---

name: anysite-company-sourcing description: The craft of sourcing companies from Anysite's 70M+ LinkedIn company database (search_sql_companies) - turning a fuzzy ICP into per-field DSL filters that return real matches instead of token soup. Fixes the default failure mode where naive keyword queries return companies from the wrong country, wrong industry and stub pages. Use when the user asks to find/source companies, build a company list, complains that company search results are bad or irrelevant, or needs a target-account universe - "найди компании", "плохие результаты поиска компаний", "source accounts". For people at those companies use anysite-people-sourcing; for funding-stage filters use crunchbase; for pushing into a CRM use anysite-crm-prospect.

Company Sourcing

linkedin/search/search_sql_companies searches 70M+ companies and is the single best bulk company tool in the catalog — and its naive use is the single most common source of garbage lists. Measured live, same intent, same day:

• {keywords: "AI startup San Francisco"} → 1 relevant of 5: a Tel-Aviv gaming-data firm, a Sydney fintech and a Ho-Chi-Minh beauty e-commerce all matched — each had the token "startup" somewhere in its description and "San Francisco" among secondary office locations.
• The same intent as structured filters (below) → 5 of 5 genuine SF AI companies in the right size band.

The difference is the whole skill. Never ship results from a naive query.

Why naive queries fail (mechanics, not opinion)

1. keywords matches whole words across ALL text fields — name, description, specialities, hashtags and the locations array — so "San Francisco" as a keyword still matches a Hanoi company that lists an SF sales office in locations. (Whole-word matching removed the inner-substring noise, but not the wrong-FIELD problem — HQ is a separate field, use it.)
2. sort changes ordering, not the candidate set — a naive keyword query is still not an ICP list. With relevance (the default), "AI startup San Francisco" surfaces Startup Weekend AI, Bitcoin AI Startup Lab, 1-employee shops and a Phoenix-HQ company (all measured) — they score high because the words sit in their NAME. The fix is per-field decomposition below, not a sort flag.
3. Millions of company pages are stubs. Without hygiene filters they dominate.
4. employee_count_range can contradict employee_count in the same record (measured: 305 employees with range "11-50"). Never filter or segment by the range string.

The method: decompose intent into fields

Take the user's ICP sentence apart and map each fragment to its OWN field. No ICP in the request → take it from the ICP section of anysite-gtm-profile when it exists (and say so in one line), including its exclusions.

Intent fragment	Field	Notes
"based in X"	country_hq: ["US"] + headquarter_location: "\"san francisco\""	token-aware; NEVER the locations field — that matches branch offices
"present in X" (offices count too)	country_any	this is the only right use of the locations array
"does AI / fintech / logistics"	specialities OR industry_name OR description together, not specialities alone	industry array wants URNs; industry_name resolves labels — see the specialities caveat below
"in the orbit of company X"	similar_organizations: "\"fsd_company:<id>\""	queryable filter, not just an output field — the reverse-graph expander; see below
"startup / SMB / enterprise"	employee_count_min / employee_count_max	integers; ignore employee_count_range entirely
"founded recently"	founded_on_min	year
named company lookup	name or alias DSL + exact verification	never trust first hit; a known domain → companies/resolve (anysite-mcp)
always, every query	is_active: true, has_website: true, min_description_length: 100	the hygiene trio kills stubs and dead pages
ranking	sort = relevance (default) or last_modified	relevance for sourcing; last_modified for "what's new since last run" (monitoring). Scoring weights a term by field: name 5× > specialities/hashtags 3× > short_description 2× > long description 1×, length-normalized; ties broken by recency. Caveat: the score is built only from keywords/name/specialities/description — a query filtered ONLY by non-text fields (e.g. just industry + employee_count_min) has nothing to score, so it falls back to recency order

Industry labels are often wrong or empty (auto-created LinkedIn pages especially). When missing a company matters more than extra noise, run a second query with the description terms and NO industry filter, merge the two (merge_data with dedupe_by: ["urn"]) and decide each row with one yes/no question — the wide-net method in anysite-crm-lookalikes.

DSL in every text field: whitespace = AND, | = OR (no spaces around it), "phrase" = exact phrase / substring, -token = NOT. Example: specialities: "\"artificial intelligence\"|\"machine learning\" -agency". IRON RULE: quote every multi-word alternative in an OR chain. Whitespace binds tighter than | — name: "Level Infinite|Proxima Beta" parses as (Level) AND (Infinite|Proxima) AND (Beta) and returns 0 (verified live on the sibling people endpoint, same parser). Lint before sending: a space inside an OR alternative without quotes → fix first.

Matching semantics (changed — this is now the biggest lever):

• A bare word matches as a WHOLE WORD, not a substring (verified live: keywords: "sdr" returns "SDR Academy", "SDR Foundry", "Vida SDR" — companies with the standalone word, and no longer "adviseur"-style inner hits). Case doesn't matter.
• No stemming. integration does NOT find integrations — write both: integration|integrations. Same for singular/plural and verb forms.
• Quotes = substring (the old behaviour, kept). "integr" matches both integration and integrations; "machine learning" matches that exact sequence. Use quotes deliberately when you WANT a fragment.
• Terms with separators (b2b-saas, S.E.E.D., x.com) and Cyrillic are auto-substring — nothing to change for them.
• alias and website are untouched — still substring (alias: "openai" finds openai-inc); hashtags is exact element match.

