Build Your First AI Project This Weekend
Stop consuming tutorials. Start creating. Get the free step-by-step guide.
2026 Update:
The IAB Tech Lab finalized v1.0 of the Agentic Real Time Framework (ARTF) on August 12, 2026. It is a container/MCP layer on OpenRTB so bidding agents can sit next to the auction instead of calling home across the internet (IAB Tech Lab). Public comment started in November 2025. Netflix, Paramount, The Trade Desk, and Yahoo were in the early coalition.
You do not need to deploy a bidder-in-a-container. You do need to understand that “AI bidding” is becoming plumbing, not a slogan.
The other 2026 fact people still get wrong: Chrome did not kill third-party cookies. Google retired most Privacy Sandbox ad APIs after low adoption. Cookies remain, with user settings and consent laws still in the way (Privacy Sandbox status). First-party data still matters. “Cookieless future” as a 2024 talking point is stale.
This article is for people buying push, native, pops, or a DSP seat.
If you are a dentist or a plumber, skip RTB. Buy Google Ads and do local SEO. Programmatic display is how you buy leftover attention. Search is how you buy a broken AC.
What real-time bidding is
Every ad slot can be sold in an auction that finishes before the page paints. Roughly 40-120 milliseconds.
- User hits a page or app
- Publisher / SSP sends a bid request (URL, device, geo, sometimes user IDs)
- Exchange fans it out
- Advertisers bid
- Winner’s creative is returned
- You get charged on the auction’s terms (often second price or a close cousin, depending on the exchange)
You are not buying “the top of CNN” as a package. You are bidding on one impression, then the next.
Why it matters: you can start at $50/day on a self-serve network. You can also light $50/day on bot impressions and learn nothing. The auction does not care.
eCPM is the ranking, not your CPC
You might bid CPC. The server still ranks on estimated revenue per thousand impressions.
eCPM = 1000 x CTR x CPC
$0.15 CPC at 1% CTR = $1.50 eCPM. You beat a $1.00 CPM buyer.
Same $0.15 CPC at 0.5% CTR = $0.75 eCPM. You lose. Your bid did not change. Your creative did.
When CTR falls, you drop a position. Lower positions get worse CTR. Death spiral. Raising the bid on a rotting creative is paying more to lose slower.
Fix the creative. Then bid.
This is why I obsess over creative volume on RTB sources. Not because I like making banners. Because CTR is the lever the auction actually pulls.
DSPs, SSPs, exchanges (only as much as you need)
Advertiser -> DSP -> Exchange <- SSP <- Publisher
Auction
Winner serves
DSPs (you buy here at scale): The Trade Desk, Google DV360, Amazon DSP. These are $500/day to “call us” money, plus contracts. Not a first campaign.
SSPs (publishers sell here): Magnite, PubMatic, OpenX, Index Exchange. You will not log into these.
Self-serve networks (PropellerAds, Taboola, PopAds) sit on top of this mess and give you a simpler UI. Start there.
MediaMath: filed Chapter 11 in June 2023 and shut off. Infillion bought assets. It is not a “major DSP” you should plan a stack around. If an old blog post lists it next to TTD, the post is stale.
When to use a DSP: you are already spending serious money, you need PMPs or CTV, and you can staff it. Until then, one network done well beats three DSPs done badly.
The multi-armed bandit (why 10 creatives beat 1)
New campaigns have no eCPM yet. The ad server explores.
Think slot machines. Most traffic goes to the current winner (exploit). A slice goes to everyone else (explore). Each creative ID is a separate arm.
If Advertiser B uploads 9 creatives and C uploads 1, B gets more exploration shots. One of those IDs will print a lucky CTR. Lucky CTR -> higher eCPM -> it can steal the top slot even at a lower bid.
The server does not know your 9 banners are almost identical. It sees 9 IDs.
That is the actual AI edge for a solo buyer: generate 20-50 variations, let the bandit work, cut losers. You are not “gaming” anyone. You are feeding the algorithm the only thing it eats.
Bidding without donating
Your rep says “top CPC is $0.40.” That number might be one placement, one device, or a brand that does not care about CPA. Matching it with an untested campaign is how beginners die.
Do this instead:
- Start at or near the platform minimum
- See if you get traffic at all
- Raise in small steps only after you have CTR and conversion data
- Never “start high and optimize down” on a new account
Structure beats a big budget. $100/day as one “US mobile” campaign is a blender. Split:
- US iPhone 300x250
- US iPhone 320x50
- US Android 300x250
- US Android 320x50
Each with multiple creatives. Scale the cell that works. Pause the rest.
