All resources

· 14 min read

How to Build an LLM Advertising Stack: Tools, Workflow, and Budget (2026)

A practical build guide for advertising across AI assistants. Covers the four layers of an LLM ad stack, the right tools at each budget level, a reference workflow, and how Lapis serves as the creative and campaign hub.

Why You Need an LLM Advertising Stack

For twenty years, “paid ads” meant search engines and social feeds. That assumption is breaking. In 2026, hundreds of millions of people begin their buying research inside an AI assistant instead of a search bar, and the assistants have started to sell ad placements against those conversations. ChatGPT alone crossed 800 million weekly active users (OpenAI, 2026), and in April 2026 OpenAI opened self-serve advertising through ads.openai.com. A channel that did not exist eighteen months ago is now a place you can buy.

This is the same structural shift that created search advertising and social advertising, and it rewards teams that build for it early. It also introduces a new problem. Advertising is no longer one destination; it is a growing set of answer engines, each with its own inventory, targeting model, and creative requirements, sitting alongside the channels you already run. If you bolt a fresh point tool onto your workflow for every one of them, you recreate the exact fragmentation that has plagued marketing operations for a decade.

88%

of marketers have already begun optimizing for AI-generated responses in places like ChatGPT and Google AI Overviews

Source: Salesforce State of Marketing, 2026

The average marketer already stitches together roughly 10 separate data sources just to see a single customer clearly, and only 31% are satisfied with their ability to unify them (Salesforce State of Marketing, 2026). Layer answer-engine advertising on top of that and a typical operation ends up juggling 6 to 8 disconnected tools: one for static design, one for copy, one for video, a separate manager for each ad platform, an analytics tool, and a competitor tracker. Every handoff loses context, and none of those tools were built for LLM-native placements.

10 data sources

stitched together on average by marketers to build a single view of the customer; only 31% are satisfied with their ability to unify them

Source: Salesforce State of Marketing, 2026

The opportunity is that most competitors still treat AI assistants as an organic-visibility problem, not a paid channel. But spend follows attention. The AI-in-advertising market is projected to reach $36.34 billion by 2030, growing at a 26.7% compound annual rate (Research and Markets, 2026). Getting a repeatable stack in place now, while inventory is cheap and competition is thin, is the same bet early Facebook and Google advertisers made.

$36.34B by 2030

projected size of the AI-in-advertising market, growing at a 26.7% CAGR

Source: Research and Markets, AI in Advertising Market Report, 2026

This guide is specifically about the LLM-native, cross-assistant stack. If you want the broader picture of a modern paid-ads operation across Meta, Google, and TikTok, start with our AI-powered paid ads stack guide. For a channel-by-channel look at where ads are appearing inside each assistant, see advertising in AI assistants. Here we focus on how to assemble the tools, workflow, and budget to advertise across all of them at once.

The Four Layers of an LLM Ad Stack

Every LLM advertising operation, regardless of budget, is built from the same four layers. Think of them as a build spec: four capabilities you must cover, whether you cover each with a platform, a tool, or a person. We break down the underlying mechanics in LLM ad infrastructure explained; here is the version you build against.

LayerWhat it doesThe question it answersWhere it comes from today
Demand & auctionServes the ad and sets the priceWhich assistants and surfaces can I actually buy?Ad networks: OpenAI (ads.openai.com), plus emerging Perplexity, Copilot, and Gemini, alongside Meta, Google, Reddit, and LinkedIn
Context & targetingMatches the ad to the right conversationHow do I reach the right intent?Platform context signals (context hints, not keywords) and first-party audiences
Creative generationProduces the ad itself, per surfaceHow do I make on-brand ads for every format, fast?Lapis (the creative and campaign hub)
Measurement & attributionConnects spend to outcomesDid it work, and which creative drove it?Platform pixels and UTMs, Lapis Web Analytics, and GA4

The layers are not equal. The demand and context layers are largely provided for you: the assistants and ad networks own the auction and the targeting signals, and they improve on their own. The measurement layer is hard but solvable with discipline. The creative layer is the one you actually have to build, and it is where most of your time, money, and competitive advantage will concentrate. That is why the rest of this guide keeps returning to it.

Layer by Layer: The Right Tools at Each Level

Here is the practical mapping of each layer to the tools that cover it in 2026. Note that the demand and context layers are mostly bought, not built: you plug into the assistants and ad platforms and inherit their auctions and targeting. The creative layer is the one you own outright, and it is the bottleneck.

LayerWhat to useNotes
Demand & auctionads.openai.com for ChatGPT; Meta Advantage+, Google AI Max, Reddit, LinkedInBuy where your audience already asks questions. ChatGPT self-serve starts around $5,000 per month.
Context & targetingPlatform-native context matching plus your first-party audiencesLLM ads target intent through context hints, not keyword lists. Feed clean product and audience data.
Creative generationLapis (one prompt, every format)The bottleneck. Volume, on-brand consistency, and pre-launch forecasting decide the whole stack.
Measurement & attributionPlatform pixels and conversion APIs, UTMs, Lapis Web Analytics, GA4The conversation gap makes last-click unreliable. Use creative-level tags and holdout tests.