Practical consequence: short single-word queries used to be the WORST case (inner substring noise); they are now precise. If you actually need "a piece of a word", reach for quotes.

Specialities caveat: self-declared tags are the highest-signal field WHEN present, but hot young startups often leave them blank — measured: 2/10 in a US/Software slice had empty specialities[], one of them Hebbia ($160M-funded, empty short_description too). So specialities is a widening OR alongside industry_name and description, never the sole gate, or you silently drop exactly the fresh-funded targets a list is built for.

The loop (never skip step 3)

1. Build the per-field query from the decomposition above.
2. Probe with count: 10.
3. Validate against the intent, not the filters: for each of the 10 check HQ country/city, what the company actually does (short_description), and employee_count. Fewer than ~8 on target → the query is wrong, not the data: move misused fragments to their proper field, add -tokens, tighten phrases. On target → widen coverage with | synonyms in the same fields ("machine learning"|"computer vision"|"nlp"), re-probe.
4. Fetch the real volume (count up to 1000). More than 1000 matches → split by size bands / countries / founded ranges into disjoint queries.
5. Free re-cuts: query_cache on the result for sorting, counting, sub-segmenting — don't re-execute.
6. Show it (clients that render MCP Apps): show_entity_table(cache_key, title=<intent>) instead of pasting rows; when the user wants to qualify the list one by one, review_leads(cache_key) — the approved companies come back as a cache_key for people sourcing or CRM push (anysite-mcp → Tables, joins and lead review).

Two hard limits to state up front (TAM planning)

Unlike people search, company search has no bucket_total, and its dry_run count-only mode does not reach you through the MCP (the count travels in a response header; the call returns an empty list). That means:

• You cannot ask "how big is my ICP universe" cheaply — there's no total count without pulling rows. Size the split blind, or probe representative sub-slices and extrapolate; tell the user the number is an estimate. (Still true regardless of sort.)
• Plan the universe from the target backwards when the user asks for TAM or tiers: accounts needed ≈ revenue goal ÷ deal size ÷ the win rate from contacted account to closed deal (ask for theirs; say it is an assumption if they have none). Then tier what you found — a few dozen 1:1 accounts, a few hundred 1:few, the rest 1:many — rather than handing over one flat list.
• A >1000 match returns only the top 1000 by rank, so full coverage still needs splitting by size/country/founded into sub-queries each < 1000. With the relevance default the top is at least relevance-ordered rather than a pure recency head, but you still don't see rows 1001+. Note card freshness lags (employee_count drifts from reality on some records). Dedup on band boundaries by urn (query_cache uniq). last_modified_after + sort: last_modified is the right combo for a "new since last run" sweep.

Read the bonus fields — they're the handoff

Every row carries, at no extra cost (names as in the call's table rows; full records from get_page use organizational_urn / website / crunchbase_link):

• company_id → feeds anysite-people-sourcing (current_company_id) and search_jobs directly;
• domain → domain for CRM matching and current_company_domain people filters;
• crunchbase_alias → free crunchbase alias, skip the live 20cr search;
• similar_organizations[] → both an output list AND a queryable filter (the bigger lever — see below);
• specialities[] → the company's own vocabulary, reuse it to widen synonyms.

similar_organizations as a filter — the reverse-orbit expander

Verified live: similar_organizations: "\"fsd_company:1441\"" (Salesforce) + country_hq:["US"] + size band returned 8 companies, all 8 carrying Salesforce in their own similar-orgs graph. This is "who sits in the orbit of company X" — the right tool for two jobs no keyword query does well:

• ICP expansion from a seed account — feed your best customer's fsd_company id, get its competitive/adjacent set.
• Competitor-adjacency lists — the closest this skill gets to "companies like my competitor". (It is NOT the competitor's customers — for that, anysite-crm-competitor-intel via wappalyzer/reviews.)

Noisy — roughly 3/8 were on-target in the test — so always combine with industry_name/specialities and run the probe-validate loop. Batch-fetching the output similar_organizations[] by urn:[...] (≤12 at a time) is the weaker, one-hop version; the filter is the scalable one.

When a different tool is right

• Funding stage / investors / valuation → crunchbase/db/db_search (dates as unix ts, count ≤100) or live crunchbase/search (hiring, it_spend filters).
• One known company by domain → companies/resolve (exact domain; recipe in anysite-mcp — several candidates can claim one domain, pick the right one).
• Early-stage / launches → yc/search/search_companies, producthunt, betalist.
• People at the sourced companies → anysite-people-sourcing with the URNs/domains from step "bonus fields". CRM push → anysite-crm-prospect (dedup + create rules).