1. INTRODUCTION

Uncovering the Hidden Effects of Advertising

“The sales of a brand are like the height at which an airplane flies. Advertising spend is like its engines: while the engines are running, everything is fine, but, when the engines stop, the descent eventually starts”.

Simon Broadbent, ‘The Advertising Budget’

When advertising success stories are reported, for example in submissions to awards competitions, they usually show dramatic gains in sales or brand share, with ambitious targets comfortably met or exceeded. But most marketing directors in competitive markets, faced with having to justify their annual ad budgets, may well feel exasperated by such tales which seem to have little to do with them. Very often, they are not launching, relaunching or repositioning from a poorly supported low point. On the contrary, their brand has been maintaining a stable share against equally stable competitors, in a largely saturated market. Sales or share movements, where they occur at all, are small and transitory. Over the medium term, nothing much seems to happen.

In this situation, which is common in the developed world, it is almost impossible to demonstrate the efficacy of advertising from aggregate sales or consumer purchase data alone (Jones 1992; Ehrenberg 2000; Ehrenberg, Barnard, Kennedy and Bloom 2002). Econometric modelling is of little use, because it needs sales changes to work on, and these are not happening. When little blips in sales do occur, they are much more easily related to price promotions than to advertising. The ability to observe these promotional effects, but few similar links for advertising, is one of the main reasons why increasing proportions of marketing budgets have been steered away from advertising towards price promotions or highly targeted direct response campaigns. It is not because these tactics are known to be more effective or less costly (indeed they have often been shown to be the reverse), but because the sales effect seems faster, more immediate, and easier to see. Any belief that the advertising is working for the brand is based, on the other hand, more on faith than observation.

We argue in this report that marketers need not, and should not, have to put up with this. We explain that the inability to see advertising effects (in full) from aggregate data alone follows naturally from our modern understanding of how advertising works to influence buyers’ propensities to buy the brand. We then introduce and describe an alternative approach, in which sales can be analysed at the level of the individual buyer, by observing whether he/she is more likely to buy the brand after being exposed to advertising than at other times. This individual level (or single-source) method has been shown capable of revealing the hidden effect of advertising, even when the outcome is not growth but simply to maintain a stable share in face of competition.

The single-source method has already been used by several researchers to discover generalised learnings about advertising effects on purchasing behaviour, which we shall briefly describe. Why should it not also be a practical management tool for analysing the effectiveness of specific campaigns and media planning approaches? We consider how such a tool could now be developed, and the benefits that would arise.

The Starting Point: How Advertising Supports Brands When They are Established and Stable

Let us put aside the big, exciting, structural changes which feature so often in the literature (but so seldom in real life). Successful brands, in most markets most of the time, maintain a roughly stable brand share: if they really change market share, it is only by very small amounts, when for some reason they have managed a temporary advantage. Their action is essentially defensive as they have many competitors who are just as strong, and equally defensive, and these brands usually sell to the same sort of customers.

Unless advertising is radically pulsed or recommenced after a long absence, it is a common complaint that brand awareness and image tracking measures, as well as sales, tend to show up simply as an essentially horizontal line on the time graph. We do not usually interpret this lack of movement as evidence that our advertising isn’t working. On the contrary, we assume it is, but so is our competitors’ advertising (and PR etc) and they cancel out. There is evidence from many case histories that if advertising support is removed or weakened for long enough, in comparison with the competition, a brand’s sales will eventually decline, not at once, but with gradually increasing speed.

Similarly, a change in share of voice or in the comparative quality of advertising results in only slow changes in sales levels. This sluggishness in market response to advertising is partly because buyers are creatures of habit. They have well established propensities of how often to buy and which brands (Ehrenberg 2000; Ehrenberg et. al 2002)^. Advertising can only nudge these propensities, and when it does so, this takes a long while to filter through to aggregate sales levels because most buyers, even of the brand leader, buy infrequently. This is why pulsed advertising campaigns do not often create yo-yo sales patterns.

The part of any brand’s advertising that falls on buyers who do not come into the market for some while can have no effect on current sales. But if the advertising successfully nudges these buyers’ propensity to buy that brand then it delivers future sales, though largely by reducing the impact of competitors’ advertising to steal (by nudging) sales. Much of advertising’s total sales effect occurs in this way, so even a burst of advertising shows little of its real sales effect in current sales volume.

