AI Advertising · History

The History of Artificial Intelligence Advertising: From Rule-Based Bidding to Andromeda

Every step from manual bids to Meta's October 2026 Andromeda update, dated and sourced. It also shows which habits belong to an era that has already ended.

Ajay Kumar • •12 min read
Illustration of a stepping-stone path where manual tools give way to an AI funnel that sorts ads, showing how advertising moved from manual control to AI

Quick answer: how did AI take over advertising?

AI in advertising developed in four eras. Manual, rule-based bidding gave way to machine-learning bids (Google's Enhanced CPC, reported as 2010, and Smart Bidding, reported as 2016). Whole-campaign automation came next (Performance Max, Meta Lattice, Advantage+). The current era is retrieval and foundation models: Meta Andromeda (December 2024), GEM (2025), next-gen Andromeda (October 2026) and Google's AI Max.

Every step of ad AI on Google and Meta, dated

From 2010 to 2026, Google and Meta handed one manual lever after another to their AI. The table dates each step and shows where the date comes from. 'Official' means Meta or Google published it. 'Reported' means a trade source citing the platform.

The scope is narrow on purpose: how ads get bid on, picked and ranked. For the basics of how the models work, start with our plain-English explainer on how AI in advertising works.

DatePlatformWhat changedSource
2010GoogleEnhanced cost per click (eCPC) adjusts manual bids by how likely a click is to convertReported
Jul 2016Google'Smart Bidding' groups Target CPA (cost per acquisition), Target ROAS (return on ad spend) and eCPCReported
May 2017GoogleMaximize conversions addedReported
2023MetaLattice, a unified ranking model architectureReported
Oct 2024GoogleNew Search and Display campaigns can no longer choose eCPCReported
2 Dec 2024MetaAndromeda ads retrieval engine announcedOfficial (Meta Engineering)
Mar 2025GoogleRemaining eCPC campaigns moved to Manual CPCOfficial (Google Help)
May 2025GoogleSmart Bidding Exploration for Search. AI Max announced on 6 MayReported. AI Max is documented in Google Help
2025MetaSales, App and Leads campaigns default to Advantage+Reported
Nov 2025MetaGEM (Generative Ads Recommendation Model) detailedReported
Early 2026MetaManual and Advantage+ setup merged into one creation flowReported
30 Apr 2026GoogleAI Max expands to Shopping and adds an AI Brief toolOfficial (Google blog)
7 May 2026GoogleExploration extended to Shopping and Performance MaxReported
28 Oct 2026MetaNext-gen Andromeda: model, system and GPU designed togetherOfficial (Meta Engineering)

Why this matters at ₹10 lakh+ a month: a team on a pre-2024 playbook keeps pulling levers that no longer exist. At ₹33,000 or more a day, that waste adds up within a week.

Rules first, then machine-learning bids (up to 2020)

Until about 2010, most advertisers set bids by hand. Over the next seven years, Google moved bid-setting to machine learning, first as a nudge and then as the whole decision.

Era 1: rules you wrote yourself

Early marketing AI was rule-based. The Springer reference history places expert systems in the 1950s to 1980s, long before today's ad platforms.

In ad accounts, that meant Manual CPC (cost per click) for each keyword. You then added bid adjustments for device, location and time of day. Meta-focused sources describe early social ad delivery in similar terms: rule-based filtering plus small models.

Era 2: the machine nudges, then sets, the bid

Trade sources date Google's launch of Enhanced CPC to 2010. It kept your manual bid but raised or lowered it based on how likely a click was to convert.

According to trade sources such as PPC Land, the Smart Bidding name arrived in July 2016, grouping Target CPA, Target ROAS and eCPC, and Maximize conversions followed in May 2017. We could not confirm these dates on a Google page. Over the same years, programmatic real-time bidding spread across the open web.

The lesson from this era still holds: a bidding model is only as good as the conversions it sees. Tracking stopped being a reporting job and became a bidding input.

Whole-campaign automation took placements and audiences

Next, both platforms automated the whole campaign, not just the bid. You stopped choosing where ads ran and, more and more, who saw them.

