Searching X Bios for "Meta AI" Returns 14,673 Hits. So Does Searching for Just "AI."

"Which AI lab has the most people on X?" sounds like a question you could answer by searching bios for company names and counting. We had 697,439 non-empty X bios in a search index and tried exactly that. The search technology broke the question before the data could.

The quotes don't do anything

The X users search endpoint takes a q parameter documented as full-text search over username, name, bio and location. It is OR-token search — it matches bios containing any of the words you typed, not the exact phrase, and wrapping the query in quotes changes nothing. Search "Meta AI" and you are really searching Meta OR AI:

"AI" alone (control)14,112

"Meta AI" (quoted)14,673

"Stability AI" (quoted)~14,200

"Google DeepMind" (quoted)2,218

"Meta AI" returns more hits than plain AI — the query is matching every bio with the word "AI" in it, plus a handful with "Meta." "Stability AI," "Perplexity AI," "Mistral AI," "Moonshot AI" and "Microsoft AI" all land in the same 14,000–15,000 range for the same reason: any two-word name where one word is "AI" is functionally indistinguishable from searching "AI" by itself. "Google DeepMind" is a smaller number but the same failure — 2,218 hits, dominated by the common word "Google," not the distinctive one. None of these are usable as a company-mention count. Quoting a phrase, an intuitive way to search for an exact company name, silently does nothing on this endpoint.

What's left: single distinctive tokens, and how much of each is actually real

Drop the common-word half of the name and query the distinctive token alone — DeepMind instead of "Google DeepMind", OpenAI instead of "Open AI" — and the numbers become small, specific, and checkable by hand:

OpenAI373

DeepMind150

xAI79

Anthropic51

Cohere31

Mistral24

DeepSeek10

Qwen8

Small numbers are countable, so we read every hit for each token — up to 30 per token, or the full set for tokens under 30 — and sorted each into "genuinely lab-related" (the official account, or a bio naming a role there) versus "collision" (an unrelated person, place or product that happens to share the word):

Two things worth sitting with. First, verification rate does not track how "real" a token is — Mistral and xAI post similar blue-check rates to the clean tokens, because verification tracks whether someone paid $8, not whether their bio is about the company you searched for. Second, the fraction of hits that are an actual employee, rather than a fan, a user or an unrelated collision, falls off a cliff for the two smallest tokens: DeepSeek and Qwen both source-check at roughly 20–40% employee-related, versus 90%+ for OpenAI and DeepMind.

Two tokens we are not publishing a count for

Kimi (Moonshot AI's model) returns 17 bio hits. We are not reporting that as a mention count, because reading the hits shows almost none of them are about the AI lab — Kimi is a common given name and a common X handle, and the sample was dominated by unrelated people, a Japanese actor, a music label fan account and a professional-dominatrix clip producer using it as a nickname. Two hits were genuinely Moonshot's official product accounts.

Moonshot (as in Moonshot AI, the lab behind Kimi) returns 8 hits — and reading them shows the token is worse than useless: a podcast network, a film production company, an African startup-conference brand, a biology research nonprofit, a nanotechnology hobby project, and Google X's own "the moonshot factory" tagline all outrank the AI lab. Only one of the eight is genuinely Moonshot AI. Publishing "Moonshot: 8 mentions" without that context would misrepresent the finding more than not publishing a number at all.

The honest shape of this, not a leaderboard

It would be easy to turn the token table above into "OpenAI beats Anthropic 373 to 51 on X" — and that framing would be wrong on every axis that matters. The counts are not comparable across labs (OpenAI is older, bigger and has more ex-employees who kept the mention in their bio), they are not a headcount (a 373-bio count against a lab with thousands of employees is a rounding error, and most bios never mention an employer at all — see our study of how the X bio itself is disappearing), and — as the section above shows — a meaningful share of even the "clean" tokens are not employees at all.

What the numbers do show, reliably: X bios are a weak and heavily skewed signal for "who works at an AI lab." The absolute counts are tiny against a 697,439-bio corpus. The search technology cannot separate a two-word brand name from the common word inside it, which rules out most of the labs anyone would actually want to look up. And the labs that show up at all skew hard toward Western, English-language, developer-and-researcher circles — DeepSeek, Qwen, Kimi and Moonshot are barely visible here not because Chinese labs have fewer staff, but because this index is seeded from Wikidata, GitHub and X's English-language search graph, which under-samples exactly that population. That skew, not a real headcount gap, is the more honest story.

What this shows, and what it doesn't

What we think holds up: single distinctive brand-name tokens (OpenAI, DeepMind, Anthropic, Cohere) return small, mostly-genuine sets of bio mentions, and the share of genuine hits varies a lot by lab — from ~90%+ down to ~20–40% for the smallest tokens. Two-word names collapse into search noise and are not usable at all. Two tokens (Kimi, Moonshot) are dominated by unrelated collisions and we report them as unmeasurable rather than give a misleading count.

What limits it:

• This is a bio-mention count, not an employee census. A mention is not proof of current employment (many hits say "prev" or "ex-"), and the absence of a mention is not proof someone doesn't work there — most bios don't name an employer at all.
• The employee-relevance shares in the table are hand-read, small-sample estimates, not a systematic classifier — read against 8–30 hits per token, with wider uncertainty for the smallest tokens (DeepSeek, Qwen).
• The dataset is a seeded index of established/discoverable X accounts, not a random sample or census — see the X users dataset for the seeding and its Western/English skew. That skew plausibly explains most of the gap between Western and Chinese lab visibility here.
• We publish aggregates only — bio-mention counts and verification rates by token, no individual accounts or bio text tied to named people.

Sources

Methodology

We queried the X users search endpoint's q parameter (full-text over username, name, bio and location) with has_bio=true for each candidate AI-lab name, first as a quoted two-word phrase, then as its single distinctive token. The endpoint documents q as full-text search; empirically it is OR-token search over the words in the query, so a quoted two-word phrase where one word is common (like "AI" or "Google") returns a result set dominated by that common word rather than the phrase. We confirmed this by comparing quoted two-word queries against a control query for the common word alone, and found near-identical totals ("Meta AI": 14,673 vs. AI alone: 14,112).

For each single-token query that returned a total under 30, we read every result; for tokens with 30 or more, we read a 30-result page sorted by relevance. Each hit was classified by hand as genuinely lab-related (the official corporate account, or a bio naming a current or former role there) or a collision (an unrelated person, place, product or use of the same word). We also pulled the is_blue_verified facet for each token. Corpus size (796,611 total profiles, 697,439 with a non-empty bio) and all counts were queried live and are current as of this study; a dataset this size and query surface changes over time, so a re-run may return slightly different totals.

Want to run your own cut? The X users dataset documents the query pattern and its Western/English-language skew in more depth; our study of how the X bio itself is changing covers the broader decline of the "job title + employer" bio format this study depends on.