The funnel didn’t disappear. It went invisible.
Marketers spent decades perfecting the funnel: awareness, consideration, conversion. We built personas. We mapped content to stages. We watched users click, scroll, bounce, convert. Everything was visible.
But GenAI doesn’t show its hand.
The funnel still exists, it’s just hidden inside the model. Every time someone prompts ChatGPT or Perplexity, they reveal their place in a decision journey.
Not by filling out a form or triggering a pixel, but through the prompt fingerprint embedded in their question.
That’s the new funnel. You’re still being evaluated. Still being chosen. But the targeting is now invisible, inferred, and dynamic.
And most marketers have no idea it’s happening. In fairness, I think only the cohort portion of this is actively happening today.
The ad system I explore here is purely theoretical (though Google appears to be working in a similar direction currently, and its rollout could be realistic, soon – links below).
TL;DR: This article doesn’t just explain how I think GenAI is reshaping audience targeting; it introduces three new concepts I think you’ll need to understand the next evolution of paid media: Prompt Fingerprints, Embedding Fingerprints, and Intent Vector Bidding.
The funnel isn’t gone. It’s embedded. And it’s about to start building and placing ads on its own.
About the terminology:
Prompt Fingerprint and Intent Vector Bidding, I believe, are net-new terms for our industry, coined here to describe how future LLM-based systems could group users and auction ad space.
Conceptually, Intent Vector Bidding aligns with work already being done behind the scenes at Google (and I’m sure elsewhere), though I don’t believe they use this phrase.
Embedding Fingerprint draws from AI research but is reframed here as a brand-side construct to power targeting and retrieval inside GenAI systems.
This article was written over the last three weeks of July, and I was happy to find an article on August 4 talking about the concepts I’m exploring for a future paid ads bidding system.
Coincidental, but validating. The link to that article is below.
Image credit: Duane ForresterWhat Cohort Targeting Used To Be
In the pre-AI era, cohort targeting was built around observable behaviors.
- Retargeting audiences built from cookies and pixels.
- Segments shaped by demographics, location, and device.
- Lookalikes trained on customer traits and CRM lists.
We mapped campaigns to persona types and funnel stages. A 42-year-old dad in Ohio was mid-funnel if he clicked a product video. An 18-year-old in Mumbai was top-funnel if he downloaded an ebook.
These were guesses, good ones, often, but still blunt instruments. And they were built on identifiers that don’t necessarily survive the GenAI shift.
Prompts Are The New Personas
Large language models don’t need to know who you are. They don’t really need to track you. They don’t care where you came from. They only care what you ask, and how you ask it.
Every prompt is vectorized. That means it’s turned into a mathematical representation of meaning, called an embedding. These vectors capture everything the model can glean from your input:
- Topical domain.
- Familiarity and depth.
- Sentiment and urgency.
- Stage of intent.
LLMs use this signal to group prompts with similar meaning, even if they come from completely different types of people.
And that’s how new cohorts can form. Not from identity. From intent.
Right now, most marketers are still optimizing for keywords, and missing the bigger picture. Keywords describe what someone is searching for. Prompt fingerprints describe why and how.
Someone asking “quietest portable generator for camping” isn’t just looking for a product, they’re signaling lifestyle priorities (minimal noise, portability, outdoor use) and stage (comparison shopping).
That single prompt tells the model far more than any demographic profile ever could.
And crucially, that person is joining a cohort of other prompters asking similar questions in similar ways. If your content isn’t semantically aligned with that group, it’s not just less visible. It’s excluded.
New Concept: Prompt Fingerprint
A unique embedding signature derived from a user’s language, structure, and inferred intent within a prompt. This fingerprint is your new persona.
It’s what the model actually sees and what it uses to determine which answers (and potentially which ads) you receive. (More on those ads later!)
When Context Creates The Cohort
Let’s say the Toronto Maple Leafs just won the Stanley Cup (hey, a guy can dream, right?!). Across the city, thousands of people start prompting:
- “Where to celebrate in Toronto tonight?”
- “Best bars near Scotiabank Arena open late?”
- “Leaf’s victory parade time and location?”
None of these users knows each other. Some are teenagers, others are retirees. Some are local, others are visiting. Some are hardcore fans, some just like to party. But to the model, they’re now a momentary cohort; a group connected by real-time context, not long-term traits.
This is a fundamental break from everything digital marketers are used to. We’ve always grouped people by identity: age, interests, behavior, psychographics. But LLMs group people by situational similarity.
That creates new marketing opportunities and new blind spots.
Imagine you sell travel gear. A major snowstorm is forecast to slam into the Northeast U.S.
Within hours, prompts spike around early departures, snowproof duffel bags, and waterproof boots. A travel-stress cohort forms: people trying to escape before the storm hits. They’re not a segment you planned for. They’re a moment the system saw before you did.
