A bid crosses a few exchanges and lands at a trade that’s already counting. Consumers shopping for real-time bidding services (RTB) usually ask about the reach and the price per 1,000 concepts. Hardly anyone asks how long the machinery takes to answer, though RTB service providers live or die by that number. The auction doesn’t wait politely.
Delay is estimated to cost about 10% of real-time bids, but to be exact, not impressions, the offers that never made it into the room before the door closed.
Say, purely for picture, that a buyer plans to spend $200,000 a month across open bidding. If 10% of the bids arrive after the exchange stops listening, about $20,000 of that budget loses its seat. Nobody gets an alert; a bid that never reached leaves no row in the report, so there’s nothing to inspect. Over 12 months, the gap grows to $240,000 of deliberate reach. It’s a dripping tap in an empty house: nobody hears it, and the bill still shows up.
The money itself rarely disappears. It drifts into other auctions at whatever price those happen to ask, which is rarely the price the plan had in mind. Budgets behave like water that way. They find the next open channel, and the record tells an orderly story afterward. A finance director sees the total spend and moves on.
Extend the clock for a moment. Suppose the same bid takes 1 second to answer rather than 100 milliseconds. PPC Land reports that Google’s Authorized Buyers documentation sets a response time limit between 80 and 1,000 milliseconds, depending on format and bidding type. So a 1-second reply sits at the very edge of the most charitable window. Against the tightest, it’s hopeless. A 100-millisecond reply fits most windows, though the 80-millisecond ones still turn it away.
Somewhere, on a Tuesday, an analysis meeting is running long. The team studies a sliding win rate. They blame the creative, then a competitor with deeper pockets. Nobody brings up the clock, because the clock leaves no fingerprints. It isn’t on the schedule, and nobody thinks to put it there. By Thursday, the creative has been rewritten. The attempts are still arriving after the doors have closed.
Exchange logs file a late bid under delays. That category sits next to errors, and people scan it on the way to the win rate. The win rate looks healthy whenever the auction that did arrive happens to be a good one. The ones that never arrived belong to no report. Only the bid rate, reasoned against requests received, shows the gap. Few control panels put it on the front page. Most reviews never ask.
Late bids rarely come from a single dramatic failure. Mostly, a slow accumulation, a few milliseconds here and a few there, like small debts. The leaks hide in places a contract rarely mentions:
Speed buys thinking time, too. Every audience score and brand-safety search eats into the same window. A faster pipeline leaves room for cleverer bidding, while a listless setup cuts corners quietly, in the code, where nobody outside engineering looks. Engineers feel the swap each time a new data source arrives asking for a few milliseconds of its own.
Sellers have noticed that on October 23, 2025, Amazon Web Services launched RTB Fabric, a managed private network. The company says it delivers a delay of just a few milliseconds between ad tech partners and cuts back cloud networking costs by up to 80%. Weeks later, IAB Tech Lab put an independent RTB requirements out for public comment. Its chief executive pegged typical response times at 400 to 600 milliseconds, and the plan trims on containers that let buyer and seller technology share a server to cut the setback. Choices like these shape what real-time bidding services can honestly promise. They also explain how 2 providers citing the same 100 milliseconds can deliver very different bidding.
A provider asked for its common response time will answer fast, with a number rounded kindly. Averages commend everyone. The honest figures tend to be the ugly ones. Better to ask for 95th-percentile response times and break rates by exchange. The slow tail is where windows get missed.
Then there’s the test. Similar budgets sent to 2 providers for 2 weeks, with every break logged, make the comparison hard to spin. A fair test also fixes the exchanges and the hours of the day, so luck has less room to toady either side. Providers confident in their speed infrequently object.
Faster isn’t free. Servers near swap cost more than servers parked anywhere cheap, and thinner decision logic gives up some precision for speed. Cheap racks in an isolated region look like savings on the invoice and read as lost bids in the logs. A buyer chasing confined audiences may accept slower replies, while a volume buyer after broad reach likely shouldn’t. Neither choice is wrong, as long as it’s made on purpose.
Geography earns a question too. Screencore, for example, works from offices in Warsaw and Singapore and says it focuses on Central and Eastern Europe and Southeast Asia. That tells a buyer where its focus lies. Whether its servers sit close to those audiences is the sequel worth making. A map on a sales page proves little; a log file proves more. Among real-time bidding services, the dependable ones publish respite rates without being chased, and they explain the quiet days as gladly as the busy ones.
A bid that takes 1 second mostly never arrives. The 10% loss infrequently announces itself. That’s why the choice of a bidding partner deserves more attention than any pitch card invites. Somewhere tonight a window will close a few milliseconds before a bid is still on its way. The auction, as ever, won’t wait.
Ans: 100 milliseconds means a system handles 10 bits per second per channel. If it jumps to 1 second per bit, it handles only 1 bit per second-a massive performance drop.
Ans: Transmitting a simple 8-bit text character would take a full 8 seconds instead of less than a second. Standard web pages, which consist of millions of bits would take days or weeks to loas a single page.
