# Seeds, sites and places (/docs/prism/discover-seeds) Resolve names to accounts, let Discover find the sites that hold a topic, and search near a place A Discover run usually starts from keywords. These options start it from something else: names you already know, sites Discover finds for you, or a place. Each section opens on a request, and `dry_run=true` prices any of them for free. Every other request field is on [Discover request fields](/docs/prism/discover.md). ## Names to accounts: seeds `seeds` resolves names you already have (companies, creators, publishers) to their own accounts, instead of searching a topic. Seeds imply `mode=resolve` and answer account rows (`view=accounts`, the only view a resolve run answers). ```bash title="cURL" curl -G "https://www.socialcrawl.dev/v1/prism/discover" \ -H "x-api-key: YOUR_API_KEY" \ --data-urlencode "seeds=notion.so,figma.com,BMW Group:bmwgroup.com,Trevor Noah" \ --data-urlencode "platforms=linkedin,instagram,twitter" \ --data-urlencode "max_credits=80" ``` On the GET, `seeds` is a comma-separated list, and each entry is one of three forms: | Entry | Resolved as | |---|---| | `notion.so` or `https://www.figma.com/` | a company with that domain, named from it (`Notion`, `Figma`) | | `BMW Group:bmwgroup.com` | a company with that name and that domain (`Name:domain`; the domain may be a URL) | | `Trevor Noah` | a name alone, searched as a person (a creator) | Write a name and its domain only as `Name:domain`. `Name (domain)`, `Name|domain`, `Name domain`, `Name - domain` and `domain:Name` are free 400s whose `did_you_mean` is the `Name:domain` form to send. Use `Name:domain` when the brand name is not the domain's own word (`BMW Group:bmwgroup.com`). On the POST, `seeds` takes the same text entries, or objects `{ name, domain?, country?, platforms? }`, whose `name` is always read as a name: ```json title="POST /v1/prism/discover" { "seeds": [ { "name": "Siemens", "domain": "siemens.com", "country": "DE" }, { "name": "Trevor Noah", "country": "ZA", "platforms": ["instagram", "twitter"] } ], "platforms": ["linkedin", "twitter", "instagram"], "max_credits": 100 } ``` - Each seed is resolved by `prism/find-accounts` once per platform, on LinkedIn, Instagram, TikTok, X, Threads and YouTube (its channel search): the seed's own `platforms`, else the request's, else find-accounts' default (LinkedIn, Instagram, TikTok and X for a seed with a `domain`, which is searched as a company; Instagram, TikTok and X for a name alone, searched as a person). A platform outside that set is a free 400 naming it. - A seed with a `domain` is checked against each leading account's own website, and else against the links in its bio: `match.evidence.domain_match` and `domain_source` (`external_url` or `bio`). The account's own website counts most; a mention in the bio counts less, because a reseller or a fan names the site too; an email address at the domain does not count. A `domain` that is a profile, link or site-builder page (a LinkedIn or Google Sites URL, a link page) is a free 400 naming it: send the company's own site. Its leading accounts on Instagram, TikTok, Threads and LinkedIn are looked up on their profiles first, because those searches return no bio or website (1 credit a lookup, 5 for a LinkedIn company page, refunded when it fails). `country` is passed as the place to judge against. - Every account row carries `match { seed, level, confidence, match_confidence, basis, evidence }`, the same `level`, `confidence` (the weaker of the ranking and the check) and `match_confidence` a direct `prism/find-accounts` call answers, and `matches`, every seed that resolved to that account (two seeds naming one company make one row), strongest first. Only accounts at or above a `match_confidence` of 0.6 are listed: a platform where no account reaches it lists none, so you get no account rather than a guess. These are automated labels for review (`judgments.automated_decision: false`), never a statement of who runs an account. - When a seed's lookup on a platform was missing a part (the judge or the check was unavailable, or one of its searches failed), that platform may list fewer accounts for the seed or none, and `data.platforms..error_codes` and the leg on the legs route name why: `judge_unavailable`, `verification_unavailable` or `search_partially_failed`. Such an answer is never cached, so the same request sent again with `new_run=true` tries those lookups again. - Each lookup is a search of the run, billed at find-accounts' own price (for example 2 credits on Instagram, TikTok or X for a name, 16 on LinkedIn for a company with its domain). A lookup keeps that price when no account matches, because the searches and the judging ran; a part that fails or could not be judged is refunded. The estimate and `dry_run` price every one, and a call holds at most 36 of them, so 12 seeds on 3 platforms answer in the call and more continue in the background. A resume never looks a seed up twice. Nothing else is judged or billed unless you ask for `label`, `include=contacts` or a filter that needs a profile. - `seeds` with `keywords`, `query`, `brands`, `groups`, `hashtags` or another subject, with a `preset`, or with a view other than `accounts` is a free 400 naming both fields. ```json title="A resolved account row (a name alone on Instagram, trimmed)" { "account": { "key": "instagram:…", "platform": "instagram", "username": "example.creator" }, "match": { "seed": "Example Creator", "level": "same", "confidence": 0.94, "match_confidence": 0.9, "basis": "judged", "evidence": { "website_match": false, "cross_link": false, "bio_consistent": true, "fan_or_parody": false, "domain": null, "domain_match": null, "domain_source": null } }, "matches": [{ "seed": "Example Creator", "level": "same", "match_confidence": 0.9, "basis": "judged" }] } ``` ## Sites found for you: sites=auto `sites=auto` (or `search_in.web=discover_sites`) finds the forums, blogs and article sites that hold your topic, so you do not have to name them. For each keyword and language, discover runs up to 4 searches on the web: the keyword alone, then the keyword with the words people use for a review, a forum and a question in that language (for example 후기, 커뮤니티 and 질문 in Korean, or 口コミ, 掲示板 and 知恵袋 in Japanese). Their results are rows of the run like any web row. The judge reads them, and discover ranks their sites by kept rows per credit, with half a credit added to each site's credits so that a site found once among twenty results never outranks one found several times, then searches the top `sites_auto_max` sites (default 5, at most 20) the way listed `sites` are searched. Sites discover already searches natively (YouTube, Reddit, Naver and the other platforms, the review sites and the marketplaces such as eBay, AliExpress, Gumtree, Walmart and Amazon), news publishers and sites that contain one of your `exclude_terms` are never chosen. ```bash title="cURL" curl -G "https://www.socialcrawl.dev/v1/prism/discover" -H "x-api-key: $SOCIALCRAWL_API_KEY" --data-urlencode "keywords=excessive sweating treatment" --data-urlencode "platforms=web" --data-urlencode "sites=auto" --data-urlencode "country=US" --data-urlencode "period=365d" ``` ```json title="Two entries of data.sites (from keywords=bonsai care&platforms=web&sites=auto&country=US, trimmed)" [ { "domain": "bonsaiempire.com", "rows_kept": 24, "bodies_asked": 0, "dated_share": 0, "discovered": true, "discovery": { "hits": 6, "kept": 5, "kept_share": 0.83, "kept_per_credit": 5.44, "credits": 0.92, "searched": true, "excluded": null } }, { "domain": "reddit.com", "rows_kept": 6, "bodies_asked": 0, "dated_share": 0.5, "discovered": false, "discovery": { "hits": 6, "kept": 6, "kept_share": 1, "kept_per_credit": 10.75, "credits": 0.56, "searched": false, "excluded": "native_source" } } ] ``` - `data.sites` lists every site the search found, one entry per domain. Each has `discovered` and a `discovery` block with `hits`, `kept`, `kept_share`, `kept_per_credit` (the rate without the half credit), `credits` (the discovery credits shared out to the site), `searched`, and `excluded` (`native_source`, `news`, `excluded_term` or `named`, else null), beside its kept rows and what its bodies gave. - `search_in.web=walk` also walks each named or found site: the site's map is filtered by the keyword, up to `walk_urls` pages per site and keyword (default 50, at most 500), and each page that no other search found is read once. The rows are judged before their pages are read, so an off-topic page costs nothing. - The sites are chosen once, when