Notice the pattern. For demand, context, and measurement, you are mostly configuring services that already exist. For creative, you are responsible for producing a constant stream of on-brand assets in every format each surface demands: conversational sponsored answers for ChatGPT, feed and story sizes for Meta, responsive assets for Google AI Max, and native formats for Reddit and LinkedIn. Do that manually and the creative layer becomes the rate limiter for the entire operation. Solve it, and everything downstream gets easier.

The Creative Hub: Where the Stack Lives or Dies

Creative is the highest-leverage layer for a simple reason: it decides the largest share of performance. Meta’s own analysis attributes roughly 56% of ad performance variation to creative quality, more than targeting, bidding, and placement combined (Meta for Business, 2025). Across LLM surfaces the effect is, if anything, stronger, because an assistant renders your ad inside a conversation where relevance and tone matter more than in a scrolling feed.

56%

of ad performance variation is driven by creative quality, more than targeting, bidding, and placement combined

Source: Meta for Business, 2025

The problem is throughput. The old creative process (brief, design, review, resize per platform) produces three to five variants over a week or more. An LLM ad stack needs the opposite: many on-brand variants, refreshed continuously, formatted correctly for every assistant and channel. Manual production cannot keep up, and generic AI image tools produce off-brand one-offs that a human still has to assemble, resize, and re-copy for each surface. This is why so many teams stall: their demand and measurement layers are ready, but their creative layer cannot feed them.

Lapis is built to be the creative and campaign hub that removes this bottleneck. It starts with Brand Intelligence: paste your website URL and Lapis extracts your logo, colors, typography, and voice automatically, so every asset is on-brand by default. From a single natural-language prompt, it generates production-ready ads for ChatGPT plus Meta, Google, Reddit, and LinkedIn in under three minutes, correctly sized for each placement. There is no manual resizing and no reformatting between tools.

Three capabilities turn that generation engine into a hub rather than another point tool. Performance Forecasting scores each variant on predicted click-through, cost, and return before you spend a dollar, so you launch only the strongest concepts. Campaign Studio lets you refine any asset with plain-language instructions (“make the headline punchier, swap in the blue product shot”) instead of a design tool. Catalog Import pulls products straight from Shopify, Amazon, or a CSV, and multilingual generation produces native copy in more than a dozen languages. Competitor Tracking and Web Analytics round it out, and we return to both in the data-flow section.

8 hrs / week

reclaimed by marketing teams that use AI to automate content variation and analysis, time reinvested into strategy and creative

Source: Salesforce State of Marketing, 2026

A Reference Workflow, End to End

Here is a reference workflow that ties the four layers into a repeatable loop. It works whether you run one assistant or five, and a single marketer can execute it.

  1. Brief. Write one plain-language brief describing the offer, audience, and angle. Store it where your creative hub can reuse brand context, so you are not re-explaining who you are every time.
  2. Generate multi-format creative. In Lapis, turn the brief into a full set of on-brand variants for every surface you plan to run (ChatGPT sponsored answers plus Meta, Google, Reddit, and LinkedIn) in one pass.
  3. Forecast. Score the variants and keep the top performers. Kill weak concepts before they cost anything.
  4. Launch across surfaces. Push the ChatGPT set through ads.openai.com (self-serve, roughly a $5,000 per month minimum), and the rest through Meta Advantage+, Google AI Max, Reddit, and LinkedIn. For a full ChatGPT setup walkthrough, see our ChatGPT ads complete guide.
  5. Track. Tag every variant with consistent UTMs, fire platform pixels and conversion APIs, and connect on-site behavior through Lapis Web Analytics and GA4 so you can attribute outcomes to specific creatives.
  6. Iterate weekly. Feed what worked back into the next brief. The loop compounds: each cycle produces better creative because it is built on real performance data.

The launch step looks different on each assistant. Here is the state of play in 2026:

AssistantAds status (2026)How buying worksTypical economics
ChatGPTLive: self-serve via ads.openai.com (April 2026)Context-matched sponsored answers; about $5,000 per month minimumCPM $25 to $60, CPC $3 to $5; roughly 1.5 to 4× higher conversion rate than Google Search
PerplexityEmerging: sponsored follow-up questionsManaged and beta placementsPremium CPMs against high-intent research traffic
Microsoft CopilotEmerging: ads within Copilot answersVia Microsoft AdvertisingTied to Bing and Edge inventory
Google Gemini / AI OverviewsRolling out: ads in and around AI answersVia Google Ads and AI MaxBlends into the existing Search auction

Two things make LLM placements different from search. First, targeting runs on context hints, not keywords: you describe the conversations and intents you want to appear in, and the assistant matches your ad to relevant exchanges. Second, attribution is harder because of the conversation gap. A user may ask several follow-up questions, leave the chat, and convert later, so last-click models undercount the channel. Plan for it by leaning on creative-level UTMs, platform conversion APIs, and holdout tests rather than a single last-touch report.