If most advertising is defensive, especially for brand leaders, how does it work? It has been convincingly argued (eg Ehrenberg et. al 2002) that it will seldom be by persuasion (i.e. changing people’s opinions, beliefs or emotive feelings towards the brand). Persuasion is not easy to do, because attitudes and perceptions are conditioned by habits of thought and behaviour, including previous usage, that have grown up over time and are resistant to change. Even advertising which is intended to persuade people to change can only deliver its message, as new, once to each individual (it may of course take time to reach all reachable individuals, but that is another story). After that, it can only reinforce by repetitive presence. Once the ad has been noticed and responded to, repeated viewing can only refresh, remind, and reinforce that original response. Reinforcement by repetition is therefore the mechanism for all advertising, whether it is intended to be persuasive or not.

Much advertising does not even seek to be persuasive. Instead advertising seeks to build and refresh memory structures which enhance the brand’s salience – making it more likely to come to mind or be noticed in a buying situation (see Ehrenberg-Bass Institute Report 16 for Corporate Members). Established brands, both large and small, typically keep a proven theme alive by regularly changing the executions in detail: noteworthy examples have been Andrex, PG Tips, Nescafé, Kenco, Stella Artois, Audi, Absolut vodka and Walkers crisps. This kind of strategy is intended precisely not to change perceptions but to reinforce (i.e. prevent decay) existing perceptions most effectively: the advertising demonstrated in such examples is reinforcement and refreshment of the already familiar.

Advertising can occasionally have a greater effect on salience than the usual nudge, but such outstanding creativity and media placement is unfortunately not common. Fortunately salience maintenance is possible even without outstanding creativity (though great creativity is desirable). Whilst people can be quite resistant to learning new information, they are usually happy to be reminded of things they already know. Competitive advertising has a two-sided task: to reinforce favourable habits of behaviour, and try to insert alternative perceptions (extending memory structures) where the habits are less favourable. Both tasks can only work at the level of an individual responding to stimuli, probably unconsciously. The response will be easiest if it is consonant with an already favourable disposition: advertising will more easily reinforce a current habit than it will instigate a new one. Repetitious reinforcement is the obvious mechanism for both tasks – the slow drip of water wearing away the stone.

What advertising can achieve in the short term is limited, by the context just described. It is however essential, not merely because markets are competitive (other brands’ advertising will eventually push your brand out of mind (and out of market) if you don’t match them). If brands lose their salience for buyers, their habits of loyalty eventually break down. Thus regular nudging of salience is necessary, especially in the face of attempts by competitors to nudge their brand to a higher probability salience.

Reinforcement is the process of advertising, and it works only through individual perceptions of individual ads. But sometimes it does not work. To be able to establish whether a campaign or execution is working for the brand (or is still working) seems essential for marketing management, but it can only be done at the individual level.

2.  What are individual-level AD sales effects, and how could we measure them?

We define it as an ‘individual-level advertising effect’ when we are able to show that buyers are more likely to buy our brand when they have recently seen advertising for it than when they have not. 

This requires ‘single-source’ panel data (purchasing and ad receipt over time from the same people).

These individual-level effects are sales effects: there is a correlation between seeing ads and the likelihood that the brand will be purchased. Yet these sales effects do not imply that there will be an overall, aggregate increase in the brand’s current sales level. On the contrary, we are asserting that these individual effects can be used to establish whether the advertising is working even when (indeed especially when) the brand remains at a stable share and sales level. It is very important to understand clearly the difference between these micro-level effects on purchases and macro-level sales. Macro-level sales may show none of the advertising effect, or at least understate it, and macro-level sales shifts may be more to do with media strategy and spend than advertising effectiveness.

In recent years, mainly through single-source consumer panel data, we have learnt that individual level advertising effects on buying can be observed, and have learnt something about the conditions under which they vary. Later in this paper we shall review this evidence briefly. But first, we want to consider the benefit to marketing managements of having a single-source research tool that could reliably and quickly measure individual level effects of advertising campaigns. We discuss later how we think such a tool might be developed.

We stick here to behavioural measures (actual buying probabilities), because that is the hardest evidence. This is not to discount awareness, salience, attitudes, brand perceptions, claims about purchasing or usage habits and other measures, which have at least a diagnostic importance and may sometimes be the only measures available. We may want these for a range of reasons that are important for marketing, but none of them have the immediacy of behavioural measures, if we are able to relate them to receipt of ads.

