Algorithmic media buying is the practice of letting software decide what a single advertising impression is worth and how much to pay for it, instead of a human setting a fixed price in advance. A model reads the signals attached to an ad request - device, location, time of day, page context, audience membership, past behaviour - estimates the probability that this particular impression produces the outcome an advertiser has paid for, and turns that estimate into a bid. The whole sequence runs in the fraction of a second before a page or an app screen finishes loading. It exists because digital advertising produces more pricing decisions per hour than any trading team can make, and because the value of two impressions that look identical on a media plan can differ by an order of magnitude once the underlying signals are read.

The phrase covers search and social platforms, where the algorithm is a bidding strategy inside a closed auction, and the open programmatic market, where it is a bidder competing through a demand-side platform. The mechanics differ. The principle does not: the price is computed per impression, by a model, against a stated objective.

How a bid gets built

The transaction begins when a publisher or app sends an ad request. In the open market that request becomes an OpenRTB bid request, a structured object describing the placement, the device, the geography, the content and whatever identifiers are available. It reaches a demand-side platform, which has milliseconds to answer.

Inside the platform, a prediction layer scores the opportunity. Google defines its own version, Smart Bidding, as bidding strategies that use Google AI "to optimize for conversions or conversion value in every auction, a feature known as 'auction-time bidding'", according to the company's advertiser documentation. Four strategies fall under the label: Target CPA, Target ROAS, Maximize conversions and Maximize conversion value. The system reads more than 20 signals at the moment of the auction, among them device, operating system, physical location, location intent, day and time, remarketing list membership, browser, the search query itself and, for Shopping, product attributes.

The prediction becomes a price through the advertiser's objective. A target cost per acquisition converts predicted conversion rate into a maximum acceptable click or impression price. A target return on ad spend does the same with predicted revenue rather than predicted conversion count. Maximize strategies drop the target and spend the budget wherever expected volume is highest. Google product managers Kristina Park and Carlo Buchmann said in March 2026 that the models train across account-level conversion data rather than campaign-level data alone, which is why a new campaign can bid from its first impression.

On the open web the objective can be written by the buyer. Custom bidding in Display & Video 360 lets a trader score impressions with a script, weighting Floodlight conversion variables, Google Analytics goals, sales revenue, viewability or video completion rates, then trains the algorithm to buy toward the highest scoring inventory; it does not apply to YouTube, Gmail, audio or programmatic guaranteed inventory. The Trade Desk's equivalent, Koa, sits inside the Kokai interface as the layer adjusting bids and inventory prioritisation in real time.

Two constraints sit on top of the price. Pacing spreads a budget across the flight so the algorithm does not exhaust it in the first hours. Frequency and inventory rules bound where the model is allowed to look. Everything else is delegated.

Origin and evolution

The infrastructure arrived before the algorithms. Google's 2011 white paper on real-time bidding dates the exchange era to 2007, when Yahoo bought Right Media in April, Google bought DoubleClick in May and Microsoft bought AdECN in August, and notes that five demand-side platforms were founded that year. Exchange and supply-side support for real-time bidding spread through 2009 and 2010. The paper also describes the conceptual jump that defines the category: a buyer "goes beyond 'yes/no' decisions and develops a scale by which the impressions with the most value to the advertiser receive very high bids".

Standardisation followed. The IAB's OpenRTB specification reached version 2.0 in January 2012, 2.4 in March 2016 and 2.5 in December 2016, with version 3.0 published in September 2017 and the current 2.6 line beginning in April 2022, according to IAB Tech Lab's version list. Each revision widened what a model could read: video, audio, native, connected television, supply chain objects.

Auction design changed the arithmetic. Exchanges including AppNexus, Index Exchange, OpenX and PubMatic began testing first-price auctions in September 2017; Google announced a unified first-price auction for Ad Manager in March 2019 and completed the rollout on 10 September 2019. Winners now paid exactly what they bid, which forced buyers to add a second model on top of the first: bid shading, predicting the lowest price that still wins.

Platforms have kept narrowing the manual alternatives since. Microsoft Advertising folded Target CPA and Target ROAS into its maximise strategies as optional goals on 4 August 2025. Google moved the other way in June 2026, restoring Target CPA and Target ROAS as standalone labels in the interface, a change the company described as visual only.

Why it matters for marketers

Manual price setting is being removed as an option, not merely deprioritised. Microsoft will retire maximum CPC fields on new campaigns using automated strategies from 1 October 2026, citing simplification and pacing irregularities, while OpenAI made "Maximize results" the default in new ChatGPT Ads ad groups in mid-August 2026 with documentation conceding the strategy "does not guarantee delivery against a specific CPA, CPC, ROAS, or other cost-efficiency target at this time". Google adjusted campaign behaviour on 17 August 2026 so that campaigns beating their efficiency target drift up toward it rather than banking the difference. Three platform changes in a single week moved the same direction.

Scale explains the pull. Meta's Advantage+ reached a $20 billion annual run rate with 70% year-on-year growth in the fourth quarter of 2024, and advertisers using its creative enhancement features recorded 22% return-on-ad-spend improvements. Amazon, introducing natural-language targeting recommendations for DSP campaigns on 12 November 2025, reported internal tests cutting bid optimisation workflow time by 26%.

The consequence is a change in what the job is. Objective definition, conversion value accuracy and inventory boundaries determine outcomes, because those are the inputs the model reads.