the run starts, and stored with the run, so a resume never searches for them again. `dry_run` prices the searches of the sites to be found as conditional legs and the walk's pages in `estimate.by_stage.bodies`. - A walk of more than 50 pages, or more found sites than one call can search, runs in the background. - `expand_from=suggestions` (with `expand` from 1 to 3) plans the extra searches from what people type into the YouTube and TikTok search boxes for each keyword: both suggestion lists are read first, at 1 credit each per keyword, and given to the planner. Without it (`expand_from=keywords`, the default) `expand` plans from your keywords alone, as before. `expand_from` without `expand` is a free 400. ### Sites found by sites=auto Measured on 17/10/2026 with `sites=auto` on one hyperhidrosis query in the United States and one in Korea, with `search_in.web=walk` on the United States group. "Found and walked" means the site was chosen and its pages about the keyword were read; "found and searched" means it was chosen and searched without a walk. | Domain | Market | How it was used | Kept rows | Bodies read | What it is | | --- | --- | --- | --- | --- | --- | | `pmc.ncbi.nlm.nih.gov` | US | Found and walked | 27 | 9, full text | Research articles | | `pinnacleskin.com` | US | Found and walked | 10 | 9, full text | A clinic's blog | | `sweathelp.org` | US | Found and walked | 18 | 7, full text | A patient association | | `ubiehealth.com` | US | Found and walked | 21 | 1, full text | Symptom questions and answers | | `hyperhidrosisnetwork.com` | US | Found and walked | 13 | 5, full text | A patient community | | `baekrokdam.com` | KR | Found and searched | 13 | None | A clinic | | `bd.mipretty.co.kr` | KR | Found and searched | 7 | None | A beauty review board | | `beautystone1.com` | KR | Found and searched | 7 | None | A clinic | | `cafe.daum.net` | KR | Found, not chosen | 1 | None | A community | In Korea most of the sites found were clinic and review pages rather than communities, so when what people say in their own words matters, also name the communities you want in `sites` in a second group. `sites=auto` cannot be mixed with listed domains in one group. ## Listings and places near a place `near` is the place the listings, places and web searches look in (a web search sends a place name as its location): a place name (`near=Dubai`) or `lat,lng` (`near=25.2,55.27`). It does not filter other rows; a request where no search reads it says so in `_warnings`, and so does a place name that names several places, or none (which one was searched, or that nothing was). `search_in=listings` searches marketplaces by keyword: Facebook Marketplace around `near`, and the marketplaces named in `platforms` (`ebay`, `amazon`, `aliexpress`, `gumtree`, `tiktokshop`, `walmart`, each searched only for listings). Each row is `record_type: listing` with `price { current, currency }`; where the marketplace rates the product, its stars and its number of ratings are in `rating { value, max, count }`, and `engagement.comments` on a listing row is that same number of ratings. ```bash title="cURL" curl -G "https://www.socialcrawl.dev/v1/prism/discover" \ -H "x-api-key: YOUR_API_KEY" \ --data-urlencode "query=rolex" \ --data-urlencode "platforms=facebook,aliexpress,walmart" \ --data-urlencode "search_in=listings" \ --data-urlencode "near=Dubai" \ --data-urlencode "max_credits=20" ``` ```json title="A listing row (from query=rolex&platforms=amazon&search_in=listings&country=US, trimmed)" { "id": "as_wrxfc6057q4b", "platform": "amazon", "record_type": "listing", "url": "https://www.amazon.com/dp/B0HJ2ZVRNZ", "published_at": null, "text": "Rolex Pre-Loved Oyster Perpetual Explorer II Automatic Watch 40", "price": { "current": 11690, "currency": "USD" }, "author": null, "judgments": { "relevance": { "p": 0.71 } }, "matched_keywords": ["rolex"], "provenance": { "endpoint": "/v1/amazon/product-search", "request_id": "req-…", "captured_at": "2026-10-18T10:24:58.829Z" } } ``` - Facebook Marketplace needs `near`: without it the request is a free 400 naming `near`. A place name is looked up first (1 credit), then searched at its coordinates; its listings are dated, carry the place they are listed in as `geo.place`, and are walked newest first when you send