3. Benefits of a single-source individual-level research tool

First, we know that not all advertising works equally well (otherwise, companies would not need to spend the money they do on pre-testing it). The purpose of measuring individual level effects on sales is simply to determine whether it is producing a return or not . We look for the probability that buyers will buy our brand after they have been exposed to our advertising recently compared with when they have not. We would expect to quantify this difference: for example, depending on previous knowledge of how our advertising performs, we could set a control level of, say, 10% or 20%, so that if the probability of purchase after advertising was less than 20% above the baseline (the probability without advertising), that would be a signal to consider whether we should change the content.

This quantified positive effect, if established, is a real return: the presence of advertising produces 20% more sales dollars than are achieved in its absence. The return on investment can thus be measured, even (we stress again) when aggregate sales and share are not changing.

Having established that there is a positive effect, we can see how it may vary under different conditions. Details can be explored which are important for creativity assessment and media planning. We can see whether increased dosage of advertising increases the purchase probability, and if so whether it is at a linear rate or diminishing returns (thought to be the typical response function). We can see what difference proximity of the advertising to the next purchase makes, and whether this is affected by increased dosage. On the other hand, we can see whether the effect of the advertising decays according to the length of time elapsing between the purchase and the latest exposure. We can look at interaction with price promotions and other marketing activities. We can compare different media opportunities-to-see (or hear), and see whether they produce different effects in combination than they do separately. All these would provide valuable, practical information which could feed through into media scheduling plans and optimisers.

All these detailed variations would follow once the basic measure was established, linking purchase probability to advertising exposure or opportunities-to-see (OTS). We now review the evidence that such a measure (a ‘single-source method’) is indeed possible.

4. The evidence for individual-level effects

STAS

We start with John Philip Jones and his STAS measure, since this is the best-known single source measure and the most influential evidence (Jones 1995).

STAS (Short-Term Advertising Strength) is a ratio of two brand shares. From a single-source panel, it is possible to take all occasions when a product category was purchased and divide them into two groups according to whether or not there had been an opportunity-to-see (OTS) for the brand of interest in the previous few days (Jones used a standard of 7 days, but this can of course be varied). The share of category purchases devoted to the brand can then be calculated for both groups, the ‘exposed’ and ‘unexposed’, and compared. If the two are equal, this gives an index of 100. If the brand share of purchase occasions in the ‘exposed’ group is higher than in the ‘unexposed’ group, the ratio between the two shares will index at higher than 100, and it can be inferred that the advertising has increased the purchase probability (i.e. the probability that the next purchase will be for our brand). For example (figures are invented):

Jones analysed 78 brands in 12 product categories from the Nielsen Household Panel in the US. STAS was found to vary considerably between brands. When the advertised brands were divided into deciles, 70% of them were found to be positive, of which about 50% had a STAS value over 110. 30% were negative (below 100). ‘Negative’ here is interpreted to mean that the advertising of other, competitive brands were more effective, luring consumers away. That is, not that the brand’s advertising has lowered the propensity to buy of those who viewed it, but rather that it was not sufficiently effective to counter competitors’ advertising – even amongst those who were exposed to the advertising (and less still of course amongst those who were not).

This range of STAS effects from the US Nielsen panel has been found to be closely replicated in data from the UK (Adlab panel, 67 brands), Germany (Nielsen, 28 brands), France (IRI BehaviorScan, 144 brands), and Denmark (ASTAS, 23 brands), in spite of differences in country, time periods, brands and methods of measurement (Jones 2002; McDonald 1996). We can thus conclude that the finding is robust. Other researchers, most notably Andrew Roberts working with the Taylor Nelson Sofres single-source panel TVSpan in the UK (Roberts 1998; Roberts 1999; Roberts 2000), have produced similar evidence that exposure to OTS increases the likelihood of a brand being purchased.

