1. The Setup
In March 2020, puzzle brand Seltzer Goods launched a cold-audience Facebook and Instagram campaign under the tagline "Keep Busy, Stay Indoorsy." The timing was important: the first weeks of COVID lockdown, a moment when millions of people were suddenly home with nothing to do and no clue how to fill the time.
The creative was a warm overhead shot of hands assembling a puzzle, paired with three copy elements: the headline, an "In Stock & Ready To Ship" urgency cue, and a nod to the artist collaboration story behind the designs. It ran against interest-based and lookalike audiences across six placements on Facebook and Instagram.
The campaign worked. Revenue increased by 785% in 30 days, with a real-world ROAS of 9.68x.
The question Imago set out to answer: could a simulation have told the brand, before launch, which audiences would resonate with this ad, which creative framing would perform best and for whom, and where the ad would leave intent on the table?
How Imago Approached This
Imago's method is built around the idea that before you test an ad in market, you can test it against simulated audiences that behave like real people. Not simple A/B testing or focus groups. A simulation that generates the kind of reactions, hesitations, and language your actual customers would have when they scroll past your ad. Here is what that looked like in practice for this campaign.
1. Inputs
We fed the system the ad creative, the full copy, the brand context, and the campaign brief. This includes everything a real person would see: the image, the headline, the supporting copy, the product price range, and the platform context. The simulation needs enough signal to construct a credible impression of the ad as it would appear in a real feed.
2. Defining the Personas
Based on the brand context and campaign targeting, we defined three audience segments. Each one represents a meaningfully different type of buyer the ad was likely to reach.
- The Nesting Homebody: comfort-driven, home-as-sanctuary mindset, already puzzle-curious, looking for personal enrichment during lockdown.
- The Thoughtful Gift-Giver: socially motivated, physically separated from loved ones, looking for something to send that communicates effort without being generic.
- The Family Activity Organizer: a parent under real pressure to find screen-free, multi-generational activities that would keep kids engaged for more than ten minutes.
Each persona is a fully described individual with specific motivations, prior experiences, and decision criteria.
3. Running the Simulation
1000 AI agents, distributed across 3 groups, were exposed to the ad across 100 rounds and six placement contexts: Facebook Feed, Marketplace, and Stories; Instagram Explore, Feed, and Stories. Each agent could react the way a real person would: posting, commenting, ignoring, liking, clicking, or purchasing; with reactions driven by how much the ad resonated with their individual profile.
Agents don't see the ad in every round, mirroring the reality of real-world impression frequency. The simulation also tracks each agent's growing likelihood to buy and simulates actual purchase events, so the output reflects not just who engaged with the ad but who would have acted on it.
4. What the Simulation Captures
The output is two things working together. First, a set of quantitative metrics: CTR, engagement rate, purchase rate, and sentiment scores per segment. Second, and more importantly, the language agents used when they reacted. The real signal lives not in the numbers alone, but in exactly what made someone stop scrolling, exactly what made them hesitate before buying, and exactly what objection the ad never answered.
That combination is what the following sections break down.
2. What the Simulation Found
The three groups reacted very differently to the same ad. The divergence was not really about who noticed it, response rates were almost identical across all three, but about what happened after they stopped scrolling. Sentiment, intent, and the language agents used to describe the ad split along persona groups.
The Three Groups at a Glance
| Segment | Purchase Rate | Positive Sentiment | CTR | Key Signal |
|---|---|---|---|---|
| The Nesting Homebody | 14% | 92.1% | 2.1% | Immediate identity match, strong self-purchase intent |
| The Thoughtful Gift-Giver | 8% | 89.6% | 2.0% | Low negative reactions, gifting reframe was organic |
| The Family Activity Organizer | 10% | 73.6% | 2.2% | Strong warmth, blocked by missing specs and age info |
Inside the Simulation
Each of the 1,000 agents runs as a fully described individual, not a demographic average. They have occupations, interests, bios, and a consistent point of view that carries across every round they appear in.
The campaign delivered a strong ROAS of 5.44x. The overall response rate of 3.42% sits at the upper edge of what is typical for a cold audience. Sentiment was strongly positive at 85.6%, with only 6.4% of reactions hostile to the creative, and sentiment improved from an early average of 0.76 to a late-round average of 0.95 (out of 1).
"The Nesting Homebody"
The framing around staying indoors, keeping busy, and finding comfort in a beautiful object mapped almost exactly onto how this persona already thinks about home and time. Nearly nine in ten reactions were positive.
The resonance came through early:
"Seeing this in my feed right now feels like a sign. I've been looking for something to do with my hands while I work from home. The modern designs are a bonus. Definite scroll stopper."
That impulse strengthened over rounds. By round 80, one tracked agent had moved from vague interest to active anticipation:
"Friday night plans: herbal tea, cozy lamp, and my newest botanical puzzle arriving tomorrow. There's nothing quite like that first pour of pieces and the promise of a calm evening."
Emily Hartwell, Round 80One interesting pattern emerged with prolonged exposure: the same agent who started out thinking about a personal purchase was, by round 85, thinking about sending one to a housebound relative. The ad's meaning drifted from self-treat to gift without any change in the creative itself.
