The Ai4 roster exercise proved a sequence: take a population, gate it on a hard ICP test, enrich it, then work it. That population arrived pre-qualified, because everyone on it had chosen to attend an AI conference. This time I had to build the population myself, from a database, for a North America ABM programme. What follows is the full log: the saved searches, an honest read of what each signal proves and misses, and the six rounds of testing that moved net-new precision from roughly 2 percent to 31.
Ask for technology instead of industry labels
An industry field is somebody's description of a company. A technology install is
the company's own behavior. Ask Apollo for the “computer games” industry and you get
whoever chose that label. Ask for companies currently running AppsFlyer or Unity and
you get companies that pay real money to measure app installs or build in a game
engine. That was the founding idea of this list: filter on
currently_using_any_of_technology_uids, US and Canada, and let the tech
stack testify. All counts below are Apollo company-search totals for the exact
filters described, pulled 2026-08-03, so treat them as point-in-time estimates of a
database rather than of the market.
The saved searches and what each one proves
The mobile attribution stack (AppsFlyer, Adjust, Branch, Kochava, Singular) returned 2,179 companies and is the best single filter, because attribution exists only to measure app installs. Game engines (Unity, Unreal) returned 2,607, and the honesty note matters here: an engine proves games or real-time 3D, without distinguishing mobile from console and PC, and Unity now shows up in industrial simulation, architecture and automotive. On its own that search is no mobile-game list. Mobile ad monetization returned 203, the sharpest and narrowest signal, with a correction I only caught later: of the seven ad networks I listed, Apollo recognized just AppLovin and ironSource, so the 203 reflects two vendors rather than seven.
The combined ABM-sized search, a union of those signals with sub-11-employee
shells removed, returned 3,163. A union rather than a sum: the individual searches
overlap heavily, which is why 2,179 + 2,607 + 203 lands nowhere near 3,163. And one
search got thrown out entirely: mobile dev frameworks (React Native, Flutter,
Swift, Kotlin and friends) returned 81,119, because Apollo resolves
swift to the SWIFT banking network rather than Apple's language. That
filter is contaminated beyond use.
The exclusivity test sorts every signal
A signal only works as a filter if a company cannot have it without shipping an app. Product analytics (Amplitude, Mixpanel, Heap, Pendo), engagement (Braze, Airship, Iterable, CleverTap), experimentation and CDPs are the strongest commercial signals we have at ThinkingAI, since a company paying for them has already bought a piece of the loop we sell whole. But web-only SaaS companies are heavy users of all of them, so filtering on them pollutes the list badly. They belong downstream, as tags within an already-exclusive set.
Apollo also has a quiet tell worth knowing. A technology it recognizes comes back
with a proper display name (appsflyer returns “AppsFlyer”), while an
unrecognized one is echoed back raw (revenuecat returns “revenuecat”)
and silently matches nothing. A zero means only that Apollo does not track the
tool; it says nothing about whether companies use it. Sorted on exclusivity and recognition: ASO and store
intelligence scored 1,220 and is clean, since you only buy store intelligence if
you are in a store. Mobile build and release (Bitrise, TestFlight) scored 583, very
clean and smaller. Mobile crash tooling is effectively unavailable at 8 companies.
And subscription and IAP infrastructure (RevenueCat, Adapty, Qonversion), the
highest-exclusivity signal there is, has zero recognized vendors: a real blind spot
for subscription apps. The verified exclusive union of the 13 signals that passed
both tests, at 11+ employees, counted 2,671. Hold that number loosely; it comes
back up.
Ad monetization cuts both ways
The ad-network signal looks precise and is a trap on its own. AdMob sits in news, weather, utility and content apps, close to noise by itself. AppLovin spans games and apps. Only Chartboost, Tapjoy, Unity Ads and ironSource skew hard toward games. Worse, the signal misses the games we most want: mid-core and hardcore titles (RPG, strategy, simulation) monetize primarily through in-app purchase and often carry no ad SDK at all, and those are exactly the studios whose economy, retention and LTV problems match our product best. Filtering on ad tech would screen out our strongest ICP.
So the IAP studios have to be inferred, because “mobile game monetized by IAP” is no technology tag. Game backend and live-ops stacks are the best proxy: a company running PlayFab, Photon, Nakama or Xsolla operates a live game with an economy, ads or no ads. Engine plus attribution (Unity or Unreal alongside AppsFlyer or Adjust) is a strong mobile-game signal, because a console-only studio has no reason to buy mobile attribution. Attribution and engine with no ad SDK is the tell for a pure IAP title. One tooling constraint shapes all of this: Apollo's technology filter is OR within a single query, with no way to express “Unity AND AppsFlyer” in one call. Intersections have to be computed locally, which is why the sequence has to be pull broad, then classify.
Six rounds of testing reversed the conclusion
Everything above was written before the filters met real output. When I sampled 100 companies at a time and hand-checked them, the technology filters produced lists that were 70 to 98 percent junk, and the “2,671 verified universe” turned out to be a count nobody should quote. The reason: Apollo's technographic detection is far looser than its clean display names suggest. A recognized display name only proves Apollo knows the word. Attribution tags fire on any company that advertises, and Apollo attributes any app in a large enterprise's portfolio to the entire enterprise, which is how $10B+ conglomerates, an international aid organization and a federal government department ended up on a “mobile app” list.
The rounds, each measured on a deep page for net-new precision: v1, technology plus 11+ employees, 2,671 companies at roughly 2 percent, with 75 of the 100 sampled being $10B+ enterprises. v2 added a revenue cap and hit roughly 30 percent, on a pool of 421 that was too small to run a programme on. v3 kept headcount only: 746 companies, 15 percent, since dropping the revenue cap let heavy industry back in. v4 gated on NAICS 51321 and scored 24 percent on just 98 companies, the wrong gate because “Software Publishers” admits all B2B SaaS and excludes game studios. v6 dropped technology for keyword tags: 11,572 companies at 11 percent. v7 trimmed the keywords and excluded services SIC codes: 1,013 companies at 31 percent, adopted, 2.8x better than v6 on the number that matters.
The cost math before any export
Counting companies is cheap. Getting people is charged per person. Covering the 3,163-account combined search at two or three decision-makers each runs to roughly 6,000 to 9,000 enrichment credits, against a balance in the low thousands: the full universe is unaffordable in one pass. That is precisely why the order is classify first, enrich second. Tier the accounts on free company data, then spend credits only on the top tiers. The Ai4 run is the precedent, where roughly half of attempted matches returned a usable contact, so the budget has to assume a meaningful share of spend returns nothing.
What I would do instead of a perfect query
Stop optimising the filter. Six rounds moved net-new precision from 2 percent to 31, and the remaining error is Apollo's data quality rather than filter syntax; another round is unlikely to beat 31 by much. Pull broad with the adopted filter (1,390 companies pulled and kept locally), then have a cheap model label every row: mobile game, PC/console game, mobile app, vendor, agency, out of ICP. Sorting a thousand rows costs cents, and a model reading a company name and domain judges better than any keyword tag. And remember where the good list came from: Ai4 worked because the population carried its own qualification. For the next list I would start from a source that does the same, an app-store publisher ranking, a conference roster, an industry association roll, and save the database for filling gaps.