Purchase-Viewing and other biases

STAS, unfortunately, is not free from certain kinds of bias, which must be corrected for in any measurement. The most important bias to be aware of, as has been pointed out by Broadbent (1999; 1996) and others, is Purchase-Viewing (PV) bias. PV bias occurs because, in a sample of purchase occasions (from different people) distributed into a contingency table as shown above, there is no control over which people the purchase occasions in each cell come from. If there is a positive correlation between buying a brand and TV viewing, say, so that the heaviest buyers are also the most likely to be exposed to TV OTS, those buyers will contribute above average to the top right-hand cell, and below average to the bottom left-hand cell. This will mean that the `STAS effect’ cannot be attributed to the advertising alone. Such PV bias has been shown to occur widely. It will tend to inflate STAS if positive, and moderate STAS if the bias is negative (i.e. when a brand is used by and advertised to mainly light viewers).

Fortunately, it is possible to correct for PV bias (and other possible biases) by doing separate calculations for different relevant subgroups such as light and heavy viewers and averaging the results, or using some other form of weighting. Andrew Roberts, for example, controls for confounding effects by “first allocating households into separate groups depending on their overall level of TV viewing and loyalty to the brand” (Roberts 1998); each group is then analysed separately, before the results are averaged to calculate overall effects. Roberts thus uses a contingency table analysis similar to STAS but calculates the advertising effect in a slightly different way.

It is also possible to design specific analyses so as to avoid PV bias altogether. In the very first analysis of single-source data, for JWT in the late 1960s (McDonald 1970), Colin McDonald studied how individuals switched brands. In sequences of three purchases where the third purchase was a change of brand (A – A – B), it was found that more ads for A occurred in the first interval (A – A) than in the second (A – B), and the reverse for B ads. Since the same individuals were involved in both purchase intervals, there could be no PV bias here: we must be observing a real response to advertising. Extending this principle, when only switches (into or out of any brand X) were counted, it was found that switches to X exceeded switches from X only when two or more OTS had occurred: since switches in and out are equal for any individual panel member, by definition, PV bias is again removed. Such unbiased analyses, unlike STAS, come at a price of using only part of the data, but they serve to prove the case that there are real effects which can be seen and measured⁽¹⁾.

It has thus been proved that real short-term effects (correlations between OTS and purchase probabilities) can be observed and measured, and that confounding biases such as PV can be estimated and allowed for to arrive at more accurate figures.

5. How do individual-level effects vary?

Establishing that we can observe and measure individual-level effects is only the first step. The question that immediately occurs is, do they vary under different conditions, and if so how? From this we can learn details of great tactical value to media planners, such as:

• whether there are benefits to be gained from increasing the frequency and/or concentration of ‘exposures’,
• how quickly effects decay when there are no exposures,
• the effect of bursts or pulses versus continuity of exposure,
• the degree of synergy obtainable when advertising is combined with promotions,
• the importance of share of voice in the category,
• effects of different media combinations
  … and more.

We can consider these as potential variations or modifications to the overall sales ‘effectiveness’ measure, which can be translated into dollar returns in the same way. Work is underway on these issues, but researchers have already made a number of discoveries that bear on these questions. In this next section, we summarise what we now know, its marketing implications, and where our knowledge is still weak.

i) Creative treatment

The primary question is, does this creative or branding execution work at all? If we establish, by STAS or other means (and after accounting for PV or other biases) that there is an effect, we take this as evidence that the ad is indeed working, in that ‘exposure’ is followed by a higher probability of purchase than non-exposure. We stress, again, that this is not to be confused with a sales increase overall, whether or not it coincides with one. What we observe is that exposure affects the probability of the next purchase, for some individuals.

Sometimes we see no effect, or even a ‘negative’ one. Do we automatically interpret this as showing that the ad is not working? This is slightly less easy to say. Jones (1995; 2002) is definite on this point: a nil or negative STAS means that the ad is clearly not helping the brand and it should be jettisoned without delay, whilst on the other hand an ad with a positive effect should be continuously supported as much as can be afforded. However, there may be some brands, and products, where the advertising may still be contributing to long-term maintenance even though there may seem to be little or no short-term stimulus. In such a case we might look for evidence of perceptual shifts even if overall salience does not rise, i.e. repositioning effects rather than sales effects. Alternatively there may be ads with a high STAS, little competitive effort, and yet still no aggregate level sales shift. This would strongly suggest a media strategy problem, thus individual level advertising effects can provide a useful diagnostic.

ii) Diminishing returns

Assuming that there is a positive effect, does it vary with different numbers of ‘exposures’ (or OTS)? This question has been crucial to the arguments during the past three decades about ‘effective frequency’.

