Repeated ad setups reported different message rates
A Colombian university ran 16 Facebook ad setups repeatedly. The results show why a single reported rate can leave out useful context.
Influence IQ analysis of 384 ad records from one Colombian university recruitment campaign in May and June 2026.
Abstract
A repeated ad setup can produce different reported results across runs. We analyzed 384 Facebook ad records from one Colombian university's recruitment campaign: 16 setups, each run 24 times in May and June 2026. Within each setup, we ranked the ads by reported messaging conversations per 1,000 impressions and compared the 6th and 18th values. In 12 of the 16 setups, the higher value was at least twice the lower one. Across setups, the median gap between those two values was 0.60 conversations per 1,000 impressions. This is a descriptive analysis of one campaign with few messages per ad, not a benchmark for other advertisers. Reporting the counts and a range alongside a rate makes that distinction easier to see.[1][2][3]
A small team sends a client the results of a Facebook ad. One run delivered three reported conversations; another run of the same setup delivered six. The second count is twice the first, but the team still needs the number of impressions, the reporting window and the context of the runs before explaining the difference. The same comparison can look different when the exposure changes.
Repeated runs can report different results

The same setup across repeated ads
Researchers studying university recruitment ads published the ad-level records from a Facebook campaign in southwestern Colombia. Their campaign used 16 setups, each defined by a creative version and its audience targeting. The authors say they launched one ad for each setup on each of 24 days, from May 14 to June 6, 2026. Each ad had a budget of 10,000 Colombian pesos. The 384 records belong to one advertiser and campaign, not to 384 independent campaigns.[1][2]
We asked how reported results varied among the 24 records within each setup. We kept the paper's setups separate instead of comparing different audiences or creative versions. For each ad we divided the number of reported messaging conversations started by its impressions, then multiplied by 1,000. An impression is a display, and one person can account for more than one. The resulting number is conversations per 1,000 impressions, not the chance that an individual sends a message.[1]
The authors report a seven-day click and one-day view attribution setting, and say they exported each ad's metrics the day after it finished. A reported message can therefore be attributed after a view as well as a click; an export taken the next day is not a complete seven-day follow-up. The dataset does not identify the exact launch date of each record. We describe the records as repeated runs, not as a chronological series.[1][2]
The middle of each setup still covered a range
For each set of 24 rates, we sorted them from low to high and selected the 6th and 18th values. These are nearest-rank lower and upper quartiles: they describe the spread around the middle of the observed rates while leaving the most extreme values out of the comparison. We divided the higher value by the lower one. Every setup had a positive lower quartile, so all 16 ratios could be calculated.[3]
In 12 of 16 setups, the upper quartile was at least twice the lower quartile. The median of the 16 setup-specific ratios was 2.26. Separately, the median of the 16 differences between upper and lower quartiles was 0.60 reported conversations per 1,000 impressions. These are summaries of different calculations across setups; they are not the two ends of one typical setup.[3]
One setup makes the arithmetic concrete. Its 6th-ranked rate was one reported conversation from 3,559 impressions, or about 0.28 per 1,000. Its 18th-ranked rate was four from 4,441 impressions, or about 0.90 per 1,000. The latter rate was about 3.21 times the former. These were two ranked records, not necessarily consecutive days or a prediction of future results.[3]
The records contain a median of four reported conversations per ad. Ten ads reported zero; eight of the 16 setups included at least one zero. With so few events, a ratio and its counts tell different parts of the story. At exactly 5,000 impressions, one additional reported conversation changes the rate by 0.20 per 1,000. That arithmetic does not establish why any two actual ads differed.[2][3]
Message-rate spread across the 16 setups
In 12 setups the upper quartile was at least twice the lower quartile; in four it was below twice the lower quartile.
| Upper quartile divided by lower quartile | Setups | Share |
|---|---|---|
| At least 2 times | 12 | 75.0% |
| Less than 2 times | 4 | 25.0% |
Broad click interactions varied differently
The source also reports Meta's “clicks (all),” which includes link clicks as well as interactions such as reactions, shares and clicks on an image or profile. It is broader than link clicks. We calculated click interactions per 1,000 impressions for each ad and repeated the same within-setup comparison, ranking that metric separately from messages.[1][3]
The median upper-to-lower quartile ratio across the 16 setups was 1.15 for clicks (all), compared with 2.26 for reported conversations. None of the 16 click-interaction ratios reached 1.5; all 16 conversation ratios did. The median quartile gap was 13.84 click interactions per 1,000 impressions and 0.60 reported conversations per 1,000, respectively. The larger relative spread for conversations therefore came with a smaller absolute gap in their own units.[3]
The metrics have different definitions and counts. The comparison does not show that clicks led to the reported messages, that messages were caused by the difference in click interactions, or that the same people appear in both numerators. A team deciding whether an ad generated inquiries should look at its inquiry measure as well as any broader interaction count.