If your content or product is aligned with that moment, you need a system that detects, matches, and delivers immediately. That’s what makes system-embedded ad tech essential.
You’re not buying audiences anymore. You’re buying alignment with the now, with a moment in time.
And this part is real today.
While the inner workings of commercial GenAI systems remain opaque, cluster-like behavior is often visible within a single platform session.
When you ask a string of similar questions in one ChatGPT or Gemini session, you may encounter repeated phrasing, brand mentions, or answer structure. That consistency suggests the model is grouping prompts by embedded meaning, not demographics or declared traits.
I cannot find studies or examples of this behavior being recorded, so please drop a comment if you have a source for such data. I keep hearing about it, but cannot find dedicated data.
Looking Forward
Entire classes of micro-cohorts may form and disappear within hours. To reach them, you’ll need AI-powered, system-embedded ad systems that can:
- Detect the cohort’s emergence through real-time prompt patterns.
- Generate ads aligned with the cohort’s immediate need.
- Place and optimize those ads before the window closes.
Humans can’t move at that speed. AI can. And it has to because the opportunity vanishes with the context.
Sidebar: What I Think Is Real Vs. What I Think Is Coming
- Prompt Fingerprints – Live Today: Every GenAI system turns your prompt into a vector embedding. It’s already the foundation of how models interpret meaning.
- Cohort Clustering by Prompt Similarity – Active Now: You can observe this in tools like ChatGPT and Gemini. Similar prompts return similar answers, meaning the system is clustering users based on shared intent.
- Embedding Fingerprints – Possible Today: If brands structure their content for vectorization, they can create an embedding signature that aligns with relevant prompts. Most don’t yet.
- Intent Vector Bidding – Emerging Theory: Almost in the market today. Given current ad platform trends, this kind of bidding system is likely being explored widely across platforms.
Why Old-School Personas Will Work Less Effectively
Age. Income. ZIP code. None of that maps cleanly in vector space.
In the GenAI era, two people with radically different demographics might prompt in nearly identical ways and be served the same answers as a result.
It’s not about who you are. It’s about how your question fits into the model’s understanding of the world.
The classic marketing persona is much less reliable as a targeting unit. I’m suggesting the new unit is the Prompt Fingerprint, and marketers who ignore that shift may find themselves omitted from the conversation entirely.
The Funnel Is Still There — You Just Can’t See It
Here’s the thing: LLMs do understand funnel stages.
They just don’t label them the way marketers do. They infer them from phrasing, specificity, and structure.
- TOFU: “Best folding kayaks for beginners”
- MOFU: “Oru Inlet vs. Tucktec comparison”
- BOFU: “Oru kayak discount codes July 2025”
These are prompt-level indicators of funnel stage. And if your content doesn’t align with how those prompts are formed, it likely won’t get retrieved.
Want to stay visible? Start mapping your content to the language patterns of funnel-stage prompts, not just to topics or keywords.
Embedding Fingerprints: The New Targeting Payload
It’s not just prompts that get vectorized. Your content does, too.
Every product page, blog post, or ad you write forms its own Embedding Fingerprint, a vector signature that reflects what your message actually means in the model’s understanding.
Repurposed Concept: Embedding Fingerprint
Originally used in machine learning to describe the vector signature of a piece of data, this concept is reframed here for content strategy.
An embedding fingerprint becomes the reusable vector signature tied to a brand, product, or message – a semantic identity that determines cohort alignment in GenAI systems.
If your content’s fingerprint aligns closely with a user’s prompt fingerprint, it’s more likely to be retrieved. If not, it’s effectively invisible, no matter how “optimized” it may be in traditional terms.
Intent Vector Bidding: A Possible New Advertising Paradigm
So, what happens when GenAI systems all start monetizing this behavior?
You could get a new kind of auction. One where the bid isn’t for a keyword or a user profile, per se, but for alignment.
New Concept: Intent Vector Bidding
A real-time ad bidding mechanism where placement is determined by alignment between a user’s prompt intent vector and an advertiser’s content vector.
To be clear: this is not live today in any public, commercial ad platform that I am aware of. But I think it’s well within reach. Models already understand alignment. Prompt clustering is already happening.
What’s missing is the infrastructure to let advertisers fully plug in. And you can bet the major players (OpenAI, Google, Meta, Microsoft, Amazon, etc.) are already thinking this way. Google is already looking at this openly.
We’ve Been Heading Here All Along
The shift toward LLM-native ad platforms might sound radical, but in reality, we’ve been headed this way for over a decade.
Step by step, platform by platform, advertisers have been ceding control to automation, often without realizing they were walking toward full autonomy.
Before we trace the path, please keep in mind that while I do have some background in the paid ad world, it’s much less than many of you.