All the single source studies agree that higher OTS frequency very often produces diminishing returns. That is, if a STAS-type measure is used, the biggest increase in brand share tends to come between 0 and 1 OTS, the next biggest between 1 and 2 OTS, and so on, with the increase tailing off until it disappears (probably around 4 OTS for the most part).

Jones argued that the jump between 0 and 1 OTS was overwhelming⁽²⁾ enough to swamp any further increases, leading to his famous claim that ‘once is enough’, a major influence on the development of recency scheduling.

Plenty of other data (eg McDonald 1996) show that a diminishing returns curve with increasing OTS is the most common, though not the only, pattern. Andrew Roberts found that established brands usually show diminishing returns, but in some situations (e.g. new launches) higher frequencies may continue to bring proportionate benefits with a linear curve (Roberts 1996). Lisa Beaumont in a more recent paper (Beaumont 2003) confirms, from TVSpan analyses of 113 brands, that diminishing returns are typical: purchase probability rises by about 3% after 1 OTS to plateau at about 6% after 3+ OTS, measured over 28 days.

In summary one OTS seems to have the greatest effect but extra OTS do generally have a positive impact on propensity to buy. The effect of further exposures is probably stronger in high clutter environments where more OTS are required to really reach buyers. This is an important media strategy issue that needs testing and quantification.

iii) Thresholds

Another long-standing discussion is whether there are thresholds, i.e. whether purchase probabilities increase much more after 2 or 3 OTS than they do after just one, making an S-shaped response curve. There is some support for the idea that new brands or campaigns may require a learning period before the message takes effect, following the model proposed by Krugman (1972). Evidence for this, from South African tracking data, is quoted by Du Plessis (1995). But this is not necessarily true for established and familiar brands (which do not need to be ‘learnt’).

Roberts (1996) found that out of seven launches and previously unadvertised brands only one showed an S-shaped curve. Beaumont’s more extensive analyses (Beaumont 2003), however, have shown that the patterns vary both in shape and steepness of curve. Launch campaigns tend to produce much higher uplifts (naturally, since they start from zero) than maintenance advertising. Beaumont quotes one example of an established brand in a non-seasonal market where the response curve over a 4-week period was mildly S-shaped. For another brand, four evaluations were conducted over a two-year period. In the first, when the brand was new to television, the response curve (for OTS over 28 days) was S-shaped as familiarity was built. After that, a second phase of advertising aimed to remind people, and optimal exposure occurred sooner, with returns diminishing after 2 OTS. By the third phase, advertising was now working like that for an established brand, and showed a classic diminishing response curve, maintaining and building sales. A fourth phase changed the creative strategy and incorporated a seasonal message, and this behaved like the first phase, again requiring a threshold of three or four OTS in the four weeks before uplift occurred.

She also quotes two cases which seem to contradict the learning theory, one for a new launch and one for a brand extension launch, both of which showed rapid response to one OTS in 4 weeks, but no significant uplift from higher numbers of OTS (a step-shaped curve).

The lesson would appear to be that advertising largely works by refreshing memory structures, and that only really new material has a wear-in before it starts refreshing. The best creativity may therefore very well be that which does not have to be learned, even if it is new, fresh and interesting. It is easy to think of advertising for long standing successful brands that shows wonderful continuity in spite of regular updating and changes in execution.

Again the possibility of thresholds is something that needs more R&D, but the available single source evidence to date does not suggest that ‘wear-in’ is very prevalent.

iv) Proximity to the purchase

There is evidence that larger numbers of OTS can, in some cases, have more ‘effect’ when they are received shortly before the purchase occasion – an interaction between frequency and proximity (McDonald 1996; McDonald 1970; Roberts 1999).

Roberts (1999), analysing 113 brands in ten markets from TVSpan, has found that concentrated OTS immediately prior to purchase had a far greater effect than the same number spread over a week or more. If only one OTS was seen before the purchase, whether 1 day before or over the previous 28 days, the percentage increase in purchasing was about 4-5%, but if 3 or more OTS were seen the effect rose from 5% (seen over 28 days) to 20% (over 3 days) and a huge 50% if the 3+ OTS were concentrated into just one day before the purchase.

Purchase probabilities steadily decrease as the interval since the last OTS increases. Roberts fitted an exponential curve to his data showing a decay rate of 4.4% per day, indicating a half-life of about 16 days. He argues that this implies that advertising effects are likely to continue beyond 28 days: “only 72% of the full effect of advertising will occur within the first 28 days”.