Budget Tiers: Starter, Growth, and Scale

Your stack should scale with spend, not the other way around. The table below maps three tiers (Starter, Growth, and Scale, plus Enterprise) to recommended tools, approximate monthly tool cost, and the matching Lapis plan. Tool cost is deliberately small relative to media: your dollars belong in ad spend and creative, not in a pile of subscriptions.

TierBest forCreative hub (Lapis)Demand layerMeasurementApprox tool cost / mo
StarterGetting started, first testsLapis Free ($0, 5 credits)Meta and Google nativeGA4 + platform pixels$0
GrowthEarly startupsLapis Basic ($99)Meta, Google, Reddit; ChatGPT when readyLapis Web Analytics + GA4~$99
Scale (recommended)Growth teams running cross-assistantLapis Pro ($599): forecasting + Campaign StudioChatGPT (~$5K+/mo) + Meta, Google, LinkedInLapis + GA4 + attribution tool~$599
EnterpriseMulti-brand, high spendLapis Enterprise (custom)Full cross-assistant + programmaticLapis + media mix modeling / NorthbeamCustom

A few notes. At the Starter tier, Lapis Free (5 credits, no credit card) plus native Meta and Google campaigns is enough to learn the loop at zero tool cost. Growth teams move to Lapis Basic at $99 for unlimited generation, competitor tracking, and web analytics, and add ChatGPT once they can commit to its minimum. The Scale tier is where most serious teams land: Lapis Pro at $599 unlocks Performance Forecasting and Campaign Studio, exactly the capabilities that make cross-assistant volume manageable. Enterprise adds custom seats, multi-brand workspaces, and dedicated support. Whatever the tier, the creative hub stays constant; you add media and complementary tools around it rather than replacing it.

The Data Flow: Performance Into Better Creative

The difference between a stack that plateaus and one that compounds is whether performance data flows back into creative decisions. In a fragmented setup, that loop is broken: your analytics live in one tool, your competitor research in another, and your creative in a third, so insights never reach the prompt. In a hub-centered stack, they do.

Two inputs make the next round of creative smarter. Competitor Tracking watches what rivals run across surfaces and surfaces their messaging and format patterns, so you can generate counter-positioning without leaving the platform. Web Analytics ties each visit back to the specific creative, headline, and CTA that drove it. When both feed the same creative engine, your prompts stop being guesses. Instead of “make a summer sale ad,” you brief “make three variants like our top performer, using the social-proof angle that beat the discount angle, and counter the competitor promo we just detected.” For a deeper treatment of closing this loop, see our AI ad optimization playbook.

83%

of high-performing marketing teams use every major type of AI (predictive, generative, and agentic), versus a minority of underperformers

Source: Salesforce State of Marketing, 2026

None of this works without disciplined tagging. Use consistent, lowercase UTMs, and always populate utm_content with a descriptive variant name so analytics can report at the creative level, not just the campaign level. That single habit is what lets you say “variant B beat variant A three to one” and feed the answer straight back into the next generation cycle. The compounding effect is the whole point: over three to six months, the stack that learns pulls away from the stack that merely ships.

Team and Process: Who Owns What

The best argument for a hub-centered stack is organizational: it collapses a team’s worth of roles into a process one or two people can run. Traditionally, cross-channel advertising required a designer, a copywriter, a media buyer, and an analyst. With the creative, forecasting, competitor, and analytics layers consolidated, most of that work becomes configuration and judgment rather than production.

A workable division of labor for a small team: one person owns strategy and briefs (what to say, to whom, and why); the creative hub owns production and formatting (turning briefs into on-brand variants for every surface); the ad platforms own bidding and delivery; and a weekly review owns the feedback loop (reading performance and updating the next brief). In practice, a single marketer can hold all four, spending a few hours a week generating, launching, and iterating.

This is what the data supports. High-performing teams are not the ones with the most tools; they are the ones that reclaim time from production and reinvest it in strategy. A consolidated stack is how a lean team gets there, and it is why one marketer with the right hub can now out-produce the three-person team of a few years ago.

Common Mistakes and How to Get Started

Most failures in LLM advertising are not exotic. They are the same three mistakes, repeated.

  • Disconnected tools. A pile of single-purpose AI tools recreates the fragmentation you were trying to escape. Consolidate around a hub that covers creative, forecasting, competitor tracking, and analytics, then add media and one or two complements.
  • Manual creative. Hand-producing assets for each assistant and channel caps your volume and starves the platforms’ optimization. Generate at volume and let forecasting pick the winners.
  • No attribution. Launching without creative-level UTMs, conversion APIs, and a plan for the conversation gap means you cannot tell what worked, so you cannot improve. Instrument before you scale.

Getting started costs nothing. Create a free account, paste your website URL so Brand Intelligence can learn your identity, and generate your first cross-surface campaign. Review the forecasts, launch the top variants on the surfaces you can afford today, and tag everything. Then run the loop weekly. Start with Lapis: the creative and campaign hub is free to try (5 credits, no credit card), and it is the one layer you should get right first. Lapis is a Y Combinator (F25) company rated 5.0 on G2, with more than 10,000 campaigns generated across 30-plus industries.

Related guides

Frequently Asked Questions