a window. - Where a marketplace's rows name their seller (AliExpress, TikTok Shop and Walmart), the seller is the row's `author`, keyed by the marketplace's seller id, and rolls up as an account under `view=accounts`. Facebook Marketplace search rows name no seller, so under `view=accounts` each kept Marketplace listing is looked up once on `marketplace/item` (1 credit a listing, in `credits.by_stage.enrich` and priced by `dry_run`), and its seller becomes the row's `author`, keyed by the seller's profile URL, and rolls up as an account. eBay, Amazon and Gumtree search rows name no seller. A seller is linked to an Instagram account only when its listing links that profile, never by a matching name. - Listings are checked against your keywords alone, free, even when you send a `context`: a listing for sale never reads as an opinion or an experience, so the context check would drop real listings. - Each marketplace searches its own markets from `country` (eBay 19, Amazon 13, TikTok Shop 16, Walmart US and CA, Gumtree GB, AliExpress any). A country a marketplace does not serve is named in `_warnings` and not searched there. Etsy has no keyword search, so it is not a listings source. `search_in=places` searches places at `near`: the businesses a keyword names there (`record_type: place`, with the name, category and address in `text` and `geo.place`, its stars and its number of reviews in `rating { value, max, count }`; on a place row `engagement.comments` is that same number of reviews, not comments), and, when `instagram` is among the platforms, the posts at the Instagram places named for `near`. ```bash title="cURL" curl -G "https://www.socialcrawl.dev/v1/prism/discover" \ -H "x-api-key: YOUR_API_KEY" \ --data-urlencode "keywords=dentist" \ --data-urlencode "search_in=places" \ --data-urlencode "near=Lisbon" \ --data-urlencode "max_credits=10" ``` ```json title="A place row (from keywords=orthodontist&search_in=places&near=Porto, trimmed)" { "id": "gq_xprk95xfx015", "platform": "google", "record_type": "place", "published_at": null, "text": "Impress Porto\n\nOrtodontista\nR. de Sá da Bandeira 676, 4000-431 Porto", "rating": { "value": 4.9, "max": 5, "count": 2824 }, "engagement": { "comments": 2824 }, "geo": { "country": { "value": "PT", "source": "post_geotag", "confidence": 0.9 }, "place": { "name": "Impress Porto", "country": "PT", "source": "post_geotag", "lat": 41.1520273, "lng": -8.6059457, "address": "R. de Sá da Bandeira 676, 4000-431 Porto", "category": "Ortodontista" } }, "judgments": { "relevance": null }, "matched_keywords": ["orthodontist"], "provenance": { "endpoint": "/v1/google/business/search", "request_id": "req-…", "captured_at": "2026-10-18T10:30:30.378Z" } } ``` - A places search needs `near`. Instagram finds its places by name, so with Instagram `near` must be a place name. - Businesses are kept without a relevance check, so a `context` never drops them and nothing is charged for one; a `label` you ask for still judges them, and the posts at an Instagram place are judged like any post. `summary.record_types.place` counts them. - Instagram's places named for `near` are checked for being there (a condo in Florida named after Bali is not walked; 1 credit per 25 checked), and the posts at up to three of them are walked newest first back to `since`. Each post carries its place as `geo.place` and its country from it. A busy place fills a page in about a day, so a month in the past is reached through quieter places or a background run; `data.platforms.instagram.window` says how far each walk reached, and when Instagram's later pages fail upstream a `_warnings` line says how many walks stopped at page 1 and that a resume with `retry_failed=true` reads further. - Prices are the lanes' own: a place search 5 credits for up to 20 businesses, an Instagram place search 5, and 5 a page of about 60 posts at a place. The estimate and `dry_run` price every page a walk may take. ## Next steps - [Market research](/docs/prism/discover-research.md): Compare one topic across markets, with sites and search interest. - [Accounts and communities](/docs/prism/discover-accounts.md): Find who writes about a topic and where it lives. - [Discover request fields](/docs/prism/discover.md): Every request field, preset and canonical request.