Testing price and landing before spending on traffic
Before burning budget on traffic, you can discover the price range the market accepts and which version of your copy convinces more, using a panel of synthetic customers. This lesson teaches you the simulated Van Westendorp to find the price window and copy A/B testing with AI personas, always within the golden rule: synthetic explores, human confirms (Train Synthetic, Test Real).
You have a new product to launch. Two questions keep you up at night: how much to sell it for, and which landing will convert. The old way to find out was expensive and slow: put it live, throw about three thousand dollars of traffic at it, wait a week, read the result, and pray. If the price was wrong or the copy was weak, you already burned the budget to find that out. Today there's a step before that: you simulate a panel of customers (AI personas that think like your audience) and ask them, in minutes, which price range feels fair and which version of the page convinces more. It's not meant to replace the real test. It's meant to get you to the real test already having eliminated the obviously bad options, spending media only on the two or three that survived.
You're going to launch a premium plan for your consulting practice and don't know whether to charge 800, 1,500, or 2,500 a month. Testing each range with real traffic is weeks and a few thousand dollars for an answer. Before that, you run a simulated Van Westendorp: you build a synthetic panel of your customer (the CFO of a mid-size company) and ask the four price questions. In minutes you see the window where the price stops sounding expensive and still doesn't sound so cheap it smells like a scam. Then you take only the two ends of that window to the real test, not the five prices from your initial guess.
The firm is going to offer a preventive advisory subscription and has two versions of the landing: one that sells "legal peace of mind," the other that sells "avoid the lawsuit that sinks the company." Which converts the SME owner? Before paying to find out, you run both against a synthetic panel of your customer's personas and read which argument bites harder, and why. The synthetic points you to the favorite and the reason; real traffic, afterward, confirms it with real money.
You're launching a course and have three candidate prices and two headlines. That's six combinations to test, and testing all of them with traffic is expensive and slow. You run a synthetic panel: first the Van Westendorp to find the price window, then the A/B of the headlines within that window. In one afternoon you cut from six combinations to two. Only those two go to paid media. You didn't outsource the decision to AI, you used AI to avoid paying to test options that never had a chance.
You're going to open a senior role and have two versions of the job post: one that sells "cutting-edge technical challenge," the other that sells "quality of life and flexibility." Which attracts the right candidate? Before publishing on paid channels and waiting two weeks to see who applies, you run both descriptions against a synthetic panel of your ideal candidate's personas and read which argument pulls harder, and why. The synthetic points you to the favorite version and the reason; the real applicant funnel, afterward, confirms for real who shows up. Synthetic explores cheaply, real hiring bangs the gavel.
You have three candidate features for the next release and two ways to present the main one in onboarding. Prioritizing everything with real user testing and product metrics takes weeks of roadmap. Before that, you build a synthetic panel of your user's personas and run the pre-test: which feature they say would solve the real pain, and which of the two onboarding messages communicates the value better. In one afternoon you cut from three features to one bet and from two messages to one. Only that finalist goes to the real user test. You didn't outsource prioritization to AI, you used AI to avoid spending a sprint testing what never had a chance.
You're going to launch a new proposal and don't know whether to anchor at 30, 50, or 80 thousand. Testing each range for real is weeks of pipeline and negotiation for an answer. Before taking it to the client, you run a simulated Van Westendorp: you build a synthetic panel of your buyer (the purchasing decision-maker at a mid-size company) and ask the four price questions. In minutes you see the window where the value stops sounding expensive and still doesn't sound so cheap it smells fishy. Then you take only the two ends of that window to the negotiation table, not the three numbers from your initial guess. The synthetic narrows it down, the real proposal closes it.
You're going to offer a new SLA tier and have two ways to communicate the package to the internal customer: one that sells "total predictability," the other that sells "less rework and fewer queues." Which convinces the department that's going to sign on? Before running the expensive pilot and measuring efficiency for weeks, you run both versions against a synthetic panel of your internal customers' personas and read which argument bites harder, and why. The synthetic shows you the favorite and the reason; the real operational pilot, afterward, confirms it with real data. Synthetic explores cheaply, real operations confirms.
