If you felt your in-app campaigns started “finding” better users faster this summer, you’re not imagining it. Two signals stand out: (1) Unity’s new AI platform Vector helped its Ad Network grow +15% QoQ in Q2 and now accounts for ~49% of Unity’s Grow revenue; (2) AppLovin posted blowout Q2 results on the back of its AXON 2 engine, with revenue up 77% YoY and EBITDA margins above 80%. Together, they confirm that AI-driven optimization is moving the goalposts for how we buy mobile traffic.
What’s actually changed
- Faster, deeper modeling. Vector and AXON 2 leverage broader cross-app signals and self-learning models to predict value earlier in the funnel (beyond day-1 installs or single postbacks). That means smarter bid/placement decisions at impression time and more stable CPIs at scale.
- Stronger sequential performance. Unity explicitly tied its Q2 ads growth to Vector’s lift, a rare, clean attribution from a network to its AI stack useful for forecasting.
- Network economics favor AI. AppLovin’s results indicate that when models hit critical mass, margins expand and inventory clears more efficiently often translating to steadier eCPMs for publishers and more predictable ROAS for buyers.
What it means for your buying strategy
- Restructure budgets around AI “learning momentum”.
Instead of thin, many-geo testing, concentrate early spend to help models stabilize:
- 60–70% budget into 3–5 priority geos per OS until you hit stable CPI and pLTV variance <10%.
- Only then expand breadth. Vector’s sequential lift data supports this phased scaling logic.
- Optimize to value, not vanity.
Both stacks reward signals that correlate with payback (subscriber trials, mid-core milestones, high-intent events). Replace generic “install” goals with composite events (e.g., Install + Session2 + Key Action within 24h) and align tROAS windows to when your model can first separate whales from window-shoppers. - Shorten learning loops.
Move from weekly to 48–72h decision cadences. Use rolling cohort readouts (D0–D3 value density) and kill under-performing bundles fast; the networks now adapt quickly enough that slow buyers simply fund competitors’ models. - Creative feeds the model.
AI thrives on variety. Maintain 5–8 concurrently active concepts per network (UGC, utility demo, problem-solution, social proof), each with 3–5 variations. Tag creatives to post-install behavior (e.g., “saves content,” “completes level 3”) so you can map which narratives generate the highest predicted LTV. - Segment iOS vs Android by signal reality.
On iOS, pair network AI with privacy-preserving measurement (AAK/SKAN); on Android, prepare for Sandbox constraints. Expect networks to lean harder on their own modeling, making first-party conversion quality and clean event taxonomies non-negotiable. (Unity and AppLovin commentary suggests networks will keep closing the gap with their internal signals). - Re-evaluate remarketing vs UA split.
As models improve at predicting value from top-funnel signals, some verticals will see prospecting ROAS close the gap with re-engagement. Test aggressive prospecting in the same geos where your remarketing already works; monitor overlap and incrementality. - Raise the bar for forecasts.
Stop using static CPI/ROAS tables. Your Q3 planning should incorporate model velocity: how quickly each network reaches stable CPA for a new geo/app bundle and how ROAS curves change after the first $5–10k of spend.
A practical 14-day playbook
- Day 0–2: Define 2–3 value events and instrument clean postbacks; prep 6+ creative concepts.
- Day 3–7: Concentrate spend in 3 geos/OS; target $3–5k per bundle to reach model stability; pause bottom decile placements daily.
- Day 8–10: Shift 15–25% budget to look-alike geos with similar monetization; keep creative refresh cadence at 20–30% per week.
- Day 11–14: Promote winning bundles; widen ROAS window if payback skews late; lock budgets for scale.
Bottom line
AI-powered networks are no longer a “black box” gamble, they’re becoming the default path to efficient scale. Treat Vector and AXON-era buying as a system problem: cleaner signals, tighter feedback loops, and creative breadth. Teams that adapt their operating rhythm to the networks’ faster learning cycles will win the auctions and keep the gains.