Relative and absolute spread by measure
The reported conversation rate had a larger relative spread; the two absolute gaps use different outcome counts.
| Measure | Median ratio (times) | Median gap per 1,000 impressions |
|---|---|---|
| Reported conversations | 2.26 | 0.60 |
| Click interactions (all) | 1.15 | 13.84 |
Put the counts beside the rate
Imagine two fictional runs that each deliver 5,000 impressions. Three reported conversations correspond to 0.60 per 1,000 impressions; six correspond to 1.20. The second rate is twice the first, a difference of three conversations and 0.60 per 1,000. If the runs had different numbers of impressions, the counts alone would not show the same comparison.
A useful report can name the setup and budget, the dates and attribution window, impressions, the reported outcome count, and the rate's numerator and denominator. If the setup has several comparable runs, it can also show the individual results or a clearly defined range. The range should say which runs are included and whether its endpoints are the minimum and maximum, quartiles, or something else.
Keep the practical decision in view. A team may need to decide whether to continue collecting results, change a creative, or discuss a budget with a client. Repeated outcomes provide more context than one run, but this dataset does not prescribe a required number of runs or identify which change would improve another campaign. Agreeing on a decision rule before reviewing new results can help a team explain what it intends to compare.
The source exported results soon after each ad ended. For a team's own report, record when the figures were retrieved as well as the attribution setting. Two exports taken at different times can cover different amounts of follow-up, even when their displayed dates or settings look similar.
Give the rate its context

Identify the setup
Name the creative, targeting, budget and reporting period.
Show the counts
Record impressions, the outcome count and the attribution and export timing.
Describe the range
Show comparable runs and explain how the range was chosen.
Read the results in their setting
The campaign promoted one university program in selected Colombian municipalities, with a small fixed budget per ad and automatic placements. Each setup was repeated, but the people reached, delivery time, placement mix and auction conditions could differ across runs. People could also have seen more than one ad. The data do not identify which of these factors, if any, explains the observed spread.[1]
The 16 setups are different combinations inside that one campaign. They are not a random sample of businesses, and the results do not establish a typical amount of variation for other markets, campaign objectives or budgets. Repeated runs can also share conditions, so this analysis does not treat the 24 records in a setup as independent replications of 24 separate campaigns.
The researchers' paper studied a different question: the effects of spokesperson type and authorship disclosure in its campaign. Our analysis uses their released records to describe variation within each setup. It does not compare human and AI creative or determine whether either version performed better.[1][2]
How we analyzed the records
The public Stata file contained 384 records with unique observation IDs, 16 setup codes and 24 records per code. Within each code, the released creative, spokesperson, authorship disclosure, gender and age fields were constant. All records had positive impressions and nonnegative click and message counts; none was excluded. All also exceeded 1,000 impressions, so that planned filter did not change the sample.[2][3]
We computed each ad's two rates from the released counts, sorted the 24 rates separately by measure within each setup, and selected ranks 6 and 18. Each setup contributed one ratio and one difference per measure to the summaries. A second calculation using SQLite and exact fractions reproduced the group quartiles from a numeric extract of the same file. The reproduction package contains source links, file hashes, code and aggregate results.[3]
As a post-hoc sensitivity check, we used linearly interpolated quartiles instead of nearest ranks. Under that convention, 11 of 16 message ratios were at least two, and their median was 2.09; the median gap was 0.56 conversations per 1,000 impressions. The alternative is a calculation on the same records, not an independent replication. No claim here depends on a significance test or on knowing the cause of the variation.[3]
The paper describes the younger target group as ages 18–24, while a label in the released file says 18–25. We do not report a subgroup comparison or resolve that discrepancy by choosing one label. The downloadable package includes the grouping and measurement checks so readers can inspect the analysis without republishing the original ad videos.[1][2][3]
The downloads contain aggregate findings, source links, a hash and code. The ZIP does not include individual ad records or videos.
Sources & notes
- Arango-Espinal, Rodríguez Orejuela and López-Herrera, Human Spokesperson Versus Artificial Intelligence Avatar and Authorship Disclosure in Digital Advertising, 2026. Campaign design, measures, attribution and export timing.
- López-Herrera, authors’ data and replication material, Zenodo, 2026. Source dataset identified as CC BY 4.0.
- Influence IQ analysis package: source link, hash, calculation code and group-level results.
Every language edition describes the same Colombian campaign and May–June 2026 collection period. Translation does not make it evidence for another market.