You're going to publish a new data-usage policy and have two versions of the announcement: one that emphasizes "protection and LGPD compliance," the other that emphasizes "less bureaucracy in your day-to-day." Which one actually gets the employee to read and comply? Before running the whole internal campaign and measuring adoption for weeks, you run both versions against a synthetic panel of the affected areas' personas and read which argument engages more, and why. The synthetic points to the winner and the reason; real adoption, afterward, confirms whether the control actually stuck. Synthetic explores, the human pilot bangs the gavel.
You're going to launch a paid API feature and don't know whether to charge per request, per user, or per volume, nor at what range. Testing each model with real customers is weeks of deploy, telemetry, and adjustment for an answer. Before pushing to production, you run a synthetic panel of your developers' and platform teams' personas: which billing model sounds fair, which range stops seeming expensive and still doesn't smell like a suspiciously cheap deal. In one afternoon you cut from three models to one and narrow the range. Only that finalist goes to the real-customer test measuring adoption. You didn't outsource the decision to AI, you used it to avoid burning a release testing what never had a chance.
You redesigned the signup flow and have two prototypes: one that opens asking for the minimum data, the other that explains the value before asking for anything. Which one does the user find clearer and less tiring? Before running the expensive usability test with real people and waiting for the schedule, you present both flows to a synthetic panel of your users' personas and read which journey confuses less, where each one stalls, and which one would make the person complete it, and why. The synthetic hands you the favorite and the why, which is what you use to fix the loser; real user research, afterward, confirms who actually completes the flow. Synthetic explores cheaply, the real user confirms.
You're going to propose the valuation range in an acquisition negotiation and have two narratives to anchor it: one that sells cost synergy, the other that sells market expansion. Which one convinces the board to approve the ticket? Before spending weeks of due diligence to find out, you run a synthetic panel of your board members and ask the four anchoring questions: at what multiple does the deal start to look too expensive, and at what multiple does it start to look cheap enough to hide risk. In minutes you see the valuation window that survives scrutiny, and test both narratives inside it. You didn't outsource the M&A decision to AI, you used AI to avoid taking six scenarios to the board that never had a chance.
Alright, let me start this lesson with the scene that hurts the most at launch. You spent weeks on the product, picked a price by guessing (usually "whatever the competitor charges, more or less"), wrote a landing page you thought was beautiful, and threw traffic budget at it. Then the result arrives: it converted poorly. And now comes the cruel part: you don't know if the problem was the price, the copy, or both. You spent money to get more confused. Think about it: what if you could find out the acceptable price range and the best version of the copy BEFORE spending the first dollar on media? You can't have absolute certainty without the real world, but you can eliminate the bad guesses for free. That's exactly what this lesson installs in the way you launch.
The core idea of this lesson. Before spending on traffic, you run two cheap tests with a panel of synthetic customers (AI personas simulating your audience). First, a simulated Van Westendorp: four price questions that reveal the window where your product doesn't sound too expensive nor so cheap it seems broken. Second, a copy A/B test: you run two versions of the landing against that panel and read which convinces more, and why. Both techniques have a ceiling you need to respect, and it's the golden rule of this entire track: synthetic explores, human confirms (Train Synthetic, Test Real). AI narrows down the options cheaply; the real test, with real people and real money, bangs the gavel. Whoever reverses that order saves the wrong way and decides worse.
01Why this is different from just asking AI
Before showing you the technique, I need to pull you out of a trap, because it's where almost everyone slips. "Synthetic panel" isn't you opening ChatGPT and asking "how much do you think I should charge?". That gives you a generic opinion and, worse, a flattering one. AI tends toward what researchers call sycophancy: it wants to please you, so it finds your price reasonable and your copy good. If you ask it badly, it becomes a mirror that flatters you.