On Google, Performance Max (PMax) launched to all advertisers in November 2021 as one goal-based campaign across Google's inventory, and Smart Shopping and Local campaigns were upgraded into it in 2022.

On Meta, Lattice arrived in 2023 as a unified ranking architecture, and in 2025 Meta began making Advantage+ the default for Sales, App and Leads campaigns, according to trade reports. We could not confirm either date from a Meta page. By early 2026, manual and Advantage+ setup sat in one creation flow, according to Jon Loomer's changelog.

Levers lost: placements and stacked interest audiences. To see what each platform now does without a paid tool, read what Advantage+ and Smart Bidding already do for free.

Andromeda, Lattice and GEM: Meta's retrieval era (2024-2026)

Since December 2024, Meta's AI has decided which of your ads even reach the ranking stage for each person. That shift made creative variety the main lever left on Meta.

Three models, three jobs

The October 2026 update

On 28 October 2026, Meta Engineering described next-generation Andromeda. Meta designed the model, the system and the GPU (graphics processor) hardware together. Trade coverage describes the result as bringing retrieval close to ranking-level performance.

The whole retrieval pipeline now runs on a single GPU. Trade coverage adds that a mixture-of-experts gating design brings this extra detail at what it calls an 'essentially free' serving cost. We have not confirmed that wording in Meta's post. Meta says this builds on Andromeda's 2024 launch on NVIDIA GH200 Grace Hopper hardware.

Our reading for advertisers

If retrieval now judges ads almost as finely as ranking does, ten ads that say the same thing give it little new to choose between. Picking one 'hero' ad is no longer your job. Supplying genuinely different concepts is. This is our interpretation, not a rule Meta has published.

Google's parallel path: eCPC retired, Exploration, AI Max

Between 2024 and 2026, Google removed its last hybrid bidding option and moved keyword control into AI. The pieces are easy to mix up, so here they are in order.

Enhanced CPC is gone

From October 2024, new Search and Display campaigns could not choose eCPC. Remaining eCPC campaigns were moved to Manual CPC from March 2025. Google Ads Help confirms that campaigns not switched to another strategy now effectively run on Manual CPC.

Exploration widened Smart Bidding

According to PPC Land, Smart Bidding Exploration launched for Search in May 2025, and Google announced its expansion to Shopping and Performance Max on 7 May 2026. We could not confirm these dates on a Google page.

AI Max took over from DSA

AI Max was announced in May 2025 (trade reports give 6 May) and moved out of beta in April 2026, when Google said Dynamic Search Ads would be upgraded to it automatically. It is not a new campaign type. Google Ads Help describes it as a layer inside Search campaigns. It evolves Dynamic Search Ads (DSA) and PMax Final URL expansion, and it keeps the controls of a Search campaign.

When you upgrade, existing DSAs become Responsive Search Ads. If a search qualifies for both AI Max and PMax, the one with the higher Ad Rank serves.

Google says advertisers who turn on AI Max typically see 14% more conversions or conversion value at a similar CPA or ROAS. We understand it is based on Google's internal data for non-retail advertisers in 2025, which we have not re-checked, so treat it as a claim to test, not a promise.

A year after launch, Google extended AI Max to Shopping campaigns and travel ad formats. It also added AI Brief, which rolls out in English first.

What most guides get wrong about Andromeda, GEM and AI Max

Most errors come from pages written before late 2026, or from practical guesses presented as platform rules. These are the ones we see most often.

