AI Can Now Simulate Purchase Intent with 90% Accuracy
Here’s how to build a digital twin of your customer to test campaigns before they ever launch.
A study dropped this week that should be a fascinating read for marketers.
LLMs can now simulate human purchase intent with 90% accuracy.
Researchers at PyMC Labs and Colgate-Palmolive showed that, with the right method, called Semantic Similarity Rating (SSR), large language models like GPT-4o can mirror how real consumers respond to product surveys.
That means:
- You can run “synthetic focus groups” with realistic variance and reasoning.
- These ‘AI customers’ can explain why they’d buy your product or not, offering both quantitative and qualitative insight.
- It’s like having your own personal focus group for marketing assets like landing pages, emails and more.
The Main Discovery
Traditional “AI surveys” failed because they asked LLMs to choose a number (1-5) directly as to if they’d purchase a product, which led models to get stuck at “neutral” (3) or to be overly optimistic (4-5).
Semantic Similarity Rating (SSR) fixes this:
Instead of asking the AI to pick a number like “3” or “4”, the researchers had to write a short answer like a human would in a survey.
They compared the wording of that answer to five example replies ranging from “definitely not” to “definitely yes.”
For example, if the AI said:
“I’d probably buy it — it sounds convenient and the price seems fair,”
that response would sit close to “definitely yes.”
But if it said:
“It doesn’t seem useful for me, and I’m not sure it’s worth the money,”
that would map closer to “definitely not.”
The closer the answer sounded to “definitely yes,” the higher the score.
It’s like reading between the lines for intent to understand how strongly someone feels, rather than just asking for a number.
Why does it matter for B2B marketers?
If you’re a B2B marketer, you can get instant feedback on product positioning, sales copy, prospecting emails, and more. You have your own AI focus group.
You can simulate prospect/customer reactions with high fidelity before you ever launch a campaign.
Run your own synthetic B2B buyer panel:
- Feed your ICP details (role, company size, pains, budget).
- Present your offer, value prop, or ad copy.
- Get realistic purchase intent scores and qualitative feedback like:
How to Apply This (Step-by-Step Tutorial)
Here’s how any marketer can build their own digital twin of a customer using Claude/ChatGPT Projects and LLM embeddings.
Step 1: Collect Real-World Customer Data:
Gather qualitative signals about your buyers. To help you follow this tutorial, I’ll provide steps to create a demo dataset.
We’ll simulate three data sources:
a. Sales Call Transcripts (5 examples)
b. Public Voice Data. We’ll use a prompt to gather this (Reviews, Case Studies, G2 Quotes)
c. CRM Notes (Objections, Motivations, Deal Killers)
i. Generate 5 Synthetic Sales Call Transcripts
The following prompt will create example sales call transcripts you can use. Edit to reflect your company.
Act as a B2B SaaS sales rep at [HubSpot]. Simulate five 20-minute sales calls with decision-makers evaluating our [CRM]. Each call should include:
The buyer’s role (e.g., VP Marketing, RevOps Lead, Head of Sales)
Their key pain points and goals
Questions about pricing, ROI, and integrations
The rep’s responses and discovery notes
Natural dialogue structure (approx. 20-25 turns per call)
Output each transcript with: Call # | Persona | Company Size | Transcript.
ii. Aggregate Reviews / Case Studies / G2-Style Quotes
You can use the following prompt with any of the AI assistants. I like to use Perplexity Labs. You can tailor the below for your company.
Act as a research assistant analyzing public user sentiment for HubSpot CRM.
Synthesize 25 short, realistic review-style snippets that capture what real users might say on platforms like G2, Reddit, and case studies.