Price was the dominant friction point. The affinity was real; the hesitation arrived at the checkout consideration, not before:
"No way I'm paying $30+ for cardboard pieces!"
Jennifer Miller"The Thoughtful Gift-Giver"
This group produced the cleanest sentiment signal in the simulation.
The Thoughtful Gift-Giver produced strong positive sentiment yet converted at the lowest rate of all three groups. The purchase rate of 8% and a ROAS of 4.08 against a CAC of $6.50 tell a story of affinity without commitment: agents warmed to the ad consistently, but stopped short of buying.
"There's something about sending a piece of art that says 'I'm thinking of you' louder than words."
Priya Kapoor"Nothing says 'thinking of you' like a good puzzle and some handwritten notes."
Elena RodriguezThe simulation makes this visible in a way standard analytics cannot.
One friction was the absence of gift-logistics scaffolding in the ad. Direct shipping and gift messaging were cited as "make-or-break" features by multiple agents, yet the copy does not surface them. Agents are doing the interpretive work themselves, translating a self-purchase ad into a gift context through their own inference. That is a conversion step the ad is currently leaving to chance:
"Hard agree. 'Exercise your mind' is such a generic tagline. Tell me the pieces are sturdy, the print is crisp, or the box is giftable. Give me something concrete or I'm scrolling past."
Sarah MitchellSentiment strengthened over the course of the run, early-round average 0.82, rising to 0.97 by late rounds (delta +0.15), meaning the message compounds positively with repeated exposure. The 'In Stock & Ready To Ship' copy element appears to be a key compounding driver, with agents who initially hesitated on price returning with stronger gifting intent once shipping reliability registered. One agent described the direct-shipping-with-gift-message feature as "the closest thing to a hug in a box" capturing precisely the emotional job this group is hiring the product to perform. The opportunity here is not to change the emotional tone, which is already strong, but to give the ad the concrete gift-logistics language that closes the final gap between intent and purchase.
Another friction, when it came, was about price certainty rather than the product itself:
"For anything over $30? I'm waiting for a promo."
Clara Davidson"The Family Activity Organizer"
This group had the most complex reaction. Overall warmth was there, 73% positive sentiment, but a consistent and vocal minority generated the clearest friction signal in the entire simulation.
The issue was less on price or the visual, but the absence of practical information:
"Modern floral packaging doesn't keep a 6-year-old interested."
Lisa Wagner"Before you buy, check for piece count, age range, and piece thickness. Learned that one the hard way."
Emily Rostova"Show me the piece count, not the aesthetic marketing."
Laura MitchellThe phrase "modern designs" in the body copy appears to have actively triggered that skepticism. Aesthetic language read as a signal that practical specs were being hidden, not as a selling point.
When the framing did land, when the quarantine context and the bonding angle came through clearly, the reactions were warm and concrete:
"Best $22 I've spent this month. Highly recommend for anyone looking for screen-free bonding time."
Laura Mitchell, Round 53The sentiment trajectory across rounds is the most encouraging signal in this group's data, and it meaningfully changes the read on what initially looks like a weak segment. Early-round average sentiment was just 0.47, reflecting the functional skepticism and spec-related objections described above. By the late rounds, average sentiment had risen to 0.73. The implication is that the friction for this group is not permanent. An ad that leads with piece count, age and durability, rather than forcing agents to resolve that information gap themselves, would likely accelerate CTR and purchase rate meaningfully without needing to change the emotional framing at all.
Where to Place the Ad Mattered
| Context | CTR | Key Finding |
|---|---|---|
| Facebook Marketplace · Gift Browsing | 2.0% | Highest CTR, gift intent already present before the ad appeared |
| Instagram Explore · Anxiety Scrolling | 1.9% | Ad offered constructive escape; strong resonance in a low-intent state |
| Facebook Feed · Working From Home | 1.9% | Deliberate mindset helped convert attention into clicks |
| Instagram Feed · Late Night Wind Down | 1.9% | Highest positive sentiment overall (94.4%) |
| Instagram Story · Coffee Break | 1.7% | Most impressions in the dataset; lowest CTR |
| Facebook Story · Quarantine Afternoon | 1.6% | 100% engagement rate among reactors, but clicks did not follow |
3. How the Simulation Compared to Reality
The real campaign delivered a 9.68x ROAS and a $4.87 CPA, with revenue up 785% in 30 days. The simulation produced a 5.44x ROAS and a $4.88 CAC, a strong match on cost-per-acquisition simulated through agent behaviour alone.
The ROAS gap comes down to one input: average order value. The simulation assigns a fixed per-unit price to each conversion. We had no visibility into Seltzer Goods' actual basket size, e.g. customers buying multiple puzzles, or one for themselves and one as a gift, in a single session. That behaviour is exactly what the agent narratives predicted. Adjusting the revenue input to reflect a more realistic average order would close most of the gap.
The more important validation is structural. The simulation gates purchases on a prior click-through and builds intent across rounds, the same logic that underlies real paid-social attribution. The CPA converging to within cents of the real figure confirms that the underlying purchase model is working correctly.
What the original report cannot tell you is why the campaign performed the way it did. The simulation identifies the Nesting Homebody as the conversion engine, the Gift-Giver's missing product proof points as the clearest efficiency gap, and the exact language driving both, all before a dollar of real spend was committed.