Ad testing platforms: choose the evidence behind your creative decision
Choose an ad-testing platform by the evidence it returns, with a complete water-alert campaign example and a clear path from research output to creative approval.
Choose an ad-testing platform by the evidence needed for your next creative approval. Sill, a fictional home-monitoring brand, is choosing a static campaign for a kit that alerts owners to water without stopping the leak. The comparison must distinguish what the composition draws attention to from what intended viewers understand.
For this brief, Sill's product owner requires actual-viewer evidence about alerting versus shutoff, alongside the branding question. A prediction about visual attention leaves that understanding question open. Earlier in development, forecasts and modeled buyer explanations can help the team explore the promise and choose a clearer presentation.
Our ad testing guide covers designing the creative study and turning findings into an edit. Here, compare the research outputs that could support that work, then ask each provider to demonstrate the relevant task and asset format.
Shortlist by the evidence returned
The table groups routes by how the answer is produced and which decision the output can inform.
Evidence route | Output to evaluate for the decision |
|---|---|
Dragonfly AI: predictive visual attention | A model-based assessment of the visual execution. Useful when the unresolved choice concerns the composition; it leaves actual viewer understanding to appropriate response evidence. |
System1 Test Your Ad: consumer-response testing | Emotional and branding measures from people exposed to the creative, under the selected product and design. Distinguish that route from System1's AI screening product. |
CSS Predictive Intelligence: human prediction market | Forecasts about how other people will respond, with participants' supporting reasons. Preserve that task when interpreting the result. |
Ipsos Creative|Spark AI: predicted creative response | Model predictions from a trained ad-evaluation system. Confirm that the intended format and market are covered. |
Vensia: grounded modeled buyer comparison | Responses across simulated buyer groups, explanations and follow-ups around a defined creative choice. Actual viewer recall and campaign behavior require their own evidence. |
Sill's approval turns on product meaning
Sill's marketing lead must choose which execution should enter a limited media test. In this proposed buying decision, the product owner's approval requires evidence from actual viewers about the alert-versus-shutoff distinction and the campaign's branding.
Sill supplies three battery-powered sensors and batteries. The sensors send a phone notification when they detect water. Setup requires a compatible phone, the Sill app, 2.4 GHz Wi-Fi and an internet connection. The kit alerts the owner; it does not close a valve or stop a leak.
Two executions are proposed. They retain the same product facts and invitation to see how the kit works, while changing the headline and visual treatment.
Execution A: catch the first signs of water
Headline: Catch the first signs of water.
Visual treatment: A close view of a sensor beside a small puddle, with a phone notification visible alongside it. The Sill name appears beside the kit.
Body: Three sensors for the places water can go unnoticed. Get a phone notification when a sensor detects water. Requires the Sill app, a compatible phone, 2.4 GHz Wi-Fi and internet. Alerts you to water; does not shut off the supply.
Call to action: See how the kit works.
Execution B: a cupboard you rarely check
Headline: A water alert for the cupboard you rarely check.
Visual treatment: A wider utility-cupboard scene shows the sensor at floor level, with the phone notification in an inset. The same Sill name and kit appear.
Body: Three sensors for the places water can go unnoticed. Get a phone notification when a sensor detects water. Requires the Sill app, a compatible phone, 2.4 GHz Wi-Fi and internet. Alerts you to water; does not shut off the supply.
Call to action: See how the kit works.
These are written specifications for the example. A visual-attention assessment would need the actual rendered files. Because headline and visual treatment differ together, a comparison of the executions would concern the whole route. It would not isolate the effect of one line of copy.
Sill has several possible questions. Does the composition direct attention toward the product and brand? How do real viewers react? What do people forecast about the campaign? Do the intended buyers understand the function and find the use situation relevant?
Dragonfly AI: predicted visual attention
Dragonfly AI's ad-testing page describes predictive attention modeling and support for static and motion-based creative, including social, display, video and native formats. Its published approach gives Sill a relevant route for examining the rendered static executions.
The acceptance question is specific: can the returned assessment help the designer choose between the two compositions? Have the demonstration use the actual assets and intended display conditions, so the recommendation refers to the kit, brand and notification as presented.
Keep the resulting claim tied to the model. Predicted visual attention is different from recorded gaze, remembered branding or an account of what the viewer believes the sensor does. A composition recommendation can be useful while Sill's understanding question remains unresolved.
System1: consumer testing and AI screening are separate products
System1's Test Your Ad documentation distinguishes testing with real consumers from its AI-powered Test Your Ad Screen. The consumer route includes emotional measures and Fluency, which concerns correct attribution of the ad to its brand. Its published coverage includes social, outdoor and print as well as other formats.
For Sill, this makes the exact selected product important. If approval depends on actual audience response and branding, establish the audience, exposure and returned measures for that consumer study. Confirm separately how an understanding question would be captured; an emotion or branding measure cannot stand in for every product-meaning question.
System1 describes Screen as directional AI screening for finished TV and long-form video. A fast screening offer should not silently become the consumer study the buyer intended to commission.
CSS: forecasts about other people's responses
CSS Predictive Intelligence's ad-testing solution uses its HUUNU prediction-market approach. Participants place virtual-currency bets about how others will respond and explain their judgments. Its public page describes creative work from early concepts to finished ads, including print and digital.