Include a mix of:
50% positive
30% neutral / mixed
20% negative
For each snippet, include:
Persona: (e.g., Founder, VP Marketing, Sales Ops Manager, RevOps Lead)
Company Size: (Small Business, Mid-Market, Enterprise)
Sentiment: Positive / Neutral / Negative
Quote: (1–3 sentences, natural tone)
Theme: (e.g., Ease of Use, Automation, Integrations, Reporting, Pricing, Support)
Output as a clean Markdown table with the above columns.
iii. CRM notes on objections, motivations, and deal killers
You are a CRM data generator creating realistic HubSpot deal notes for internal analysis.Create 20 synthetic CRM entries that reflect real B2B SaaS deals in the CRM/marketing automation space.
Each entry should include:
Account Name (fictional company)
Industry (e.g., SaaS, eCommerce, Professional Services, Manufacturing)
Company Size (Small Business, Mid-Market, Enterprise)
Primary Contact Role (e.g., VP Marketing, Head of Sales, RevOps Lead, CMO)
Stage (Discovery / Evaluation / Negotiation / Closed Lost)
Motivations (why they were exploring HubSpot — e.g., unify data, improve pipeline visibility, automate workflows, consolidate tools)
Objections (pricing, integrations, data migration risk, security, perceived complexity)
Closed-Lost Notes (a 2–3 sentence natural-language note summarizing why the deal was lost from the rep’s perspective — tone should sound like a real CRM log, not a polished statement)
Ensure variation across:
Company sizes and roles
Deal stages
Reasons for loss (e.g., pricing, competition, internal delays, status quo bias)
Output as a Markdown table with these exact columns:
Account Name | Industry | Company Size | Contact Role | Stage | Motivations | Objections | Closed-Lost Notes
Step 2: Summarize Behavioral Insights
Analyze the data you collected in Step 1 to uncover what really drives your buyers.
In the PyMC Labs × Colgate study, the breakthrough wasn’t just that AI could give a score but that it could explain its reasoning.
Those explanations revealed purchase intent far more accurately than raw numbers, e.g., asking the LLM how likely you are to buy this product between 1 and 5.
We’ll replicate that idea here.
You’ll use AI to surface the underlying motivations, emotions, objections, and triggers that show why people buy or walk away.
This turns your raw transcripts, reviews, and CRM notes into usable behavioral intelligence for your digital-twin customer.
You’ll create:
A short, structured table that summarizes:
- Motivations – why customers engage
- Emotional Drivers – how they want to feel (safe, efficient, confident)
- Objections – what stops them from buying
- Buying Triggers – moments or conditions that push action
Do the following steps:
i. Gather your Step 1 outputs: Combine your synthetic sales calls, review snippets, and CRM notes into one document.
ii. Ask the AI to read the data and find patterns: This is similar to how the researchers analyzed open-text answers for “semantic similarity.” We’re doing a lighter version, spotting repeated themes that signal intent.
iii. Use this prompt:
Analyze the following customer data (sales calls, reviews, CRM notes). Identify recurring motivations, emotional drivers, objections, and buying triggers.
Summarize them into 5–7 clear behavioral insights using short, marketing-friendly language.
For each insight, add one sentence explaining why it matters for go-to-market or messaging strategy.
Output in this table format:
Category | Insight | Why It Matters
Step 3: Turn Insights into Anchor Statements
Now that you’ve summarized how your customers think, it’s time to translate those insights into something an AI can use.
In the PyMC Labs study, the researchers compared each AI response to five anchor statements that represented the Likert scale from “definitely not” to “definitely yes.” That’s how they measured how close an AI’s free-text answer was to a real human reaction.
We’ll do a simplified version of that for marketing. Instead of “purchase intent,” our anchors will capture customer sentiment or decision confidence toward your product or message.
You’ll create:
A short set of five anchor statements that represent your customer’s mindset from strongly negative to strongly positive. These act as “reference points” for your digital-twin customer when evaluating new ideas, campaigns, or positioning.
Do the following steps:
1. Start with your behavioral insights from Step 2.
You don’t need to analyze them yourself. Just paste your list of motivations, objections, and emotional drivers into the next prompt. The AI will use this context to figure out what “yes” and “no” sound like in your customer’s voice.