Common claimWhat is actually documented
Andromeda launched in late 2024, and that's the whole storyAnnounced on 2 Dec 2024. Meta published a next-gen update on 28 Oct 2026.
Andromeda replaced rankingAndromeda is retrieval, the stage before ranking. Lattice and other models still score the candidates.
GEM means Generative Engagement ModelMeta calls it the Generative Ads Recommendation Model.
GEM generates your ad creativeGEM predicts and recommends. Creative generation comes from Advantage+ creative tools.
Lattice was a 2025 updateTrade reports date Lattice to 2023, with later expansions. We have not seen that date on a Meta page.
A 60% similarity score suppresses your adsThere is no Meta source for this. Treat it as unverified.
Near-identical ads are merged into one retrieval entryMeta has not published any merge rule. It's a working rule of thumb.
eCPC is still a hybrid optionBlocked for new campaigns from Oct 2024, and moved to Manual CPC from Mar 2025.
DSA is Google's keywordless optionAI Max is the evolution of DSA. In April 2026 Google said DSA campaigns will be upgraded to AI Max automatically, and Google Ads Help now dates that upgrade from February 2027.

The advice behind the similarity myths is still sound: make ads that differ in concept. Just don't build a process around a percentage that Meta has never published.

What the history means in your market

The platform history is global, but billing, tax and payment signals are local. As the models train more on your conversions, these local details shape what the AI learns and what your ROAS really is.

Tax sits on top of spend

In India, Google adds 18% GST (Goods and Services Tax) on top of the spend shown in Google Ads, and tax guides report that Meta does the same in Ads Manager. You can usually claim input tax credit (ITC) if the invoice carries your GSTIN. Tax guides also say that if a foreign entity bills you, the invoice shows no GST and you self-assess 18% IGST under reverse charge, so confirm this with your tax adviser. Google sets out its rules in Google Ads Help on GST in India.

In the US there is no VAT, though some states levy digital advertising or gross-receipts taxes. In the UK, EU, Gulf, Australia and South-East Asia, VAT or GST treatment depends on your billing entity. Check the actual invoice instead of assuming.

Cash-on-delivery signals in India and the Gulf

A cash on delivery (COD) order fires a purchase event at checkout, before the customer has paid. If the parcel is refused or returned to origin (RTO), the model has still learned from a sale that never happened. The fix is at the event level: use the Meta Pixel and Conversions API (CAPI) setup to send confirmed or delivered orders where you can.

How tax and RTO change the numbers

Here is how the gap between reported and real ROAS looks once tax and failed deliveries are counted.

Worked example: ₹30 lakh a month on Meta

Step 1: Start with what Ads Manager shows

Example: a brand spends ₹30,00,000 a month on Meta. Ads Manager reports ₹1,05,00,000 in purchase value, a ROAS of 3.5.

Step 2: Remove tax and failed COD orders

Assume ₹15,00,000 of that value is GST collected from customers, which leaves ₹90,00,000. Now assume 20% of order value is cancelled or comes back as RTO. That leaves ₹72,00,000 actually delivered.

Step 3: Recalculate on delivered, net revenue
Delivered net ROAS
₹72,00,000 ÷ ₹30,00,000 = 2.4 with ITC claimed. Without ITC: ₹72,00,000 ÷ ₹35,40,000 = about 2.0

Compare this with your break-even ROAS after product cost, shipping, tax and returns. If break-even is 2.5, the '3.5 ROAS' campaign is losing money.

Step 4: Fix what the model learns

The model optimised toward the 3.5 figure, including the 20% of orders that came back. Sending confirmed orders through CAPI teaches it to find buyers who actually accept delivery.

This is the calculation ScaleWin makes before it judges any ad: profit after product cost, shipping, tax and RTO, not platform ROAS.

The pattern: what to stop, keep and check, in order

Every era took away an output you used to control and left you with inputs. Bids, placements, audiences, keywords and creative selection went to the AI. Conversion signals, creative concepts, feeds and landing pages stayed with you.

Our order of checks when performance shifts

When results move after a platform change, work through these checks in order. Each threshold is our rule of thumb, set on the conservative side.

  1. Signal. Check that the Pixel and CAPI both fire and are deduplicated. If more than 1 in 5 COD orders fail, optimise on a confirmed-order event.
  2. Creative concepts. If your main ad set has fewer than 3 genuinely different concepts (hook, format, angle), fix creative before you touch structure.
  3. Landing page. If landing-page views are below 70% of link clicks, fix page speed and the page itself before changing ads.
  4. Structure. Consolidate ad sets that split the same audience and budget.
  5. Bids last. On Google, stay on Manual CPC or Maximize conversions until a campaign records about 30 conversions in 30 days. Only then test Target CPA or Target ROAS.

