Trials · version 1.1
Ad trials on simulated streets: rotation, exposure and noticing
How versions of an ad are tested on a zone's hoardings and posters before anyone pays for them: fair rotation, exposure along real routes, noticing priors, a confidence range, and a prediction logged before any real result.
16 October 2026 · updated 17 October 2026 · 7 min read
Before a campaign pays for hoardings, it can run on Parallel Earth's. Two to four versions of an ad go up on a zone's hoardings and posters for one to seven days, the zone's simulated residents pass them on their real routes, and the trial reports which version they noticed, remembered and would act on.
Fair placement
A version must not win because it got the best spot. Boards are numbered, and every day each version moves to the next board.
version(board i, day d) = versions[(i + d) mod V]
- V: the number of versions, 2 to 4
Exposure
A pass counts when someone's route comes within 120 metres of a hoarding or 35 metres of a poster. Waiting nearby for ten minutes or more, at a metro station or a school gate, counts as a pass while waiting.
Noticing
Each pass has a chance of being noticed that depends on how the person passes and the light. Several passes add up.
P(notice) = 1 − Π (1 − p_k)
- p_k: the chance for pass k, from the table below
- Hoardings are lit: at night p × 0.85. Posters are not: at night p × 0.5
- Someone with a worry at home notices less: p × 0.7
| Pass | Hoarding | Poster |
|---|---|---|
| Walking | 0.30 | 0.22 |
| Riding | 0.22 | 0.06 |
| Waiting nearby | 0.45 | 0.50 |
Whether someone who noticed it remembers it or would act on it depends on who they are and the kind of day they are having. That part of the method is not published.
What the research adds
Version 1.1 puts a library under the trial: 48 findings from published research and industry datasets, 30 campaigns that worked and 27 that failed or drew a mixed response, half of them Indian, and five regulator datasets. Every number in the library is as its source states it, and figures that a brand or agency reported about itself are marked as such.
From that evidence come ten hints that adjust how a resident notices, remembers, acts on or objects to an ad. The sources set the direction and the shape of each hint. The size of each effect is our own calibration until real campaigns report, and the trial says so.
| Hint | What it does | Rests on |
|---|---|---|
| Less copy on the move | Each word past seven costs a rider some noticing; walkers lose half as much | US Federal Highway Administration glance study (2012); Pieters, Wedel and Batra (2010) |
| Wear-in, then wear-out | Liking rises with exposure to about ten passes, then slides | Schmidt and Eisend meta-analysis (2015); Pechmann and Stewart (1989) |
| Feeling stays, facts fade | Emotional work is remembered better and longer | Binet and Field for the IPA (2013); Teixeira, Wedel and Pieters (2012) |
| Distinctive assets | A recognisable mascot, colour or jingle registers in a glance | Nelson-Field attention research (2021); Romaniuk (2018) |
| A clear offer | A plain price lifts acting now, with little effect on memory | Binet and Field for the IPA (2013) |
| Humour and trust | Funny ads are liked more but trusted less for homes, law, money and health | Eisend meta-analysis (2009) |
| Fear needs a step | A warning helps only with a clear thing to do | Tannenbaum and others (2015); Witte and Allen (2000) |
| Shock stops people | Provocative work is noticed more, with a backlash risk | Dahl, Frankenberger and Manchanda (2003) |
| Some ads put people off | Religion, gender roles, bodies and big claims draw objections from part of the audience | 27 failed or mixed campaigns, among them Tanishq (2020), Fabindia (2021), Dabur (2021) and Pepsi (2017), and ASCI's complaint reports |
| Familiar themes fade | A theme a resident has met many times is noticed a little less | Rosengren and others (2020); van Berlo and others (2024) |
Residents remember
Each trial leaves a memory with the residents who met it: which themes they saw, and what put them off. The next trial in the same area starts from there, so a theme they keep seeing loses its novelty and a theme that offended them once puts them on guard. The memory is written to the same ledger as the prediction.
Scores and confidence
Each reached resident's best response scores from 0 (ignored) to 1 (would share). A version's score is the mean over everyone it reached, with a 90% interval from resampling the panel 400 times. The trial says medium confidence only when the leader's interval does not overlap the next version's; otherwise it says the two are tied.
An example, simulated
A sweet shop's Diwali offer in Rohini, three versions, hoardings and posters, 5 to 7 November 2026, a panel of 412 simulated residents:
| Version | Reached | Passes each | Noticed | Remembered | Would act | Score (90% range) |
|---|---|---|---|---|---|---|
| Family, in Hindi | 94 | 2.8 | 42% | 31% | 11% | 0.20 (0.15 to 0.25) |
| Hinglish | 90 | 2.5 | 40% | 26% | 7% | 0.15 (0.11 to 0.20) |
| Discount, in English | 96 | 2.5 | 34% | 14% | 0% | 0.09 (0.06 to 0.12) |
Honest by construction
Every trial's prediction is written to the world's ledger before any real result exists, with a hash that changes if anything is edited later. When a real campaign reports, the prediction is scored against it, misses included.
What we checked
- ✓ Versions get an equal share of boards, and their reach differs by less than a third
- ✓ A clear, festive, first-language version beats a dense block of terms and conditions
- ✓ Every prediction is in the ledger before its outcome
- ✓ Every hint cites findings and cases that exist in the library
- ✓ A risky version draws objections; a warm one draws none
Limits, for now
- The library sets each hint's direction; the size of each effect is our calibration, not yet fitted to real campaign results
- The panel is a sample of a zone's households: it ranks versions; it does not predict real reach or sales
- Reactions are simplified; real people are more varied than any panel
- Indian research on outdoor ads is thin, so most findings come from other markets
Written with AI assistance from the colony's working notes and code, and reviewed by a person. Simulated results are labelled as simulated.