Based on the customer insights uploaded, create five anchor statements that represent how a customer might respond to a new product or message — ranging from “definitely not interested” to “definitely yes.”
The statements should:
Sound like real customer language
Reflect emotional tone and reasoning (not just numbers)
Be short, clear, and natural
Output in this format:
1. Definitely not → [statement]
2. Probably not → [statement]
3. Unsure → [statement]
4. Probably yes → [statement]
5. Definitely yes → [statement]
Example Output
Definitely not: “This sounds too complex for our team — not worth switching.”
Probably not: “It’s interesting, but I’m not sure it solves our core problem.”
Unsure: “I can see the value, but I’d need proof it integrates with our stack.”
Probably yes: “This could replace two of our current tools — I’d want a demo.”
Definitely yes: “This is exactly what we’ve been looking for — I’d buy it today.”
Step 4: Build Your Digital Twin Customer
Now that you have your anchor statements, it’s time to bring everything together, your behavioral insights and semantic anchors, inside an AI project to build your digital twin.
In the PyMC Labs × Colgate study, this was the moment the researchers “conditioned” the model: they told it who it was and let it respond to new concepts as if it were a real person.
We’ll do the same thing just applied to B2B marketing.
You’ll create a Digital Twin Customer inside Claude Projects (or ChatGPT if you prefer).
This twin will respond to your new campaigns, offers, or positioning just like a real buyer would.
You’ll create:
An AI project that acts as a simulated version of your ideal customer informed by your Step 2 insights and Step 3 anchors.
It becomes your test environment for messaging, campaigns, and product concepts.
Do the following steps:
1. Open Claude Projects.
Create a new project called:
Digital Twin – [Your ICP Name]
Example: Digital Twin – VP of Marketing (Mid-Market SaaS)
2. Add your behavioral insights and anchor statements.
You’ll paste your tables from Step 2 and 3 into the below instructions.
3. Use this system prompt for instructions (add your details into name/role/company type):
You are a digital twin of [persona name], a [role] at a [company type].
You make decisions like this persona, based on the following insights:[Paste behavioral insights from Step 2 here.]
When you evaluate new marketing messages, products, or campaigns, respond in your own words — as if you were that buyer.
Then, rate your overall likelihood to act using these five anchors:[Paste behavioral insights from Step 3 here.]
Always explain your reasoning briefly before giving your rating.
4. Test it.
Feed it a new campaign, ad, or product description.
For example - I’ll give it the HubSpot home page above the fold screenshot and ask if it instantly communicates the value of HubSpot.
Here is the output from my digital twin, which is a VP of Marketing in an SMB company - Pretty Good!
“Okay, “Where go-to-market teams go to retain” — I get what they’re trying to say, but honestly? It doesn’t grab me. Retention is important, sure, but that’s not my burning problem right now. I’m drowning in tool chaos, I can’t get clean data to show my CEO which campaigns actually drove pipeline, and my team is constantly asking me why we can’t just have one system that actually works together.
The subhead is better: “Unite marketing, sales, and customer service on one AI-powered customer platform that delivers results fast.” That’s closer to what I need to hear — one platform, unity, fast results. But “retain” as the hero message? That feels like it’s solving tomorrow’s problem when I’m still stuck in today’s mess.
What I want to see immediately is: “Stop juggling 10 tools. One platform. Real visibility. Fast.” Show me you understand my chaos and that you’re the answer to ending it — not just optimizing retention once I’ve somehow figured everything else out.
Rating: Unsure
The unified platform promise in the subhead resonates, but the headline focuses on retention when my immediate pain is about consolidation, visibility, and speed. I’d need to dig deeper to see if this actually solves my “dueling dashboards” nightmare — it doesn’t instantly communicate that value to me.”
The applications of this are mind boggling. This is a very simplistic version of how you can build real intent models using the concept of Digital Twins.
Until Next Time,
Happy AI’fying,
Kieran