What a network outage teaches you about dependency

Ten days after the ironSource Ads network went dark, the buyers who feel it least are not the ones who spent the least there. They are the ones whose remaining channels could do the same job. That distinction is the whole argument of this piece: dependency is not the share of your budget a channel takes, it is the time and money it costs you to replace what the channel was doing. Most concentration dashboards measure the former and call it the latter.

This is a claim a sensible head of growth can push back on. Spend share is easy to compute, easy to govern with a rule ("no channel above 40 percent") and correlates loosely with risk. But the shutdown Unity announced in March and completed on April 30 has given the industry an unusually clean natural experiment, and the results do not line up with spend share. Some studios with a modest ironSource line item are still scrambling. Some with a large one had moved on inside a fortnight.

Why the budget number lies

An ad network is not a single service. When you buy through one you are consuming at least four things bundled together: access to a set of publisher inventory, a bidding model trained on your conversion history, a creative delivery format your production team has tuned for, and a stream of attribution signals (SKAN postbacks, Google referrer data, MMP callbacks) that your reporting has learned to interpret.

When the network disappears, each of those has a different replacement cost.

• Inventory usually migrates. Publishers reroute demand to other mediation partners, and the impressions reappear elsewhere at a different price.
• The trained model does not migrate at all. Every network that picks up your spend starts its learning phase from scratch, and your first weeks of cost per install reflect its ignorance, not the market.
• Creative formats migrate slowly. A playable that was built to one network's specification often needs re-export, re-QA and a fresh round of size trimming for another.
• Attribution signals migrate messily. Your SKAN conversion value schema was tuned to one network's postback behaviour and the campaign structure that produced it.

A channel at 15 percent of spend that owned 40 percent of your learned optimisation and all of your interstitial playable inventory was a much larger dependency than the budget line suggested. Spend share reported it as small. The outage reported it correctly.

The mechanism nobody budgets for

Here is the part most post-mortems skip. When a network shuts, the spend that leaves it does not vanish. It lands on the surviving networks, and it lands there at the same moment for every advertiser who was on the closed platform. Auctions clear at higher prices for a period, not because supply fell but because demand converged.

The buyer who moved early paid the old price. The buyer who waited for the deadline moved into an auction already crowded with everyone else who waited. This is a coordination problem, and it rewards the studio that treats a shutdown announcement as a start date rather than an end date. Our own migration checklist published the day Unity made the announcement said as much, and the studios that acted in the first week have reported the smoothest curves.

There is a second-order effect that runs the other way, and it is less comfortable. The networks that absorbed the demand have gained pricing power. Fewer meaningful direct-demand networks means each of the survivors has a stronger hand at rate review, and the automated bidders that now dominate the category have less competitive pressure on their margins. Diversification advice tends to assume that alternatives exist at comparable terms. After a consolidation event they exist, but the terms have moved.

A dependency score you can actually compute

Spend share stays in the model, but it becomes one input among five. For each channel, score the following from one (low dependency) to five (high):

• Substitution time. How many weeks until a replacement channel reaches comparable efficiency at comparable volume? Under two is a one, over eight is a five.
• Unique inventory. What share of the installs this channel delivers come from publishers you cannot reach elsewhere at all?
• Learned optimisation. How much of the channel's performance sits in a model you do not own and cannot export? A fully automated bidder with no manual controls scores high here.
• Format lock-in. How many of your active creatives would need rebuilding or re-QA for another network's specification?
• Data exit terms. Can you pull historical campaign, creative and attribution data after the relationship ends, and for how long? Unity's shutdown FAQ gave advertisers until May 31 to retrieve ironSource history; that window is a five-week reprieve, and not every vendor offers one.

Multiply the average of those five by the channel's spend share and you get a weighted dependency figure that behaves very differently from the raw budget number.

An illustrative example: a mid-sized puzzle studio has 18 percent of spend on a network that scores four on substitution, four on learned optimisation, three on format and one on inventory and exit. Average 2.6, weighted 0.47. The same studio has 35 percent on a self-attributing platform that scores two on substitution, three on optimisation, one on format, one on inventory and two on exit. Average 1.8, weighted 0.63. Higher spend, higher weighted score, as you would expect. But a third channel at 9 percent of spend scoring five, five, four, four, five comes out at 0.41, almost level with the 18 percent channel and well above what a spend cap would ever flag.

That third channel is the one that hurts when it goes down.

The trade-off diversification advice ignores

Reducing dependency has a cost, and it is not a small one. Every additional channel you run at meaningful scale is a set of creative exports, a QA lane, a reporting integration and, above all, a bidding model that needs enough conversions to learn. Spread the same budget over eight networks and you may find none of them ever exits the learning phase. The over-diversified buyer is paying a permanent efficiency tax to insure against an event that happens perhaps once every two or three years.

The honest approach is to decide which dependencies you are willing to carry. A channel scoring high on learned optimisation but low on everything else is a reasonable bet: the model is replaceable given time, and time is the resource you can plan for. A channel scoring high on unique inventory and exit terms is the one to reduce or contract around, because when it closes, the audience is gone and so is your record of what worked with them.

Run the outage before it runs you

Nobody knows their dependency score until a channel actually stops. Estimates made in planning meetings this spring were wrong in both directions.

So schedule the outage. Once a quarter, pick one channel above ten percent of spend and pause it for seven days while holding total budget flat. Watch three things: how quickly the reallocated spend reaches the paused channel's efficiency, whether total installs hold or drop, and which creatives cannot be moved. Record the substitution time and the efficiency gap in the dependency score.

You will lose some efficiency in the drill week. That is the premium on the insurance, and it is cheaper than discovering the answer with a deadline attached.

Related archive reading

These articles provide related context and remain subject to their stated review status.

• Unity is shutting the ironSource ad network. A 30-day migration checklist for buyers
• Mobile Ad Network Spend Share: Q3 2025, Network by Network