Latency is the delay between a request and the response to it. In digital advertising the request is an ad call and the response is a bid, a creative, or a decision to serve nothing. The interval is measured in milliseconds and it is always bounded: every system in the chain enforces a deadline after which a late answer is discarded. That deadline has money attached. A bid arriving after the auction has closed is worth nothing to the buyer who submitted it, and a page that renders slowly loses the reader before an impression can be recorded.

Where the milliseconds go

An ad call accumulates delay at four points. The first is network propagation. Signals move at a finite speed through fiber, and every router, amplifier and peering handoff adds time. Distance sets a floor no engineering removes. Low earth orbit satellite constellations operate between roughly 340 and 1,200 kilometers and deliver round trips of 20 to 40 milliseconds; geostationary links measure in hundreds of milliseconds, outside the budget most exchanges allow.

The second is connection setup. Google's best practices guide for real-time bidding states that a new connection costs an extra round trip, and that because connections open on demand, the first request on one has a shorter effective deadline and is likelier to time out than those that follow. The recommended remedy is reuse: an idle timeout of 2.5 minutes, in Apache a KeepAliveTimeout of 150 with MaxKeepAliveRequests at zero.

The third is processing. The bidder parses the request, applies targeting and frequency rules, scores the opportunity, selects a creative and serializes a response. Under load this is where latency turns unstable. The same guide describes a bidder that queues everything until latencies climb and requests time out in a steep saw-tooth pattern, against bidders that shed load by returning errors or no-bids. Google throttles callouts on an error rate above 15%.

The fourth is rendering. A chosen creative still has to be fetched, decoded and painted, and in video it may sit behind a chain of redirects before a frame plays.

The fields that carry the deadline

In OpenRTB, the IAB Tech Lab specification that standardizes programmatic bidding, the field is tmax, defined as the maximum time in milliseconds the exchange allows for bids to be received including internet latency. Version 3.0 adds an instruction for intermediaries: an exchange between a publisher and a bidder should decrease the outbound tmax from the value it received, to account for its own processing and the extra hop. Support is not universal. Microsoft's Xandr documentation states that tmax is not supported on incoming requests from supply partners, the ceiling being set per member seat and per data center instead.

Google's Authorized Buyers documentation gives the buyer-facing number: the response deadline runs from 80 to 1,000 milliseconds depending on format and auction type, with the figure for each request carried in BidRequest.tmax. The testing guide describes the same range as measured from the moment Google sends the call to the moment it receives a reply, and response latency percentiles are exposed to buyers through a dedicated graph.

On the publisher side, Prebid.js splits the budget into four settings. The auction timeout governs how long the header bidding wrapper waits for bidders. The failsafe timeout sits outside Prebid entirely, a JavaScript setTimeout() that calls the ad server if the library fails to load. The s2sConfig timeout applies to server-side bidders and, per Prebid documentation, should sit within 50% to 75% of the auction timeout, defaulting to 75%; that value is what Prebid Server writes into tmax. A timeout adjustment then has Prebid Server shave a safety buffer and answer before the original tmax expires. The project's frequently asked questions recommend 1,000 milliseconds or less for the auction timeout and 3,000 milliseconds or less for the failsafe, while bidderTimeout defaults to 3,000. One browser quirk is documented: because setTimeout only queues a callback approximately, bids can enter an auction late, though it closes immediately once the overshoot exceeds 200 milliseconds.

Google's Publisher Tag library defers requests rather than capping them. Its lazy loading configuration exposes fetchMarginPercent, the distance from the viewport at which a slot is fetched as a percentage of viewport height, renderMarginPercent, the distance at which it is rendered, and mobileScaling, a mobile multiplier.

Origin and evolution

Latency became a named constraint when auctions moved into the ad call itself. The waterfall that preceded real-time bidding was sequential by design, and its worst case was the sum of every timeout in the chain.

Header bidding replaced that sequence with a parallel auction, converting latency from an unbounded chain into a single configurable window. It created a new problem in exchange, because parallel calls multiply requests and each exchange forwards them onward. Publishers including Chegg and the Daily Mail deployed bid throttling from 2025 to cut the duplication, which the reporting linked to slower page loads and higher infrastructure costs. Will Doherty, vice president of inventory development at The Trade Desk, described the pattern in that coverage: "One issue we've encountered is that [SSPs] send 30% of the publisher's traffic - the stuff they think is best - and they send it three times."

Measurement of the user-facing half followed a separate track. Interaction to Next Paint, a responsiveness metric with a 200 millisecond threshold, entered Google's Web Vitals program in May 2022 and replaced First Input Delay as a Core Web Vital on March 12, 2024.

Infrastructure changed last. Amazon Web Services announced RTB Fabric on October 23, 2025, a managed private network for bidding traffic, on the stated basis that most real-time bidding auctions must complete within 200 to 300 milliseconds. Single-digit millisecond latency and up to 80% lower networking costs are the company's own figures. Custom domain support arrived on May 14, 2026.

Why the number is commercial

On the buy side, a response that misses the window is a bid that never competed. Vendors reporting early results from AWS RTB Fabric quantified it on their own traffic: Yieldmo reported more than an 80% reduction in timeouts alongside higher bid rates, and Azerion reported timeout and connection errors down by more than 50%. Both figures are self-reported. The mechanism is not disputed: fewer responses dropped after the deadline means more bids in more auctions.

On the sell side, delay suppresses the transaction the advertising is meant to produce. Shopify analyzed Core Web Vitals across its merchant base over a 28-day window spanning January and February 2026 and found conversion roughly 3.5% lower for every additional 100 milliseconds taken to load main content, and about 1.5% lower for every 32 milliseconds of added interaction delay. The company noted the second relationship is noisier.