I’m attempting to keep my date ranges and tech evolutions accurate, and I believe they are, but others may have a different view.
My point here isn’t historical accuracy, it’s to demonstrate a continual, directional progression, not nail down on which day of which year did Google do X.
And, I’ll add, maybe I’m entirely off base with my thinking here, but it’s still been interesting to map all this out, especially since Google has already been digging in on a similar concept.
1. From Manual Control To Rule-Based Efficiency
- Early 2000s – 2015
In the early days of search and display, marketers controlled everything: keyword targeting, match types, ad copy, placements, and bidding.
Power users lived inside tools like AdWords Editor, manually optimizing bids by time of day, device type, and conversion rate.
Automation started small, with rule-based scripts for bid adjustments, budget caps, and geo-targeting refinements. You were still the pilot, just with some helpful instruments.
2. From Rule-Based Logic To AI-Guided Bidding
- 2015 – 2018
Then came Smart Bidding.
Google introduced Target CPA, Target ROAS, and Enhanced CPC: bid strategies powered by machine learning models that ingested real-time auction data (device, time, location, conversion likelihood) and made granular decisions on your behalf.
Marketers set the goal, but the system chose the path. Control shifted from how to what result you want. This was a foundational step toward AI-defined outcomes.
3. From AI-Guided Bidding To Creative Automation
- 2018 – 2023
Next came the automation of the message itself.
Responsive Search Ads let advertisers upload multiple headlines and descriptions and Google handled the permutations and combinations.
Meta and TikTok adopted similar dynamic creative formats.
Then Google launched Performance Max (2021), a turning point that eliminated keywords entirely.
- You provide assets and conversion goals.
- The system decides where and when to show your ads, whether across Search, YouTube, Display, Gmail, Maps, and more.
- Targeting becomes opaque. Placement is more invisible. Strategy becomes trust.
You’re no longer steering the vehicle. You’re defining the destination and expecting the algorithm gets you there efficiently.
4. From Creative Automation To Generative Execution
- 2023–2025
The model doesn’t just optimize messages anymore; it writes them.
- Meta’s AI Sandbox generates headlines and CTAs from a prompt.
- TikTok’s Creative Assistant produces hook-driven video scripts on demand.
- Third-party tools and GPT-based agents build full ad campaigns, including copy and targeting.
- Google’s Veo 3 and Veo 3 Fast now live on Vertex AI, generate polished ads and social clips from text or image-to-video inputs, optimized for rapid iteration and programmatic use.
This isn’t sci-fi. It’s what’s coming to market today.
5. What Comes Next – And Why It’s Inevitable
The final leap is where you don’t submit an ad, you instead submit your business.
A fully LLM-native ad platform would:
- Accept your brand’s value propositions, certifications, product specs, creative assets, brand guidelines, company vision statements, and guardrails.
- Monitor emergent cohorts in real time based on prompt clusters and conversation spikes.
- Inject your brand into those moments if, and only if, your business’s vector aligns with the cohort’s intent.
- Charge you automatically for participation in that alignment.
You wouldn’t target. You wouldn’t build campaigns. You’d just feed the system and monitor how well it performs as a semantic extension of your business.
The ad platform becomes a meaning-based proxy for your company, an intent-aware agent acting on your behalf.
That’s not speculative science fiction. It’s a natural endpoint of the road we’re already on, I believe. Performance Max removed the steering wheel. Generative AI threw out the copywriter. Prompt-aligned retrieval will take care of the rest.
Building The LLM-Native Ad Platform
This is a theoretical suggestion of what could be our future for paid ads within AI-generated answer systems.
To make Intent Vector Bidding real at scale, the underlying ad platform will have to evolve dramatically. I don’t see this as a plug-in bolted onto legacy PPC infrastructure.
It will be a fully native layer inside LLM-based systems, one that replaces both creative generation and ad placement management.
Here’s how it could work:
1. Advertiser Input Shifts From Campaigns To Data Feeds
Instead of building ads manually, businesses upload:
- Targeted keywords, concepts, and product entities.
- Multimedia assets: images, videos, audio clips.
- Credentials: certifications, affiliations, licenses.
- Brand guidelines: tone, voice, claims to avoid.
- Business limitations: geography, availability, compliance.
- Structured value props and pricing tiers.
2. The System Becomes The Creative + Placement Engine
The LLM:
- Detects emerging prompt cohorts.
- Matches intent vectors to advertiser fingerprints.
- Constructs and injects ads on the fly, using aligned assets and messaging.
- Adjusts tone and detail based on prompt stage (TOFU vs BOFU).
3. Billing Becomes Automated And Embedded
- Accounts are pre-funded or credit-card linked.
- Ad spend is triggered by real-time participation in retrieval or output injection.