Clone, do not edit winners. Changing a live campaign changes your eCPM and can move you onto different inventory. Clone, change one thing, compare after 48-72 hours. If the clone wins, pause the original. If it loses, you still have the original printing.
AI prompts that match the auction
Creative volume:
20 ad variants for RTB.
Offer: [X]
Source: [push | pop | native | display]
Format: [size]
For each: headline, description, image idea, why it might win.
No fake urgency, no impersonating system alerts.
After you have data:
Analyze this export.
For creative, placement, device, geo:
1. Top eCPM with enough volume
2. Cut list
3. Needs more data
4. Bid moves
5. CTR decay / eCPM drop flags
[paste]
Run that every 48-72 hours on new campaigns, weekly on maintenance.
Do not ask AI for “expected ROI.” It will invent a number. You want cuts and clones, not a fantasy CPA.
KPI order
Watch in this order:
- eCPM - are you still winning impressions?
- CTR - that is what moves eCPM when bid is flat
- Conversion rate - is the traffic useful?
- CPA - the only number your accountant cares about
People invert this. They screenshot a “good CPA” on 11 clicks while eCPM is sliding and volume is about to vanish.
Spend tiers (so you do not buy a DSP for a $300 test)
Under ~$5k/month: self-serve networks. Push, native, pop. Learn eCPM there.
$5k-25k: maybe a DSP trial if you need brand-safe display or you already have converting audiences. TTD, DV360, Amazon DSP (Amazon first-party data is the point of that one).
Above that: PMPs, guaranteed, a human who lives in the seat.
CTV is real spend for brands. It is a stupid first test for a CPA offer with a $20 payout.
2026, without the press-release fog
ARTF / AAMP. IAB folded ARTF into a broader Agentic Advertising Management Protocols effort in 2026: buyer/seller agent SDKs, an Agent Registry, deals APIs. Production use is at big platforms. Your job is still: good creatives, clean targeting, conversion postbacks.
Cookies. Still in Chrome. Safari and Firefox already limited them. Consent popups still exist. Build email / customer lists anyway. Contextual targeting never left and is less fragile than a third-party ID.
CTV / DOOH / audio. Programmatic inventory expanded. Connected TV is a real budget line for brands. It is not where you test a $50 CPA offer. Learn display/native/push first.
PMPs and programmatic guaranteed. Invite-only or fixed-price deals. Relevant when you need brand safety and you can spend. Irrelevant for a first pop campaign.
Tracking
Track creative, placement, device, OS, geo, conversions, revenue. Network dashboards lie by omission.
Trackers: Voluum, RedTrack, BeMob, Keitaro. Set up before spend. Tracking guide.
Display is upper funnel. Test last-click vs a longer click window. Do not “fix” a campaign by switching attribution until you understand both views.
Pass the click ID all the way through the offer. If the affiliate network cannot fire a postback with that ID, you cannot train the source’s bidder, and you cannot trust the report. This is the unsexy work that separates people who “tried programmatic” from people who can scale it.
View-through conversions on display will make the dashboard look like a hero. Count them. Do not pay CPA Goal as if they were last-click sales unless you have tested it against real revenue.
Mistakes
- Chasing the top bid
- One creative
- Editing a live winner
- Watching CPC while eCPM (and volume) dies
- One giant campaign
- Buying programmatic “awareness” for a local service company
- Asking the rep what bid “wins the auction” and treating it as a strategy
- Spreading $50/day across five DSPs
Kill signals: CTR down 3+ days, eCPM below break-even, no conversions after a statistically boring pile of clicks (think 100+ on a simple offer), same assets for two weeks with no refresh.
Do not kill at 200 impressions. Do not nurse a loser for a month.
Should you even be here?
Yes: you buy affiliate or performance traffic, you will look at eCPM, you can generate creatives weekly.
No: you need booked jobs in a service area this month. That is Search + Maps.
The mechanics above still help you read Google’s auction (Quality Score is a cousin of eCPM). Learn them. Apply them on the channel that matches the buyer.
If a vendor demo jumps straight to CTV, identity graphs, and “agentic buying,” ask them to show a simple display or native campaign with a postback. If they cannot, you are not the customer yet. You are a slide in their deck.
Next
Starting: pick push or pops. Tracker on. 20 creatives. Minimum bid. $50/day. 48 hours.
Already live: export, cut anything under ~50% of top eCPM, clone the top three.
Local business: Google Ads and local SEO.
Related: paid traffic map | native | campaign optimization
Have questions? Contact me.