Scale amplifies both. Sundar Pichai's remarks on search infrastructure, reported in April 2026, sit against a fixed threshold: humans perceive latency in the low hundreds of milliseconds, and budgets managed well below that ceiling still aggregate into what billions of users feel.

Limitations and disputes

The first dispute concerns who absorbs the cost. Client-side header bidding runs in the browser, offering transparency at the price of page load; server-side execution moves the work off the device but adds a hop and complicates identity synchronization. IAB Spain's SSP guide, published in April 2026, sets out both positions without resolving them.

The second concerns whether latency reduction is being sold as efficiency or as lock-in. Removing hops by placing a bidder inside exchange infrastructure also places a buyer's decisioning logic inside a seller's data center. Bedrock Platform ran the first containerized demand-side platform bidder inside Index Exchange on April 21, 2026, with an execution window Index Exchange reports at under five milliseconds.

The third concerns vendor arithmetic. IAB Tech Lab states that its Agentic Real-Time Framework cuts latency by 90%. Google's published 27% reduction in banner request latency for the Mobile Ads Next-Gen SDK rests on internal measurements taken between August 21 and September 3, 2025. Neither has independent verification.

Adjacent terms

Timeout is the deadline, not the delay. Latency is what a system takes; a timeout is what it is allowed.

Throughput counts requests per second. A bidder can be fast and small, or slow and enormous. Queries per second and milliseconds constrain each other but measure different things.

Data latency describes reporting freshness rather than transaction speed. Google's consent diagnostics in Analytics carry a 48 to 72 hour detection lag, and Google Ads documentation ties offline conversion upload latency to a seven-day gateon data-driven attribution. Both are measured in days.

Conversion lag is the interval between a click and a purchase, a property of consumer behavior rather than infrastructure.

Recent developments

The compression effort has moved from the auction into the ad break. A+E Global Media cut inbound ad server requests by 84% and raised impressions by 39% by switching from slot-level to pod-level auctions on an OpenRTB 2.6 integration with Index Exchange and FreeWheel, in a case study published in March 2026. Magnite's guidance for the 2026 soccer tournament, issued on March 9, 2026, addressed audience spikes, bid timeouts and failure handling in live streaming.

Server-guided ad insertion pulls the other way. IAB Australia's explainer, covered in August 2026, notes that resolving ads just in time improves video start time while concentrating ad calls into a narrow window at the break, which can stress ad servers and supply-side platforms without prefetch and pacing.

On mobile, Google named the Next-Gen SDK preferred for Android on July 6, 2026, the legacy kit sunsetting on June 30, 2028. The rewrite removes remote procedure calls, which Google says cuts latency and crash risk together.

Timeline

  • December 2010: OpenRTB is codified as an industry specification, carrying a maximum bid response time in the bid request
  • Early 2015: Prebid.js replaces sequential waterfall calls with a parallel header auction bounded by a single configurable timeout
  • 2020: OpenRTB 3.0 instructs intermediary exchanges to decrease the outbound tmax value to account for their own processing and the extra network hop
  • April 2022: OpenRTB 2.6 is released, retaining the same timing field
  • May 2022: Interaction to Next Paint enters Google's Web Vitals program as an experimental metric
  • March 12, 2024: Interaction to Next Paint replaces First Input Delay as a Core Web Vital, with a 200 millisecond threshold
  • August 21 to September 3, 2025: Google measures banner ad request latency for the Google Mobile Ads Next-Gen SDK against the previous SDK
  • October 23, 2025: AWS announces RTB Fabric, citing a 200 to 300 millisecond completion window for most real-time bidding auctions
  • January 22, 2026: The Google Mobile Ads Next-Gen SDK enters beta for Android with a published 27% reduction in banner request latency
  • February 26, 2026: IAB Tech Lab names its agentic initiative AAMP, with the Agentic Real-Time Framework beneath it
  • March 2026: Index Exchange and FreeWheel publish the A+E Global Media pod-level bidding case study
  • April 15, 2026: Amazon donates its Dynamic Traffic Engine to IAB Tech Lab as an open-source project addressing queries-per-second waste
  • April 21, 2026: Bedrock Platform runs a demand-side platform bidder inside Index Exchange infrastructure
  • April 27, 2026: Shopify publishes conversion data linking load time and interaction delay to purchase rates
  • May 14, 2026: AWS RTB Fabric adds custom domain support
  • July 6, 2026: Google names the Next-Gen SDK preferred for Android, setting a June 30, 2028 sunset for the legacy kit

Summary

Who. Demand-side platforms and bidders answer within the deadline; exchanges and supply-side platforms set it; publishers configure the wrapper timeouts that bound the page-side auction. IAB Tech Lab defines the field that carries the budget in OpenRTB, Prebid.org documents the client-side settings, and Google, Amazon Web Services and the major exchanges publish the platform-specific limits.

What. Latency is the elapsed time between a request and its response, measured in milliseconds and enforced by a deadline after which a late response is discarded. It is expressed in OpenRTB as tmax, in Prebid as a set of four related timeouts, and in Core Web Vitals as user-facing metrics for loading and responsiveness.

When. The constraint became explicit when auctions moved inside the ad call, from OpenRTB's codification in December 2010 and header bidding's arrival in 2015 onward. Infrastructure built specifically to compress it dates from October 2025, with containerized bidding following in April 2026.

Where. In the browser, in mobile SDKs, in streaming players and ad pods, across the public internet between exchanges and bidders, and increasingly inside private networks and exchange data centers.

Why. Late responses do not compete, and slow pages lose readers before impressions are recorded. Both effects have been quantified: timeouts falling by more than 80% on a dedicated bidding network, and conversion falling roughly 3.5% for each additional 100 milliseconds of load time.