- No ad reps. No auctions you manage. Just vector-aligned outcomes billed per engagement, view, or inclusion.
- Ad creation and placement become a single-price-point item as the system manages all, in real time.
If you want some more thoughts on this concept, or one that’s closely related, Cindy Krum was recently on Shelley Walsh’s IMHO show, where she talked about whether she thinks Google will put ads inside Gemini’s answers, and it was an interesting discussion.
You should give it a listen. And this report on Google suggests this is not only here now, but expanding.
The Human Role Doesn’t Disappear – It Evolves
Marketers and ad teams won’t be eliminated. Instead, they’ll become the data stewards and strategic interpreters of the system.
- Expectation setting: Clients will need help understanding why their content shows up (or doesn’t) in GenAI outputs.
- Data maintenance: The system is only as good as the assets you feed it, and relevance and freshness matter.
- Governance and constraints: Humans will define ethical limits, messaging boundaries, and exclusions.
- Training and iteration: AI ad visibility will rely on live outputs and observed responses, not static dashboards. You’ll tune prompts, inputs, and outputs based on what the system retrieves and how often it surfaces your content.
In this model, the ad strategist becomes part translator, part data curator, part retrieval mechanic.
And the ad platform? It becomes autonomous, context-driven, and functionally invisible, until you realize your product’s already been included in the buyer’s decision … and you’ve been billed accordingly.
A Closer Look: Intent Vector Bidding In Action
Imagine you’re an outdoor gear brand and there’s a sudden heatwave hitting the Pacific Northwest. Across Oregon and Washington, people begin prompting:
- “Best ultralight tents for summer hiking”
- “Camping gear for extreme heat”
- “Stay cool while backpacking in July”
The model recognizes a spike in semantically similar prompts and data from news sources, etc. A heatwave cohort forms.
At the same time, your brand has a product page and ad copy about breathable mesh tents and high-vent airflow systems.
If your content has been vectorized (or if your system embeds an ad payload with a strong Embedding Fingerprint), it’s eligible to enter the auction.
But this isn’t a bid based on demographic data or historical retargeting. It’s based on how closely your product vector aligns with the live cohort’s prompt vectors.
The LLM chooses the most semantically aligned match. The better your alignment, the more likely your product is included in the AI’s answer, or inserted into the contextual ad slot within the response.
No campaign setup. No segmented audience targeting. Just semantic match at machine speed. This is where creative, product, and performance converge, and that convergence rewrites what it means to “win” in modern advertising.
What Marketers Can Do Right Now
There’s no dashboard that will tell you which Prompt Fingerprints you’re aligned with. That’s the hard part.
But you can start by thinking like a model until tools start to develop features that allow you to model your Prompt Fingerprint.
Start with:
- Simulated prompt testing: Use GPT-4 (or Gemini or any other) to generate sample queries by funnel stage and see what brands get retrieved.
- Create content for multi-cohort resonance: for example, a camping blog that aligns with both eco-conscious minimalists and adventure-seeking parents.
- Build your own prompt libraries: Classify by intent stage, specificity, and phrasing. Use these to guide creative briefs, content chunking, and SEO.
- Track AI summaries: In platforms like Perplexity, Gemini, and ChatGPT, your brand might influence answers even when you’re not explicitly mentioned. Your goal is to become the attributed source, not just a silent contributor.
In this new, genAI version of search, you’re no longer optimizing for page views. You’re optimizing for retrievability by semantic proximity.
The Rise Of The Prompt-Native Brand
Some brands will begin designing entire messaging strategies around prompt behavior. These prompt-native brands won’t wait for traffic to arrive. They’ll engineer their content to surf the wave of prompt clusters as they form.
- Product copy structured to match MOFU queries.
- Comparison pages written in prompt-first language.
- AI ad copy tuned by cohort spike detection.
And eventually, new brands will emerge that never even needed a traditional website. Their entire presence will exist in AI conversations.
Built, tuned, and served directly into LLMs via vector-aligned content and Intent Vector Bids.
Wrapping Up
This is the next funnel, and it’s not a page. It’s a probability field. The funnel didn’t disappear. It just went invisible.
In traditional marketing, we mapped clear stages (awareness, interest, decision) and built content to match. That funnel still exists. But now it lives inside the model. It’s inferred, not declared. It’s shaped by prompts, not click paths.
And if your content doesn’t align with what the model sees in that moment, you’re missing in the retrieval.
More Resources:
- Beyond Fan-Out: Turning Question Maps Into Real AI Retrieval
- Brand Bias For Visibility In Search & LLMs: A Conversation With Stephen Kenwright
- The State Of AI In Marketing
This post was originally published on Duane Forrester Decodes.
Featured Image: NicoElNino/Shutterstock
