How Does AI Optimise Programmatic Advertising? A Practical Guide

How does AI optimise programmatic advertising?

AI optimises programmatic advertising mainly by making better predictions, impression by impression, than a person could. For each bid request, the demand-side platform estimates how likely that impression is to produce the result you want and whether its price is good value against the market. Then it bids more, less or not at all.

That one idea runs through most of what gets called "AI" in a DSP. The same prediction engine also looks for people who behave like your best customers, spreads the budget across the day and the flight, and picks which creative to show. What it does not do is set your strategy. It optimises towards the goal and the limits you give it, nothing more.

If you need a refresher on the auction itself, read how programmatic advertising works first.

Where does AI work in a programmatic campaign?

AI shows up at several points, and you keep control of different things at each one.

How does automated bidding decide what to pay?

In DV360, automated bidding scores each impression on two things: the chance it leads to the outcome you chose, such as a click, a conversion or a viewable impression, and whether it is available at a good price compared with the market. Google says the predictions draw on signals such as the inventory, your audience lists and how your creatives perform. You can apply automated bidding at the insertion order or the line item level.

The strategies sit in two groups, under labels that use US spelling in the interface.

• "Maximize performance": for example, maximise conversions or clicks while spending the full budget, or while aiming for a target cost per result. App campaigns can maximise installs.
• "Maximize brand": for example, maximise viewable impressions, completed in-view and audible video impressions, or impressions viewable for at least 10 seconds. Connected TV line items can use an optimised reach option.

Optional settings let you cap the average CPM and prioritise deals over open auction inventory. Google's help page on automated bid strategies in DV360 lists the full set.

Two practical points matter more than the menu. First, Google says the algorithm can take up to four weeks to learn and calibrate before you judge it, so do not pull a line item after five days. Second, every big change to the goal, budget or targeting gives the model something new to learn. For more on how auctions and bids interact, see what real-time bidding is.

What is custom bidding, and when is it worth using?

Custom bidding lets you define what a good impression means for your business, instead of using a standard goal. DV360 offers two kinds.

• Rule-based: you combine predefined signals in the interface and weight them, for example giving one conversion type more value than another.
• Script-based: you write a script that scores impressions. It allows much more complex logic but needs someone who can write and maintain code.

Google recommends the rule-based version whenever it can support your KPI. Custom bidding is not available on YouTube, Programmatic Guaranteed or Demand Gen inventory. The model also needs a minimum amount of data before it can be used, such as at least 500 positively scored impressions at advertiser level and 50 per line item, and Google says training can take one to three days once that is reached.

It is worth the effort when your conversions differ in value. A lender that cares far more about completed applications than calculator visits, or an e-commerce brand that values a ₹5,000 order above a ₹500 one, gets more from custom bidding than from counting every conversion the same.

How does AI find new audiences?

Through what DV360 calls optimised targeting. You choose audiences as a starting point, and the system uses live click and conversion data to find other people likely to convert, beyond your original lists.

The useful distinction is between hard limits and hints. Location, language, first-party audience exclusions, brand safety exclusions and placement targeting are respected as hard limits. Audience inclusions, such as affinity, in-market, custom or your own lists, act as signals that guide the model but do not cap who sees the ads.

Optimised targeting switches on automatically for new display, video and audio line items that use automated bidding and target affinity, in-market or custom audiences. It works with click, conversion and custom bidding goals, not with the viewable impression strategies, and it is not available for connected TV line items. If a line item must stay strictly inside a list, check this setting before launch. Our guide to targeting audiences in DV360 covers the options in detail.

What is enhanced automation in DV360?

Enhanced automation lets DV360's automated bidding learn from aggregated, anonymised click, conversion and impression signals from participating advertisers. It helps most when your own data is thin, for example when third-party cookies are not available.

A partner-level admin has to accept the terms first. After that it can be turned on for the whole partner, for chosen advertisers or for individual insertion orders. It does not apply to custom bidding or YouTube. Google suggests allowing two to four weeks before you judge the results, ideally as an A/B test between two identical insertion orders, one with it switched on and one without.

How do other DSPs use AI?

Every major DSP now builds AI into buying, and the details change quickly.

• The Trade Desk launched its Kokai platform in 2023, spreading AI across bidding, forecasting and measurement. Traders steer it with bid factors that ask for more or fewer impressions of a given type.
• Amazon said in September 2026 that it is bringing its DSP and its ads console together as Amazon Ads Agent, with a conversational assistant. Display, video and audio buying moves into a simplified experience called DVA+, while experienced buyers keep advanced settings for targeting, deals, frequency and bidding.
• Google said in March 2026 that Gemini models now help DV360's Marketplace curate media packages before campaigns go live.

Our comparison of programmatic advertising platforms looks at how these platforms differ beyond AI.

What does AI need from you to work well?

1. A goal that matches the business. Optimise to a qualified lead or a purchase if that is what you pay for, not to a click.
2. Accurate conversion tracking. Tags should fire once, on the right page, ideally with a value. See how to track conversions in DV360.
3. Enough volume. A model that sees five conversions a week has little to learn from. Fewer, larger line items learn faster than many small ones.
4. Stable settings. Let a strategy run through its learning period before judging it.
5. Good inputs. Consented first-party data, fresh creatives and clean inclusion and exclusion lists.

An illustrative example shows why point 3 matters. A lending app has ₹6 lakh a month and a target of ₹1,500 per completed application. That is about 400 applications a month (₹6,00,000 ÷ ₹1,500). Split across 12 line items, each one sees roughly 33 a month, or about eight a week, and the models learn slowly. Run the same budget through three line items and each sees around 130 a month, which gives the bidding far more to work with.

Where does AI fall short?

• It optimises the goal, not the business. Ask it for cheap clicks and it will find accidental clickers on low-quality apps.
• It cannot judge brand fit. Which publishers suit a premium jewellery brand, or which news topics a bank should avoid, is a human call.
• It does not negotiate. PG deals on OTT apps, festive inventory and sponsorships are agreed between people.
• It sees correlation, not cause. Retargeting models love people who were going to buy anyway. Proving real impact needs lift studies and holdout tests, which we cover in measuring programmatic advertising ROI.
• It needs time and data. A short, small campaign may end before the model has learned anything useful.

So the trader's job has shifted. There is less manual bid-tweaking and more work on setting the right goals, feeding clean data, choosing supply and checking results against business outcomes.

What mistakes should you avoid with AI bidding?

1. Judging automated bidding in its first week.
2. Changing goals and budgets every few days, which keeps resetting what the model has learned.
3. Running conversion bidding on tracking that nobody has tested end to end.
4. Leaving optimised targeting on for a retargeting line item, then wondering why it reached new people.
5. Accepting every automated suggestion without checking it against the plan.

Next step

If you want automated bidding set up properly from day one, our DV360 Managed Service sets up tracking and bid strategies, checks pacing daily and optimises every week. Or start with a free media plan: share your goal, audience, budget and dates, and a DV360 specialist sends you a draft within one working day.