Roberts further notes that while increases from 3+ OTS within 3 days before the purchase occur in all viewing groups they are much greater among lighter viewers (this was also confirmed by Adlab analysis)(McDonald 1996).

Reichel (1992) reported a STAS-type analysis (using a slightly different calculation method) of 45 brands from Nielsen’s Homescan panel over a two-year period and found similar results to other single-source studies. He found a median increase of +10 to +14% among those exposed to the advertising compared to those who were not exposed. He also found that the further removed the advertising was from the purchase, the faster the effect dropped. 19 brands had substantial enough advertising budgets to make this more sensitive analysis possible. The average increase for these 19 brands declined from + 21% when OTS were last received 1 day before the purchase down to +11% when the last OTS were 28 days before the purchase, with progressive decreases in between. In other words, a delay of a month before the next purchase halved the advertising effect compared with purchases made the next day.

Beaumont (2003) finds that the uplift for an average grocery brand attributable to TV OTS is just below 5% if the ad was ‘seen’ in the 3 days prior to purchase, but “even when the advertising was seen at any time in a month prior to purchase, sales are about 3% higher than expected”. Again this is evidence that proximity can almost double the sales impact (compared with being exposed to the advertising a month ago), but also evidence that advertising effects can last for a considerable period.

v) Share of Voice

This finding that advertising concentration can be effective raises the question of ‘share of voice’ (SOV). This means comparing, not different numbers of OTS for the brand in time periods before the purchase, but different brand shares of OTS for the category. This of course interacts with the numbers: higher SOV tends to coincide with less category advertising. Relatively little SOV analysis has been done on single-source data, but there is some evidence from the Adlab panel (McDonald 1997) that purchase probabilities do sometimes increase with higher SOV up to about 50%, tailing off thereafter (where SOV is above 50%, advertising is very likely to be light).

vi) Interaction with promotions

Clearly, advertising effects can be confounded with promotions, which may also influence short-term purchase probabilities (some would expect more so). Roberts’s analysis of TVSpan specifically allows for promotion effects as well as PV bias, by the same method (analysing the purchase occasions within ‘promoted’ and ‘unpromoted’ groups separately before combining the results). So the TVSpan results reported can be regarded as free from ‘promotion bias’.

Roberts (2000) reports a comparison of repeat-purchasing rates (over the next 12 months) between those who bought at normal price and those who bought under a price promotion. He found that repeat purchase rates were lower for the promotion buyers than for those buying at normal price, and that for both groups purchases preceded by advertising (during the previous 28 days) showed a slightly higher repeat rate than purchases without advertising. So advertising can add something even when brands are bought on promotion.

Jones (1995) had earlier published evidence about his ‘top quintile’ (7 brands) showing that STAS was substantially higher when combined with price promotion than when it was not. He argues that the combination of successful advertising and promotion, combined with “advertising intensity” (continuing investment in the successful advertising), leads to the best longer-term growth.

TVSpan data have shown that advertising typically works harder if supported by in-store promotion (Beaumont 2003). That is, those who have seen the brand advertising are more likely to choose the brand, when on promotion, than those who have not seen advertising. But there are exceptions. With some ‘deep’ price promotions, the switching caused by the promotion will swamp any advertising effect.

There is much work still to be done to understand the interactive effects (if there are any) between price promotions and advertising.

vii) Brand loyalty effects

TVSpan analyses confirm that most buyers have split repertoires (Beaumont 2003). High loyals are defined as those for whom the brand accounts for at least 40% of category purchases, low loyals less than 10%, medium loyals between 10% and 40%. The low loyals were found to respond most strongly when OTS were received in the previous week (uplift of +23%): by contrast, medium and high loyals only increased by 4% and 1% respectively. Adlab analyses found similar patterns. The high loyals, however, showed very little decay effect with greater separation of the last OTS from the day of purchase. This suggests that high loyals need less stimulus to maintain their loyalty. And reinforces the importance of advertising reaching light occasional buyers of the brand.

viii) Viewing weight effects

Broadbent (1999), picking up evidence from TVSpan (see above), believes that light TV viewers tend to respond more strongly to OTS than heavy TV viewers, since the latter are subjected to greater quantities of competing ads. He suggests that if the response in each viewing group is in fact linear with increasing OTS, but at different rates (as suggested above), the apparent diminishing returns found by looking at the total sample of purchases could be misleading (simply because the higher-OTS purchases come disproportionately from the heavier viewers, who are increasing more slowly). If this hypothesis (that response within a viewing level is linear) were found to be true, Broadbent argues that increasing total GRPs could sometimes be a more effective strategy than planning for maximum exposures of, say, 2 OTS in the relevant period. We need more evidence here.