The serious synthetic panel is something else. You build specific personas of your audience (not "a customer," but "Marina, marketing manager at a 50-person company, tight budget, burned by a tool that promised and didn't deliver") and ask the SAME structured question to a group of them, reading the distribution of answers, not just one. And here's a technical detail that changes everything, one a 2025 study made clear: if you ask AI for a score of 1 to 5, it flattens everything toward the middle and the result is worthless. The way that works (the method the study called Semantic Similarity Rating) is to ask for a free-text answer, in human language, and then interpret its sentiment. That same study, running 57 surveys and more than nine thousand responses, reached around 90% of a human panel's reliability. Ninety percent is great for exploring. It's not a hundred, and that's why it doesn't decide alone.
Notice this is a cousin of what you saw back in 2.1, about context being everything: the quality of AI's answer is the quality of the context you gave it. Vague persona, vague answer. Rich persona, anchored to your real customer, an answer worth exploring. Fair?
02Simulated Van Westendorp: finding the price window
Now the first test, and it solves the most agonizing question of any launch: how much to sell for. Van Westendorp is a pricing research technique that's existed since the 70s, and its beauty is that it doesn't ask "how much would you pay?" (a question that gives a dishonest answer, because everyone wants to pay little). It asks four indirect perception questions:
- At what price does this start to feel too expensive (you'd think twice)?
- At what price is it expensive, but still worth considering?
- At what price is it a good deal, cheap without being suspicious?
- At what price does it get too cheap, to the point you'd doubt the quality?
Notice the fourth one: there's such a thing as too low a price. In finance we tend to forget this, but in marketing it's gold, because charging too little doesn't just leave money on the table, it makes the product smell fishy. Cross-referencing the four answers from a whole panel, you find a range, a window, where the price stops sounding expensive and still doesn't sound suspiciously cheap. Your price lives inside that window.
What AI does here is simulate that panel. You build 15 to 30 varied personas of your audience (high and low budget, more skeptical and more open, already a customer and never a customer) and ask each one the four questions, requesting an answer in value plus the reasoning. Then you read where the answers cluster. Instead of recruiting thirty real people and waiting two weeks, you have the sketch of the window in one afternoon.
A warning so you don't fool yourself: the window the synthetic panel gives you is a sketch, not the final truth. AI tends to erase the extremes (the study calls this model collapse, it gravitates to the middle and disappears with the rare cases). So your niche customer, the one who'd pay much more or who has a strange objection, might not show up. Use the window to cut the prices from your guess that clearly fell outside it, and take the two ends that survived to the real test. The synthetic saved you from testing five prices. It didn't hand you the final price as a gift.
03Copy A/B with a synthetic panel: which version convinces
Now the second test, which attacks the other question: which landing converts. You usually have two (or three) versions of the page: different headlines, different promises, different orders of argument. The expensive way to decide is to put both live and split paid traffic between them until one wins. It works, but it costs media and time, and during the test half your money is going to the worse version.
The synthetic panel does a pre-test. You present both versions of the copy to the same panel of personas and ask, for each one: what does this page promise you, what convinced you, what left you in doubt, and which of the two would make you move forward, and why. You're not asking for a score (remember, a score flattens everything). You're reading the reactions in text, like a focus group, just in minutes. One agency that tested this described it exactly like that: AI personas as a landing-page focus group, reading where the copy stalls and what confuses, before spending on a real audience.
What you gain from this is qualitative and precious: it's not just "version B won," it's "version B won BECAUSE the promise was more concrete and version A left the persona in doubt about the price." That "why" gives you what to fix, not just which to choose. You rewrite the losing version with the learning, or combine the best of both, and ONLY THEN take one or two finalists to real traffic.