That is a distinct human task. A participant forecasting that an ad will work is giving a different answer from describing their own first interpretation of Sill's sensor. Both can contribute to a decision, provided the report retains the question behind the measure.
This route belongs on Sill's shortlist if the marketing lead needs collective forecasts and reasons for a broader creative choice. If the immediate approval requires direct evidence of a particular misunderstanding, agree how that evidence will be obtained rather than relabeling the forecast as a comprehension result.
Ipsos Creative|Spark AI: asset and market coverage
The current UK Creative|Spark AI product page describes an AI prediction model trained to Creative|Spark survey measures. It lists TV, YouTube skippable, Meta social in-feed and short-form video assets. These predictions are a different evidence source from a new survey of viewers.
Sill is currently choosing static executions. The listed video scope is a reason to obtain explicit confirmation before assuming this product covers the proposed files. If the team later commissions a film, its format and research question would define a new selection decision.
Vensia: meaning and audience fit
Vensia builds simulated audiences grounded in purchase, behavioral and research data. Customer evidence and past research can add context. We compare alternatives, investigate differences across groups and return explanations and recommendations, with follow-up questions to the same population.
For Sill, a modeled comparison could explore how an agreed audience interprets the complete presentation and why the use situation matters. Agree the asset representation and supported task first. A comparison of written presentations should be reported at that level; it cannot certify the visual execution.
The missing approval evidence changes the supplier choice
Sill should make the unresolved approval visible to the supplier. “Choose the strongest ad” is too loose if different people mean visual composition, branding, product understanding and conversion.
Suppose the design team needs to resolve where the brand and device sit within the static frame. A demonstrated predictive-attention output can inform that composition decision. The team should retain the original assets and the exact proposed edit so it can assess the revised execution.
Now suppose the product owner is worried that the campaign implies automatic leak prevention. An attention result cannot resolve that concern. Sill needs the relevant interpretation evidence, with the first-exposure question and answering audience preserved. A modeled comparison can explore the explanation and possible wording repairs; an approval requiring actual human understanding needs a suitable human design.
Our message testing guide covers the narrower wording decision if that becomes the remaining task. The platform comparison has done its job when it identifies which evidence Sill still needs to commission.
The answer can also be to revise the asset before testing. If an agency adds a closing-valve animation to either execution, Sill's product facts already provide the objection: this kit does not close a valve. Send the asset back for correction. A high prediction or favorable audience score would leave that unsupported demonstration in place.
For Sill's stated approval, the first commission should therefore be a consumer-response study with the required understanding task explicitly agreed. System1 documents a human route that makes it a relevant conversation, but its standard emotion and branding measures alone would leave this particular acceptance condition incomplete. If the supplier cannot demonstrate that task, Sill needs a different human research engagement.
A team choosing composition or exploring the message before final approval could select a different route. Sill's stated approval still needs its actual-viewer understanding task, while its static files need format confirmation from the listed Ipsos video product.
The report needs a traceable recommendation
For this campaign, a useful supplier demonstration follows one actual execution through its analysis and proposed next action. Sill should be able to identify the version, the source of the response or prediction, the measure and the reason the recommendation follows.
If Sill receives this | The approval still needs this distinction |
|---|---|
A predicted attention assessment | Which composition change follows, and what remains unknown about understanding and branding. |
Human emotional and branding results | Which people saw the asset, how they encountered it and which question each measure answers. |
A prediction-market forecast | The outcome participants were forecasting and the reasons supporting their judgments. |
A modeled audience comparison | The grounding, represented encounter, group differences and explanation behind a proposed creative change. |
Do not combine those results into an invented common score. Their units, tasks and evidence sources differ. Read them against the approval each was commissioned to inform. Our synthetic vs human respondents guide explains the response-source choice in more detail.
Agree which work the platform supplies and which work stays with your team: preparing final assets, checking product claims and interpreting disagreements between measures. A reusable workspace is useful when those responsibilities lead to a recommendation someone can act on.
For Sill, that could mean revising a composition, commissioning a human understanding question, narrowing a message comparison or returning an unsupported depiction to the agency. Those are different actions. The selected output should make the appropriate one clear.
Creative approval starts the next campaign test
Approving an execution for media testing creates the next question. Does it bring the intended people to the destination, and does that destination meet the expectation the ad created?
Define that event before launch. For Sill's explanatory campaign, a visit to the setup explanation may be an intermediate outcome; a completed order is a different event. Keep placement, audience, delivery and creative versions identifiable when comparing results.
A prelaunch research result does not supply those campaign outcomes. Use actual delivery and behavioral records for the next decision, and investigate whether a change concerns the creative, destination, targeting or product. The platform comparison should help the team choose what deserves that live test.
The next research job sets the brief
If the campaign is still taking shape, Vensia can compare how simulated buyer groups interpret the alert promise and use occasion, then help your team decide what to revise. Set that exploration brief separately from Sill's later actual-viewer approval and live campaign test.
Book a demo to explore which wording and use situation make the water-alert offer clear and relevant to the buyers you need.