6. Future progress

We have shown that individual-level effects of advertising on purchasing behaviour really happen and can be separated out and measured.

These individual level sales effects are vital evidence as to whether, and how well, our brand’s advertising is working compared to the competition, and they vary according to factors relevant to media planning (the delivery of the advertising dosage) – which means we can learn about media and advertising strategy from observing these sales effects.

What can we do to exploit this new knowledge (i.e. single  source method)?

We propose three lines of attack. First, there is a need for further analysis and exploration of the data available in the various single-source panels, from which all our current knowledge comes. Our basic knowledge, outlined in this paper, is still sketchy. For example, very little has so far been done to investigate OTS in media other than television. Though we do know, from Jones’s work (Jones 2002), that print media show similar effects. More of such work is urgently needed.

Secondly, we need more true single-source panels. Particularly ones that track exposure to multiple media, not just TV. Procter & Gamble’s leadership in supporting Arbitron and VNU in creating ‘the Apollo Project’ is to be applauded. Aptly named, ‘the Apollo Project’ has the potential to benefit the entire marketing community as it delivers learning, not just about creating and managing such a single-source panel, but also about advertising and media effects.

Project Apollo is an exciting new initiative, one that will inevitably be followed by others around the world. Emerging portable meter technology will make true single-source multi-media panels much more feasible than they are today.

Such multi-media single-source panels will produce extraordinary amounts of data. And analysis will be extraordinarily complex. There is much to be learnt about handling single-source data and gaining knowledge from it. The pioneering work described in this report provides a foundation for project Apollo and other panels that will emerge around the world.

Thirdly, we believe the time has come to develop alternative methods of data collection for industries not easily covered by consumer panels. To provide marketers with a practical, fast and cost-effective method of checking their marketing campaigns as they run – even when maintenance, not growth, is the outcome.

True single-source panels are very expensive, afflicted by small sample problems, and can have difficulty sustaining investment. Apart from cost, small samples mean that to study any brand in a market, even a brand leader, requires a long build-up over time before enough purchase occasions for that brand are accumulated. This inhibits any tactical use of the data for ongoing marketing management. The BehaviorScan system attempts to overcome this problem by experimental control and manipulation of ad deliveries, but even so it is not immune from these two problems.

It would be valuable if we could develop a survey-based method which could link purchasing and receipt of advertising at the individual level. We believe that the Internet now makes this a practical possibility (McDonald 1970; Boon and McDonald 2005). The aim would be to develop questions through which we could link, single-source, respondents’ probabilities of exposure to their responses- not only behavioural responses but other measures of salience (essential for the growing body of media advertising which is not concerned with frequently-purchased products).

In summary, the single source method can reveal advertising’s effect on sales, even when competition and the long-term nature of advertising effects conspire to prevent these sales effects showing up in changes in overall sales volume or market share. Advertising can therefore be made accountable, and from measuring these sales we can learn which are the most effective advertising and media strategies; the only thing standing in our way is a little necessary R&D… and this has started.

 

Footnotes:

⁽¹⁾ Brand share comparisons like STAS can also be affected by the ways the measures can be defined. For example, Jones in his analysis included all the data, including from non-buyers of the brand and those who received no advertising. McDonald, in the original JWT panel analysis and in later analysis of the Adlab panel (McDonald 1996; 1970), counted only those panel members who did buy the brand at least once and saw some advertising. Because it alters the relative numbers allotted to the different cells in the contingency table, this can make a difference to the resulting ratio. When comparing results from different sources, it is important to be aware of possible differences in definition of this sort.

⁽²⁾ Unfortunately, his analysis was flawed, as Broadbent (1999) has clearly shown, because he did not separate out the number of purchases with 1 OTS from the much smaller number with more than 1. To get a correct measure of how returns diminish with increasing OTS, it is necessary to calculate the brand shares for 2, 3, 4 or more OTS (until one reaches the point where the sample is too small to break down further).