There's a trap here worth naming, because it's the same one as always in this track: don't confuse the panel's reaction with guaranteed conversion. An AI persona saying "I'd click" isn't someone putting down their card. The panel tells you which copy communicates better; only the market tells you which one SELLS better. They're correlated things, not identical ones.
04The golden rule: synthetic explores, human confirms
Here's the heart of this lesson, and it's the phrase I want pinned on your marketing wall: Train Synthetic, Test Real. In plain language: use synthetic to EXPLORE (generate hypotheses, cut bad options, find the range, choose the finalists) and use real humans to CONFIRM (bang the gavel on the decision that costs money). Reversing this is the mistake that turns a good tool into a trap.
Why is confirming non-negotiable? Because of something researchers call verification asymmetry, and it's kind of cruel: you only know if the synthetic panel got it right by comparing it to real data, which is exactly the cost you wanted to avoid. So the synthetic never proves itself. It's a calibrated bet, not a certainty. What backs that bet is a simple risk rule: a low-risk decision (which of two headlines to test first, which angle to explore) can be resolved in synthetic; a high-risk decision (the launch's list price, the brand's central promise) ALWAYS passes through a real human before becoming truth.
And look how this saves you money for real, without deceiving you. You started with six combinations of price and copy. The synthetic panel, in one afternoon, cut it to two. You take only those two to paid traffic. You didn't eliminate the real test, you made it cheap: instead of burning media on six bets, you burn it on two that already went through a serious screening. That's the honest gain. Whoever promises that synthetic replaces the real test is selling you the wrong savings, the kind that costs a lot later.
05How this fits with the rest of your marketing track
To close out the mental model, notice this lesson doesn't live alone. It's the "decide and measure" piece that protects the earlier ones. If back in the personalization module you built a synthetic panel of your audience (the digital twin of the customer), it's that SAME panel you reuse here to test price and copy, so the investment in building good personas pays off twice.
And there's a safety stitch I can't let pass, because AI has already shown up in this track as a risk, not just a tool. When you feed your landing copy and your price numbers to a synthetic panel, you're pasting content into AI's context. Worth remembering what you saw in G.2 about prompt injection: if you're running this in a tool that also reads third-party content, a hidden instruction could contaminate the test. And if the personas are built from your real customers' data, the LGPD ruler you saw in G.8 applies: a customer's personal data turned into a persona is still personal data, and deserves the same care. The synthetic panel is one of the few techniques in this track that touches sensitive customer data; treat it with the respect the Guardian taught you.
To take away: before spending on traffic, you run two cheap tests on a panel of synthetic customers. The simulated Van Westendorp asks four price questions and sketches you the acceptable window (neither too expensive, nor so cheap it smells fishy). The copy A/B runs two versions of the landing against the panel and tells you which convinces more, and why. Both techniques live within a rule that doesn't break: synthetic explores and narrows the options, real human confirms the high-risk decision (Train Synthetic, Test Real). You're not replacing the real test, you're arriving at it having already cut the bad bets for free. Fair? Next.
Do it now
Your mission for this lesson is to run your first synthetic test of price AND copy for ONE of your products that's about to launch (or that you want to reprice). About twenty minutes. Out comes a one-page artifact: the Pre-Traffic Test Sheet.
STEP 1 · BUILD THE PANEL (5 minutes) Describe 5 to 8 specific personas of your audience. Not "a customer," but people with a name, role, budget, objection, and story (e.g., "Marina, marketing manager, 50-person company, tight budget, already burned by a tool that promised and didn't deliver"). Vary them: high and low budget, skeptical and open, already a customer and never a customer. You'll reuse this panel for both tests.
STEP 2 · SIMULATED VAN WESTENDORP (7 minutes) Ask AI that, for EACH persona, it answers in text (not a score) the four questions, with the reason:
- At what price does this get too expensive (you'd think twice)?
- At what price is it expensive, but still worth considering?
- At what price is it a good deal, cheap without being suspicious?
