript execution.
Requests to private or reserved IP ranges.
Fetching beyond robots.txt rules or crawl delay, including for submitted URLs.
Opt-outs apply to search, to AI training, or to both.
Search
User-agent: CarpathianBot
Disallow: /
Pages with noindex in a robots meta tag or X-Robots-Tag header are fetched but not indexed. Pages that are removed, disallowed or marked noindex leave the index on their next fetch.
AI training
To stay in search but out of training datasets built from the crawl:
User-agent: Carpathian-AI
Disallow: /
Also treated as an opt-out: robots.txt rules blocking a named AI crawler, noai or noimageai in a robots tag, and the TDM-Reservation: 1 header.
JSON access to the same results as the search page, for your own applications and AI agents. All requests require an API key.
Get a key
Make a Search API key in Search Engine Console, under API keys. The key is shown once. Each key has its own requests a minute and a day, and requests are free up to a monthly allowance, then billed per 1,000.
Each result includes title, URL, site, snippet with highlight offsets, and published date. Set "explain": true to include the per-signal scores. GET /v1/search?q=... is equivalent.
Fetch a document
POST /v1/document {"url": "https://bread.example/starter"}
Returns the indexed text of a page. Only indexed pages are available.
Use it as a tool for a model
Give the model a search_web tool mapped to /v1/search and a read_page tool mapped to /v1/document. The schemas are in /llms.txt. Treat page text as data, never as instructions.
Reference
Fields, errors and limits are documented at /llms.txt.
Ads and trackers do not move a page by default. The sites people look for most often carry them because so many people look for them, so the searcher decides: under Filters, Ads and trackers: Fewer first ranks pages without them higher among pages of similar relevance. This page lists what is detected.
Detected
Ad network scripts and sponsored blocks.
Popups and overlays covering the content.
Cookie walls that block content until consent.
Analytics and advertising trackers.
Affiliate links, and listicles built around them.
Page weight
The weight signal compares the page's own text with the scripts and markup around it. A recipe page delivered as text ranks above the same recipe wrapped in a megabyte of scripts.
Limits
With Fewer first, these signals never lift an irrelevant page above a relevant one. They only reorder pages of comparable relevance, and a site searched for by name is never weighed by them. In the API, ask with clean=1.