Stop and keep

Don't confuse seasons with updates

Festive peaks such as Diwali, Ramadan and BFCM (Black Friday to Cyber Monday) push up CPMs, the cost per 1,000 impressions. Check the calendar before you blame a platform change for a jump in costs.

For what is likely to come next, see what has shipped and what is still a bet for 2026-2030.

Free tools for this topic

Key takeaways

  • Each era took one manual lever away: bids, then placements and audiences, then keywords, then creative selection.
  • Enhanced CPC is gone. New campaigns lost it in October 2024, and remaining ones moved to Manual CPC from March 2025.
  • On Meta, Andromeda picks the candidate ads, Lattice ranks them and GEM teaches both. GEM does not make your ads.
  • What you still control is inputs: clean conversion signals, distinct creative concepts, product feeds and landing pages.
  • Judge results on net, delivered revenue. Tax and failed cash-on-delivery orders distort both your ROAS and what the models learn.

Artificial intelligence advertising: frequently asked questions

When did AI start being used in advertising?

Rule-based marketing systems go back decades. In mainstream ad buying, Google's Enhanced CPC (reported as 2010) and Smart Bidding (reported as 2016) were among the first widely used machine-learning bid tools. Programmatic real-time bidding grew on the open web over the same period.

What is Meta Andromeda and when did it launch?

Andromeda is Meta's ads retrieval engine. It narrows the pool of eligible ads to a small set of candidates before ranking. It was announced on 2 December 2024, and Meta Engineering described a next-generation version on 28 October 2026.

What is the difference between Andromeda, Lattice and GEM?

Andromeda retrieves candidate ads, and Lattice ranks them. GEM, the Generative Ads Recommendation Model, is a large teacher model that improves both through knowledge distillation. They are three stages of one system, not competing products.

Does GEM create my ads?

No. GEM is a prediction and recommendation model. AI-generated creative variations come from Meta's Advantage+ creative tools, not from GEM.

Is Enhanced CPC still available in Google Ads?

No. New Search and Display campaigns could not choose eCPC from October 2024. Remaining eCPC campaigns were moved to Manual CPC from March 2025, and Google Help says any campaign that wasn't switched now effectively runs on Manual CPC.

Is AI Max the same as Performance Max?

No. AI Max is an optimisation layer inside your existing Search campaigns and evolves Dynamic Search Ads. Performance Max is a separate campaign type. If a search qualifies for both, the one with the higher Ad Rank serves.

Do similar ads really count as one ad under Andromeda?

Meta has not published any rule that merges similar ads into one record, and claims about a '60% similarity score' have no Meta source. The practical takeaway still holds: ads that differ in concept give the system more real choices than small text tweaks do.

Does GST apply to Meta and Google ads in India, and can I claim it back?

Yes. Google Ads Help confirms 18% GST for Indian advertisers, and Meta is reported to add 18% GST on top of the spend shown in Ads Manager. GST-registered businesses can usually claim input tax credit if the invoice carries their GSTIN. Tax guides say that if a foreign entity bills you, you self-assess 18% IGST under reverse charge, so check this with your tax adviser.

Sources

Ajay Kumar
Founder, ScaleWin · 12+ years in performance marketing

Ajay is the founder of ScaleWin. He works hands-on with Meta Ads, Google Ads and SEO for Indian D2C and service brands, and built ScaleWin so that the daily work of a performance marketer - reading numbers, pausing losers, scaling winners - happens automatically.

Drafted with AI assistance. How these articles are written

Let ScaleWin read your ad data every day

ScaleWin checks your Meta and Google Ads daily, judges every ad on profit and tells you exactly what to change.

Start free, no card

Get marketing tips that grow your business

One practical email a week on ads, SEO and growth. No spam, no fluff. Unsubscribe anytime.