Limitations and disputes

The dominant criticism is that the algorithm optimises toward what it can measure, which is not always what the advertiser is buying. The ANA's programmatic media supply chain transparency study, covering $123 million of spend and 35.5 billion impressions from 21 member companies between September 2022 and January 2023, found the average campaign running across 44,000 websites, 21% of impressions coming from made-for-advertising sites and roughly $20 billion of waste in an $88 billion market.

Control is the second dispute. Advertisers questioned whether Performance Max holds up above $100,000 a month in April 2026, describing budget drift toward remarketing and away from prospecting. Google Ads product liaison Ginny Marvin answered that the campaign type is "fundamentally built to prioritize the conversion over the channel". Channel reporting arrived in May 2025 and API access in January 2026, but channel-level budget allocation did not.

Failure modes are documented. Meta accounts saw 75% of daily budgets consumed within hours with tenfold CPM inflation in February 2024, and full budgets spent within hours that April. Marketing consultant Tom Goodwin has called the aggregate result "average-vertising", advertising that "kind of works and no-one is going to get fired".

The structural objection is that the party running the auction often also supplies the bidder and reports the outcome. Agency executives quoted by The Verge described the arrangement as platforms marking their own homework, a conflict no amount of model accuracy resolves.

What it is not

Programmatic describes the rails: the protocols, exchanges and deal types through which inventory is transacted. Algorithmic buying is the decision logic riding on those rails. A programmatic guaranteed deal at a fixed CPM is programmatic and not algorithmic.

Bid shading is a narrower algorithm concerned only with the clearing price, predicting the minimum needed to win rather than the value of winning.

Agentic buying delegates the negotiation and packaging of media to language-model agents, often across pre-negotiated deals rather than open auctions. DataBeat found agentic buyers clearing at a $6.13 CPM against $6.95 for conventional programmatic demand in May 2026, while participating in 86% fewer auctions.

Dynamic creative optimisation selects the message; algorithmic buying sets the price. The two increasingly interact: Display & Video 360 began matching creative selection to a line item's bidding goal on 12 August 2026.

Recent developments

Research has moved from per-scenario models toward general ones. Alibaba's Bid2X, presented at ACM SIGKDD in August 2025, replaces separate budget-constrained and target-ROAS bidders with a single foundation model; two months of A/B testing across a million Taobao campaigns produced 4.65% higher gross merchandise volume and 2.44% higher return on investment for 0.05 seconds of added inference latency.

Platform controls are being re-exposed selectively. The Trade Desk replaced Koa's on-off switch with three optimisation modes covering performance and delivery, efficiency and spend allocation, entering closed beta on 30 April 2026 with Koa left on by default. Chief executive Jeff Green used the 6 August 2026 earnings call to separate real-time "decisioning" from transaction execution, as rivals wrap guaranteed deals in agentic technology at far lower take rates.

Definitions are catching up to practice. IAB Tech Lab published final programmatic auction definitions on 26 June 2026, setting out 15 terms and a 12-step auction workflow across display, video, audio and connected television.

Timeline

  • April to August 2007: Yahoo acquires Right Media, Google acquires DoubleClick, Microsoft acquires AdECN; five demand-side platforms founded
  • 2009 to 2010: exchanges and supply-side platforms add real-time bidding support
  • January 2012: OpenRTB 2.0 published
  • December 2016: OpenRTB 2.5 published
  • September 2017: exchanges begin first-price auction testing; OpenRTB 3.0 published
  • March 2019: Google announces unified first-price auction for Ad Manager
  • 10 September 2019: Google completes first-price rollout
  • April 2022: OpenRTB 2.6 published
  • 4 August 2025: Microsoft folds Target CPA and Target ROAS into maximise strategies
  • August 2025: Alibaba presents the Bid2X auto-bidding foundation model at ACM SIGKDD
  • 12 November 2025: Amazon introduces natural-language targeting recommendations for DSP
  • 30 April 2026: The Trade Desk opens closed beta of three Koa optimisation modes
  • 26 June 2026: IAB Tech Lab publishes final programmatic auction definitions
  • 12 August 2026: Display & Video 360 ties creative selection to bidding goals
  • 17 August 2026: Google changes target behaviour; OpenAI defaults ChatGPT Ads to automated bidding
  • 1 October 2026: Microsoft retires maximum CPC on new automated campaigns

Summary

Who: Advertisers and agency traders on the buy side, operating through demand-side platforms such as Display & Video 360, The Trade Desk's Kokai, Amazon DSP and the bidding layers inside Google Ads, Microsoft Advertising, Meta and ChatGPT Ads. Publishers and supply-side platforms set the floors and auction rules the models bid into. IAB Tech Lab maintains the protocols and, since June 2026, the shared auction vocabulary.

What: The pricing and placement of individual ad impressions by predictive models rather than by fixed human-set prices, expressed through objectives such as target CPA, target ROAS and maximise-conversion strategies, and through custom scoring scripts on the open web.

When: The infrastructure emerged with the exchange acquisitions of 2007 and real-time bidding support in 2009 and 2010. Standardisation ran through the OpenRTB releases from 2012 onward. The shift to first-price auctions between 2017 and 2019 reshaped bid calculation. Manual price controls have been withdrawn steadily through 2025 and 2026, with Microsoft's maximum CPC retirement due on 1 October 2026.

Where: Across search, social, retail media, display, audio and connected television, in both open auctions and closed platform environments.

Why: The volume of impression-level pricing decisions exceeds human capacity, and signal-level differences in impression value are invisible on a media plan. The trade-off is contested: efficiency gains documented by platforms sit alongside the ANA's finding of roughly $20 billion of waste in the open web market and repeated practitioner complaints that the models optimise toward the cheapest measurable outcome rather than the intended one.