- At what price does it get too cheap, to the point you'd doubt it?
Read where the answers cluster and write down the WINDOW (from "suspiciously cheap" to "too expensive"). Mark the prices from your guess that fell outside it.
STEP 3 · COPY A/B (5 minutes) Paste your TWO versions of the headline or main promise. Ask the same panel, per persona: what does each version promise you, what convinced you, what left you in doubt, which would make you move forward, and why. Note the winner AND the reason (the "why" is what you'll use to improve the loser).
STEP 4 · THE GOLDEN RULE (3 minutes) Write in one line what you're going to take to the REAL TEST: the two ends of the price window and the winning copy (plus, if worthwhile, the losing one adjusted with the learning). That's the only thing that's going to spend media. Everything else the synthetic already cut for free.
At the end, ask yourself the test question: if I had jumped straight to traffic, how many bets would I have spent media on? And how many did the synthetic panel let me cut beforehand? That difference is the money you just avoided burning. And the question that keeps you honest: which of these decisions is high-risk enough that I should NOT trust the synthetic alone?
Where the synthetic panel comes from and why it isn't magic
The market name for this is the synthetic customer (synthetic respondent): an AI-generated persona that simulates how a real person would respond to a survey, interview, or offer. Analysts like NielsenIQ project that this type of input could exceed half of market research by 2027, and startups in the sector have already raised rounds valued at more than a billion dollars, so it's not a passing trend, it's a foundational shift in how research is done. What gives it technical credibility is the method: a 2025 study (Semantic Similarity Rating, or SSR) showed that asking for a free-text answer and interpreting the sentiment, instead of asking for a 1-to-5 score, reaches around 90% of a human panel's reliability. But the same research field honestly catalogs the limits, and you need to know them: sycophancy (AI flatters you too much), model collapse (it erases rare cases and gravitates to the middle), and verification asymmetry (you only prove the panel got it right by comparing it to the real thing, which is the cost you wanted to avoid). That's why the sector formalized the Train Synthetic, Test Real rule and the tiered risk framework: low risk resolves in synthetic, high risk requires human. There are even people selling Validation-as-a-Service, a third-party service that certifies whether your panel matches reality. You don't need to implement any of this under the hood. You need to recognize the jargon and demand the golden rule when the tool (or the vendor) tries to sell you the shortcut that skips the real test.
Practice
1. What's the right goal of running a simulated Van Westendorp with a synthetic panel before launch?
2. You ran the copy A/B on the synthetic panel and version B won by a wide margin. What's the correct next step?
3. Why is the rule 'synthetic explores, human confirms' (Train Synthetic, Test Real) non-negotiable for high-risk decisions like the launch's list price?
Fair? Let's close out this lesson's message together. Before spending the first dollar on traffic, you have a cheap step almost nobody takes: build a panel of synthetic customers and run two tests. The simulated Van Westendorp gives you the price window, and notice it reminds you of something we tend to forget, that there's such a thing as too low a price, which smells fishy. The copy A/B tells you which landing version convinces more, and, most valuably, tells you why, which is what you use to fix the loser. And both techniques only work if you respect the rule that runs through this entire track: synthetic explores and cuts the bad options for free, real human confirms the decision that costs money. You're not replacing the real test, you're arriving at it smarter and cheaper. Whoever skips this step burns media to discover what a synthetic panel would have whispered in one afternoon. And whoever confuses the panel's whisper with the market's final word saves the wrong way, the kind that costs a lot down the road. Run both tests on your next launch and tell me how much you didn't spend.
For the board
On the painit converted badly and you do not know whether it was the price, the copy, or both. Testing first separates the variables.
On the golden rulethe synthetic explores, the human confirms. It sketches the window cheaply; real traffic makes the call.
On the limitthe synthetic panel is a calibrated bet and cannot prove itself. An expensive decision goes through humans.
Thanks for the feedback. It helps sharpen the next lesson.